SYSTEM AND METHOD FOR INFERRING USER INTENT FROM SPEECH INPUTS

[0001] This application is a continuation of U.S. Ser. No. 16/204,467, filed on Nov. 29, 2018, entitled SYSTEM AND METHOD FOR INFERRING USER INTENT FROM SPEECH INPUTS, which continuation of U.S. Ser. No. 14/298,725, filed on Jun. 6, 2014, entitled SYSTEM AND METHOD FOR INFERRING USER INTENT FROM SPEECH INPUTS, which claims priority from U.S. Provisional Ser. No. 61/832,896, filed on Jun. 9, 2013, entitled SYSTEM AND METHOD FOR INFERRING USER INTENT FROM SPEECH INPUTS, which are hereby incorporated by reference in their entirety for all purposes.

[0002] The disclosed implementations relate generally to digital assistants, and more specifically, to processing a speech input to a infer user intent therefrom.

[0003] Just like human personal assistants, digital assistants or virtual assistants can perform requested tasks and provide requested advice, information, or services. An assistant's ability to fulfill a user's request is dependent on the assistant's correct comprehension of the request or instructions. Recent advances in natural language processing have enabled users to interact with digital assistants using natural language, in spoken or textual forms, rather than employing a conventional user interface (e.g., menus or programmed commands). Such digital assistants can interpret the user's input to infer the user's intent, translate the inferred intent into actionable tasks and parameters, execute operations or deploy services to perform the tasks, and produce outputs that are intelligible to the user. Ideally, the outputs produced by a digital assistant should fulfill the user's intent expressed during the natural language interaction between the user and the digital assistant.

[0004] In order to perform natural language processing on speech inputs, the speech input is first converted to text (e.g., with a speech-to-text processor), and the converted text is then analyzed by a natural language processor to infer the user's intent. Consequently, any errors in the speech-to-text conversion (e.g., incorrect recognition of words in the speech input) will be propagated to the natural language processor, which may be unable to infer the user's intent due to the incorrect transcription of the words. For example, if a user provides a speech input such as “find show times for ‘Hansel and Gretel,”’ and the speech-to-text processor incorrectly converts the input to “find show times for cancel and Gretel,” the natural language processor may be unable to find any movies with the name “cancel and Gretel,” and thus, unable to provide a satisfactory response to the user.

[0005] Accordingly, there is a need for systems and methods to infer user intent from a speech input so as to account for possible speech recognition errors.

[0006] The implementations described herein provide natural language processing systems and methods for determining a user's intent from a natural language input.

[0007] Natural language processing is specifically designed to free users from having to know arcane syntax and specialized command structures in order to communicate with computers. This feature of natural language processing is particularly useful in the context of digital assistants, because different people will formulate requests in different ways, and the digital assistant should be flexible enough to understand the user's intent regardless of the particular syntax or wording of a request. In particular, a digital assistant should be able to infer that speech inputs such as “set an alarm for six AM” and “wake me up tomorrow at six” are both requesting the same action. Natural language processors are, therefore, able to infer a user's intent by attempting to understand the meaning of the input, rather than by simply processing commands that are input according to a known syntax or command structure.

[0008] One technique for accomplishing this, as described herein, is to process inputs using an ontology that is organized into “domains,” where domains relate to real-world concepts and/or tasks that the digital assistant can perform. In some implementations, a text string corresponding to a speech input is analyzed in light of the ontology to determine a domain that the text string most likely implicates. After (or as part of) determining the domain, the natural language processor identifies an “actionable intent” (i.e., a task that can be performed by the digital assistant) to be performed in response to the user's input.

[0009] The domain and/or the actionable intent is determined, at least in part, based on the particular words in the text string. In particular, domains are associated with lists of words that relate to the domain. For example, the word “remind” might be included in a word list for a “reminder” domain, but might not appear in a “restaurant reservation” domain. Similarly, the word “reservation” may be associated with a “restaurant reservation” domain, but not with a “reminder” domain.

[0010] Moreover, some domains are associated with “named entity” lists that include names of particular people, places, or things of particular salience to the domain. For example, a “movie” domain is associated with a named entity list including a list of known movie names. Similarly, a “communication” domain is associated with a contact list and/or address book including names of contacts. Accordingly, the natural language processor will be able to quickly identify that the phrase “The Breakfast Club” identifies a particular movie; that the phrase “Cheesecake Factory” identifies a particular restaurant; that the name “Andy Bernard” identifies a particular person; and so on.

[0011] The natural language processor also determines a confidence score representing how well or to what extent a user input matches a particular domain or actionable intent. The confidence score can be used, for example, to help determine which of two candidate domains is most likely to accurately reflect or represent the intent of the input. The confidence score can also be used to determine whether any of the domains or actionable intents are sufficiently relevant to the user input to justify selection of that domain or actionable intent. For example, in some implementations, the natural language processor is configured so that if no candidate domain satisfies a predetermined confidence threshold, the digital assistant will not provide a response.

[0012] Moreover, the natural language processor may be able to determine a likely domain for a given user input, but be unable to determine what the user is specifically requesting. For example, the natural language processor may be able to determine that an input “movie times for Titanic” relates to a movie domain, but it is unable to satisfy the user's request because it does not recognize that the word “Titanic” is a movie name (e.g., because it is not in the list of known movie titles). These problems may be especially apparent where a speech-to-text (STT) process provides imperfect transcriptions. For example, if an STT process mistakenly transcribes the speech input “movie times for Argo” as “movie times for are go,” the natural language processor may be unable to provide a suitable response because, although it can determine that the request relates to a “movie” domain, it cannot identify what movie the user was referring to.

[0013] According to some implementations described herein, errors such as those described above are mitigated by performing a domain-specific word-matching process on the text string after the domain is determined. For example, the natural language processor receives a text string corresponding to a speech input, and determines a domain of the text string by applying a first word-matching process (e.g., comparing the words in the text string to words associated with many different domains). Once the domain is determined, if the natural language processor cannot adequately determine the user's intent (e.g., the specific action that the user is requesting), the natural language processor performs a second word-matching process of at least a part of the text string in order to better identify what the user may have said.

[0014] For example, in some implementations, the second word-matching process includes applying a more relaxed matching criteria than the first word-matching process. Specifically, the first word-matching process may require an exact match between words in the text string and words in (or associated with) the domains, whereas the second word-matching process requires only an approximate string match or a phonetic match.

[0015] Also, in some implementations, the second word-matching process is limited to words that are associated with the domain that was determined for the text string. As a specific example, for a text string “show times for cancel and kettle,” the natural language processor easily determines that the text string relates to a “movie” domain, but does not initially find any matches or relevance to the phrase “cancel and kettle.” Thus, the natural language processor re-processes the text string (or a portion thereof) to search for word matches in a vocabulary that is specific to the “movie” domain (e.g., a list of known movie titles). By limiting the search scope to in-domain words, and/or applying more relaxed matching criteria, the natural language processor can determine that the text actually refers to the movie “Hansel and Gretel,” and thus provide a suitable response to the user where it otherwise may have failed to do so.

[0016] Another way to mitigate errors such as those described above is to apply natural language processing to multiple different text strings from the STT processor, identify domains and/or actionable intents for each of the multiple text strings, and adopt the domain and/or actionable intent with the highest confidence score. Specifically, an STT processor typically generates multiple candidate text strings for a given speech input, and selects only what it determines to be the “best” to provide to the natural language processor (e.g., based on a speech recognition confidence score). However, this selection may not always accurately reflect the speech input (e.g., “cancel and kettle,” as described above). Indeed, the correct text string may have even been generated as a candidate text string, but ultimately passed over in favor of another text string.

[0017] In some implementations, the natural language processor processes several of the candidate text strings produced by the STT processor to determine a user intent for each text string. The natural language processor then selects one of the text strings based at least partially on a confidence score representing how well or to what extent the text string matches a particular domain or a particular task associated with a domain. Thus, the natural language processor can leverage its powerful intent deduction abilities to help select the text string that most likely reflects the actual words in the speech input.

[0018] The implementations disclosed herein provide methods, systems, computer readable storage medium and user interfaces for determining a user intent from a text string and/or a speech input.

[0019] According to some implementations, a method is performed at an electronic device with one or more processors and memory storing one or more programs for execution by the one or more processors. A text string corresponding to a speech input is provided, wherein the text string comprises a first portion and a second portion. A domain of the text string is determined using natural language processing by applying a first word-matching process to at least the first portion of the text string. It is determined whether the second portion of the text string matches at least one word of a set of words associated with the domain by applying a second word-matching process to the second portion of the text string. Upon determining that the second portion of the text string matches at least one word of the set of words, a user intent is determined from the text string based at least in part on the domain and the at least one word of the set of words.

[0020] In some implementations, word-matching criteria applied by the second word-matching process is more relaxed than word-matching criteria applied by the first word-matching process. In some implementations, the first word-matching process requires exact matches between words, and the second word-matching process does not require exact matches between words. In some implementations, the second word-matching process applies approximate string matching techniques.

[0021] In some implementations, the second word-matching process comprises determining an edit distance between a word of the second portion of the text string and a word of the set of words associated with the domain. In some implementations, the edit distance is a Levenshtein distance.

[0022] In some implementations, the second word-matching process applies phonetic matching techniques.

[0023] In some implementations, one or more actions to satisfy the user intent are performed. In some implementations, performing the one or more actions includes making a restaurant reservation, providing information about a movie to a user, initiating a communication with another user, etc.

[0024] According to some implementations, a method is performed at an electronic device with one or more processors and memory storing one or more programs for execution by the one or more processors. A plurality of candidate text strings for a speech input are provided, wherein the candidate text strings are generated by a speech recognition process. At least a subset of the plurality of candidate text strings is selected, the subset including at least two candidate text strings. A plurality of user intents are determined from the selected subset of the plurality of text strings. A user intent is selected from the plurality of user intents.

[0025] In some implementations, the plurality of user intents are associated with a plurality of intent deduction confidence scores. In some implementations, the user intent is selected based on the plurality of intent deduction confidence scores. In some implementations, the selected user intent is the user intent with the highest intent deduction confidence score.

[0026] In some implementations, the plurality of candidate text strings are associated with a plurality of speech recognition confidence scores. In some implementations, the user intent is selected based on the speech recognition confidence scores and the intent deduction confidence scores.

[0027] In some implementations, a plurality of composite confidence scores are determined for the plurality of user intents. In some implementations, a composite confidence score is a combination of a respective speech recognition confidence score and a respective intent deduction confidence score. In some implementations, a composite confidence score is a sum of a respective speech recognition confidence score and a respective intent deduction confidence score. In some implementations, a composite confidence score is an average of a respective speech recognition confidence score and a respective intent deduction confidence score. In some implementations, a composite confidence score is a weighted average of a respective speech recognition confidence score and a respective intent deduction confidence score.

[0028] In some implementations, the selected user intent is the user intent with the highest composite confidence score.

[0029] In some implementations, one or more actions to satisfy the user intent are performed. In some implementations, performing the one or more actions includes making a restaurant reservation, providing information about a movie to a user, initiating a communication with another user, etc.

[0030] According to some implementations, a method is performed at an electronic device with one or more processors and memory storing one or more programs for execution by the one or more processors. A plurality of text strings corresponding to a speech input are provided, wherein the plurality of text strings is associated with a plurality of speech recognition confidence scores and a plurality of user intents, and wherein the plurality of user intents is associated with a plurality of intent deduction confidence scores. One of the plurality of user intents is selected based on one or both of the speech recognition confidence scores and the intent deduction confidence scores.

[0031] In some implementations, a plurality of composite confidence scores are determined for the plurality of user intents. In some implementations, a composite confidence score is a combination of a respective speech recognition confidence score and a respective intent deduction confidence score. In some implementations, a composite confidence score is a sum of a respective speech recognition confidence score and a respective intent deduction confidence score. In some implementations, a composite confidence score is an average of a respective speech recognition confidence score and a respective intent deduction confidence score. In some implementations, a composite confidence score is a weighted average of a respective speech recognition confidence score and a respective intent deduction confidence score.

[0032] In sonic implementations, the selected user intent is the user intent with the highest composite confidence score.

[0033] In some implementations, one or more actions to satisfy the user intent are performed. In some implementations, performing the one or more actions includes making a restaurant reservation, providing information about a movie to a user, initiating a communication with another user, etc.

[0034] In accordance with some embodiments, an electronic device includes one or more processors, memory, and one or more programs; the one or more programs are stored in the memory and configured to be executed by the one or more processors and the one or more programs include instructions for performing the operations of any of the methods described above. In accordance with some embodiments, a computer readable storage medium has stored therein instructions which when executed by an electronic device, cause the device to perform the operations of any of the methods described above. In accordance with some embodiments, an electronic device includes: means for performing the operations of any of the methods described above. In accordance with some embodiments, an information processing apparatus, for use in an electronic device, includes means for performing the operations of any of the methods described above.

[0035] The details of one or more implementations of the subject matter described in this specification are set forth in the accompanying drawings and the description below. Other features, aspects, and advantages of the subject matter will become apparent from the description, the drawings, and the claims.

[0044] Like reference numerals refer to corresponding parts throughout the drawings.

[0045] FIG. 1 is a block diagram of an operating environment 100 of a digital assistant according to some implementations. The terms “digital assistant,” “virtual assistant,” “intelligent automated assistant,” or “automatic digital assistant,” refer to any information processing system that interprets natural language input in spoken and/or textual form to infer user intent, and performs actions based on the inferred user intent. For example, to act on a inferred user intent, the system can perform one or more of the following: identifying a task flow with steps and parameters designed to accomplish the inferred user intent, inputting specific requirements from the inferred user intent into the task flow; executing the task flow by invoking programs, methods, services, APIs, or the like; and generating output responses to the user in an audible (e.g. speech) and/or visual form.

[0046] Specifically, a digital assistant is capable of accepting a user request at least partially in the form of a natural language command, request, statement, narrative, and/or inquiry. Typically, the user request seeks either an informational answer or performance of a task by the digital assistant. A satisfactory response to the user request is either provision of the requested informational answer, performance of the requested task, or a combination of the two. For example, a user may ask the digital assistant a question, such as “Where am I right now?” Based on the user's current location, the digital assistant may answer, “You are in Central Park near the west gate.” The user may also request the performance of a task, for example, “Please invite my friends to my girlfriend's birthday party next week.” In response, the digital assistant may acknowledge the request by saying “Yes, right away,” and then send a suitable calendar invite on behalf of the user to each of the user' friends listed in the user's electronic address book. During performance of a requested task, the digital assistant sometimes interacts with the user in a continuous dialogue involving multiple exchanges of information over an extended period of time. There are numerous other ways of interacting with a digital assistant to request information or performance of various tasks. In addition to providing verbal responses and taking programmed actions, the digital assistant also provides responses in other visual or audio forms, e.g., as text, alerts, music, videos, animations, etc.

[0047] An example of a digital assistant is described in Applicant's U.S. Utility application Ser. No. 12/987,982 for “Intelligent Automated Assistant,” filed Jan. 10, 2011, the entire disclosure of which is incorporated herein by reference.

[0048] As shown in FIG. 1, in some implementations, a digital assistant is implemented according to a client-server model. The digital assistant includes a client-side portion 102a, 102b (hereafter “DA client 102”) executed on a user device 104a, 104b, and a server-side portion 106 (hereafter “DA server 106”) executed on a server system 108. The DA client 102 communicates with the DA server 106 through one or more networks 110. The DA client 102 provides client-side functionalities such as user-facing input and output processing and communications with the DA-server 106. The DA server 106 provides server-side functionalities for any number of DA-clients 102 each residing on a respective user device 104.

[0049] In some implementations, the DA server 106 includes a client-facing I/O interface 112, one or more processing modules 114, data and models 116, and an I/O interface to external services 118. The client-facing I/O interface facilitates the client-facing input and output processing for the digital assistant server 106, The one or more processing modules 114 utilize the data and models 116 to determine the user's intent based on natural language input and perform task execution based on inferred user intent. In some implementations, the DA-server 106 communicates with external services 120 through the network(s) 110 for task completion or information acquisition. The I/O interface to external services 118 facilitates such communications.

[0050] Examples of the user device 104 include, but are not limited to, a handheld computer, a personal digital assistant (PDA), a tablet computer, a laptop computer, a desktop computer, a cellular telephone, a smart phone, an enhanced general packet radio service (EGPRS) mobile phone, a media player, a navigation device, a game console, a television, a remote control, or a combination of any two or more of these data processing devices or other data processing devices. More details on the user device 104 are provided in reference to an exemplary user device 104 shown in FIG. 2.

[0051] Examples of the communication network(s) 110 include local area networks (“LAN”) and wide area networks (“WAN”), e.g., the Internet. The communication network(s) 110 may be implemented using any known network protocol, including various wired or wireless protocols, such as e.g., Ethernet, Universal Serial Bus (USB), FIREWIRE, Global System for Mobile Communications (GSM), Enhanced Data GSM Environment (EDGE), code division multiple access (CDMA), time division multiple access (TDMA), Bluetooth, Wi-Fi, voice over Internet Protocol (VOIP), Wi-MAX, or any other suitable communication protocol.

[0052] The server system 108 is implemented on one or more standalone data processing apparatus or a distributed network of computers. In some implementations, the server system 108 also employs various virtual devices and/or services of third party service providers (e.g., third-party cloud service providers) to provide the underlying computing resources and/or infrastructure resources of the server system 108.

[0053] Although the digital assistant shown in FIG. 1 includes both a client-side portion (e.g., the DA-client 102) and a server-side portion (e.g., the DA-server 106), in some implementations, the functions of a digital assistant is implemented as a standalone application installed on a user device. In addition, the divisions of functionalities between the client and server portions of the digital assistant can vary in different implementations. For example, in some implementations, the DA client is a thin-client that provides only user-facing input and output processing functions, and delegates all other functionalities of the digital assistant to a backend server.

[0054] FIG. 2 is a block diagram of a user-device 104 in accordance with some implementations. The user device 104 includes a memory interface 202, one or more processors 204, and a peripherals interface 206. The various components in the user device 104 are coupled by one or more communication buses or signal lines. The user device 104 includes various sensors, subsystems, and peripheral devices that are coupled to the peripherals interface 206. The sensors, subsystems, and peripheral devices gather information and/or facilitate various functionalities of the user device 104.

[0055] For example, a motion sensor 210, a light sensor 212, and a proximity sensor 214 are coupled to the peripherals interface 206 to facilitate orientation, light, and proximity sensing functions. One or more other sensors 216, such as a positioning system (e.g., GPS receiver), a temperature sensor, a biometric sensor, a gyro, a compass, an accelerometer, and the like, are also connected to the peripherals interface 206, to facilitate related functionalities.

[0056] In some implementations, a camera subsystem 220 and an optical sensor 222 are utilized to facilitate camera functions, such as taking photographs and recording video clips. Communication functions are facilitated through one or more wired and/or wireless communication subsystems 224, which can include various communication ports, radio frequency receivers and transmitters, and/or optical (e.g., infrared) receivers and transmitters. An audio subsystem 226 is coupled to speakers 228 and a microphone 230 to facilitate voice-enabled functions, such as voice recognition, voice replication, digital recording, and telephony functions.

[0057] In some implementations, an I/O subsystem 240 is also coupled to the peripheral interface 206. The I/O subsystem 240 includes a touch screen controller 242 and/or other input controller(s) 244. The touch-screen controller 242 is coupled to a touch screen 246. The touch screen 246 and the touch screen controller 242 can, for example, detect contact and movement or break thereof using any of a plurality of touch sensitivity technologies, such as capacitive, resistive, infrared, surface acoustic wave technologies, proximity sensor arrays, and the like. The other input controller(s) 244 can be coupled to other input/control devices 248, such as one or more buttons, rocker switches, thumb-wheel, infrared port, USB port, and/or a pointer device such as a stylus.

[0058] In some implementations, the memory interface 202 is coupled to memory 250. The memory 250 can include high-speed random access memory and/or non-volatile memory, such as one or more magnetic disk storage devices, one or more optical storage devices, and/or flash memory (e.g., NAND, NOR).

[0059] in some implementations, the memory 250 stores an operating system 252, a communication module 254, a user interface module 256, a sensor processing module 258, a phone module 260, and applications 262. The operating system 252 includes instructions for handling basic system services and for performing hardware dependent tasks. The communication module 254 facilitates communicating with one or more additional devices, one or more computers and/or one or more servers. The user interface module 256 facilitates graphic user interface processing and output processing using other output channels (e.g., speakers). The sensor processing module 258 facilitates sensor-related processing and functions. The phone module 260 facilitates phone-related processes and functions. The application module 262 facilitates various functionalities of user applications, such as electronic-messaging, web browsing, media processing, Navigation, imaging and/or other processes and functions.

[0060] As described in this specification, the memory 250 also stores client-side digital assistant instructions (e.g., in a digital assistant client module 264) and various user data 266 (e.g., user-specific vocabulary data, preference data, and/or other data such as the user's electronic address book, to-do lists, shopping lists, user-specified name pronunciations, etc.) to provide the client-side functionalities of the digital assistant.

[0061] In various implementations, the digital assistant client module 264 is capable of accepting voice input (e.g., speech input), text input, touch input, and/or gestural input through various user interfaces (e.g., the I/O subsystem 244) of the user device 104. The digital assistant client module 264 is also capable of providing output in audio (e.g., speech output), visual, and/or tactile forms. For example, output can be provided as voice, sound, alerts, text messages, menus, graphics, videos, animations, vibrations, and/or combinations of two or more of the above. During operation, the digital assistant client module 264 communicates with the digital assistant server using the communication subsystems 224.

[0062] In some implementations, the digital assistant client module 264 includes a speech synthesis module 265. The speech synthesis module 265 synthesizes speech outputs for presentation to the user. The speech synthesis module 265 synthesizes speech outputs based on text provided by the digital assistant. For example, the digital assistant generates text to provide as an output to a user, and the speech synthesis module 265 converts the text to an audible speech output. The speech synthesis module 265 uses any appropriate speech synthesis technique in order to generate speech outputs from text, including but not limited to concatenative synthesis, unit selection synthesis, diphone synthesis, domain-specific synthesis, formant synthesis, articulatory synthesis, hidden Markov model (HMM) based synthesis, and sinewave synthesis.

[0063] In some implementations, instead of (or in addition to) using the local speech synthesis module 265, speech synthesis is performed on a remote device (e.g., the server system 108), and the synthesized speech is sent to the user device 104 for output to the user. For example, this occurs in some implementations where outputs for a digital assistant are generated at a server system. And because server systems generally have more processing power or resources than a user device, it may be possible to obtain higher quality speech outputs than would be practical with client-side synthesis.

[0064] In some implementations, the digital assistant client module 264 utilizes the various sensors, subsystems and peripheral devices to gather additional information from the surrounding environment of the user device 104 to establish a context associated with a user, the current user interaction, and/or the current user input. In some implementations, the digital assistant client module 264 provides the context information or a subset thereof with the user input to the digital assistant server to help infer the user's intent. In some implementations, the digital assistant also uses the context information to determine how to prepare and delivery outputs to the user.

[0065] In some implementations, the context information that accompanies the user input includes sensor information, e.g., lighting, ambient noise, ambient temperature, images or videos of the surrounding environment, etc. In some implementations, the context information also includes the physical state of the device, e.g., device orientation, device location, device temperature, power level, speed, acceleration, motion patterns, cellular signals strength, etc. In some implementations, information related to the software state of the user device 104, e.g., running processes, installed programs, past and present network activities, background services, error logs, resources usage, etc., of the user device 104 are provided to the digital assistant server as context information associated with a user input.

[0066] In some implementations, the DA client module 264 selectively provides information (e.g., user data 266) stored on the user device 104 in response to requests from the digital assistant server. In some implementations, the digital assistant client module 264 also elicits additional input from the user via a natural language dialogue or other user interfaces upon request by the digital assistant server 106. The digital assistant client module 264 passes the additional input to the digital assistant server 106 to help the digital assistant server 106 in intent deduction and/or fulfillment of the user's intent expressed in the user request.

[0067] In various implementations, the memory 250 includes additional instructions or fewer instructions. Furthermore, various functions of the user device 104 may be implemented in hardware and/or in firmware, including in one or more signal processing and/or application specific integrated circuits.

[0068] FIG. 3A is a block diagram of an example digital assistant system 300 in accordance with some implementations. In some implementations, the digital assistant system 300 is implemented on a standalone computer system. In some implementations, the digital assistant system 300 is distributed across multiple computers. In some implementations, some of the modules and functions of the digital assistant are divided into a server portion and a client portion, where the client portion resides on a user device (e.g., the user device 104) and communicates with the server portion (e.g., the server system 108) through one or more networks, e.g., as shown in FIG. 1. In some implementations, the digital assistant system 300 is an implementation of the server system 108 (and/or the digital assistant server 106) shown in FIG. 1. It should be noted that the digital assistant system 300 is only one example of a digital assistant system, and that the digital assistant system 300 may have more or fewer components than shown, may combine two or more components, or may have a different configuration or arrangement of the components. The various components shown in FIG. 3A may be implemented in hardware, software instructions for execution by one or more processors, firmware, including one or more signal processing and/or application specific integrated circuits, or a combination of thereof.

[0069] The digital assistant system 300 includes memory 302, one or more processors 304, an input/output (I/O) interface 306, and a network communications interface 308. These components communicate with one another over one or more communication buses or signal lines 310.

[0070] In some implementations, the memory 302 includes a non-transitory computer readable medium, such as high-speed random access memory and/or a non-volatile computer readable storage medium (e.g., one or more magnetic disk storage devices, flash memory devices, or other non-volatile solid-state memory devices).

[0071] In some implementations, the I/O interface 306 couples input/output devices 316 of the digital assistant system 300, such as displays, keyboards, touch screens, and microphones, to the user interface module 322. The I/O interface 306, in conjunction with the user interface module 322, receives user inputs (e.g., voice input, keyboard inputs, touch inputs, etc.) and processes them accordingly. In some implementations, e.g., when the digital assistant is implemented on a standalone user device, the digital assistant system 300 includes any of the components and I/O and communication interfaces described with respect to the user device 104 in FIG. 2. In some implementations, the digital assistant system 300 represents the server portion of a digital assistant implementation, and interacts with the user through a client-side portion residing on a user device (e.g., the user device 104 shown in FIG. 2).

[0072] In some implementations, the network communications interface 308 includes wired communication port(s) 312 and/or wireless transmission and reception circuitry 314. The wired communication port(s) receive and send communication signals via one or more wired interfaces, e.g., Ethernet, Universal Serial Bus (USB), FIREWIRE, etc. The wireless circuitry 314 receives and sends RF signals and/or optical signals from/to communications networks and other communications devices. The wireless communications may use any of a plurality of communications standards, protocols and technologies, such as GSM, EDGE, CDMA, TDMA, Bluetooth, Wi-Fi, VoIP, Wi-MAX, or any other suitable communication protocol. The network communications interface 308 enables communication between the digital assistant system 300 with networks, such as the Internet, an intranet and/or a wireless network, such as a cellular telephone network, a wireless local area network (LAN) and/or a metropolitan area network (MAN), and other devices.

[0073] In some implementations, memory 302, or the computer readable storage media of memory 302, stores programs, modules, instructions, and data structures including all or a subset of: an operating system 318, a communications module 320, a user interface module 322, one or more applications 324, and a digital assistant module 326. The one or more processors 304 execute these programs, modules, and instructions, and reads/writes from/to the data structures.

[0074] The operating system 318 (e.g., Darwin, RTXC, LINUX, UNIX, OS X, WINDOWS, or an embedded operating system such as VxWorks) includes various software components and/or drivers for controlling and managing general system tasks (e.g., memory management, storage device control, power management, etc.) and facilitates communications between various hardware, firmware, and software components.

[0075] The communications module 320 facilitates communications between the digital assistant system 300 with other devices over the network communications interface 308. For example, the communication module 320 may communicate with the communication interface 254 of the device 104 shown in FIG. 2. The communications module 320 also includes various components for handling data received by the wireless circuitry 314 and/or wired communications port 312.

[0076] The user interface module 322 receives commands and/or inputs from a user via the I/O interface 306 (e.g., from a keyboard, touch screen, pointing device, controller, and/or microphone), and generates user interface objects on a display. The user interface module 322 also prepares and delivers outputs (e.g., speech, sound, animation, text, icons, vibrations, haptic feedback, and light, etc.) to the user via the I/O interface 306 (e.g., through displays, audio channels, speakers, and touch-pads, etc.).

[0077] The applications 324 include programs and/or modules that are configured to be executed by the one or more processors 304. For example, if the digital assistant system is implemented on a standalone user device, the applications 324 may include user applications, such as games, a calendar application, a navigation application, or an email application. If the digital assistant system 300 is implemented on a server farm, the applications 324 may include resource management applications, diagnostic applications, or scheduling applications, for example.

[0078] The memory 302 also stores the digital assistant module (or the server portion of a digital assistant) 326. In some implementations, the digital assistant module 326 includes the following sub-modules, or a subset or superset thereof: an input/output processing module 328, a speech-to-text (STT) processing module 330, a natural language processing module 332, a dialogue flow processing module 334, a task flow processing module 336, a service processing module 338, and a speech interaction error detection module 339. Each of these modules has access to one or more of the following data and models of the digital assistant 326, or a subset or superset thereof: ontology 360, vocabulary index 344, user data 348, task flow models 354, and service models 356.

[0079] In some implementations, using the processing modules, data, and models implemented in the digital assistant module 326, the digital assistant performs at least some of the following: identifying a user's intent expressed in a natural language input received from the user; actively eliciting and obtaining information needed to fully infer the user's intent (e.g., by disambiguating words, names, intentions, etc.); determining the task flow for fulfilling the inferred intent and executing the task flow to fulfill the inferred intent.

[0080] In some implementations, as shown in FIG. 3B, the I/O processing module 328 interacts with the user through the I/O devices 316 in FIG. 3A or with a user device (e.g., a user device 104 in FIG. 1) through the network communications interface 308 in FIG. 3A to obtain user input (e.g., a speech input) and to provide responses (e.g., as speech outputs) to the user input. The I/O processing module 328 optionally obtains context information associated with the user input from the user device, along with or shortly after the receipt of the user input. The context information includes user-specific data, vocabulary, and/or preferences relevant to the user input. In some implementations, the context information also includes software and hardware states of the device (e.g., the user device 104 in FIG. 1) at the time the user request is received, and/or information related to the surrounding environment of the user at the time that the user request was received. In some implementations, the I/O processing module 328 also sends follow-up questions to, and receives answers from, the user regarding the user request. When a user request is received by the I/O processing module 328 and the user request contains a speech input, the I/O processing module 328 forwards the speech input to the speech-to-text (STT) processing module 330 for speech-to-text conversions.

[0081] The speech-to-text processing module 330 (or speech recognizer) receives speech input (e.g., a user utterance captured in a voice recording) through the I/O processing module 328. In some implementations, the STT processing module 330 uses various acoustic and language models to recognize the speech input as a sequence of phonemes, and ultimately, a sequence of words or tokens written in one or more languages. The STT processing module 330 can be implemented using any suitable speech recognition techniques, acoustic models, and language models, such as Hidden Markov Models, Dynamic Time Warping (DTW)-based speech recognition, and other statistical and/or analytical techniques. In some implementations, the speech-to-text processing can be performed at least partially by a third party service or on the user's device. Once the STT processing module 330 obtains the result of the speech-to-text processing, e.g., a sequence of words or tokens, it passes the result to the natural language processing module 332 for intent deduction. In some implementations, the STT module 330 resides on a server computer (e.g., the server system 108), while in some implementations, it resides on a client device (e.g., the user device 104).

[0082] In some implementations, the STT processing module 330 provides more than one speech recognition result to the natural language processing module 332 for intent deduction. Specifically, the STT processing module 330 produces a set of candidate text strings, each representing a possible transcription of a speech input. In some implementations, each of the candidate text strings is associated with a speech recognition confidence score representing a confidence that the candidate text string is a correct transcription of the speech input.

[0083] In some implementations, the set of candidate text strings represents a portion of the text strings that the STT processing module 330 generates for a given speech input. For example, in some implementations, the set of candidate text strings includes text strings that have a speech recognition confidence score above a predetermined threshold. In some implementations, the set of candidate text strings includes the n-best text strings. (E.g., the 2, 5, 10 or 15 best text strings (or any other appropriate number), based on speech recognition confidence scores.)

[0084] When an utterance is received, the STT processing module 330 attempts to identify the phonemes in the utterance (e.g., using an acoustic model), and then attempts to identify words that match the phonemes (e.g., using a language model). For example, if the STT processing module 330 first identifies the sequence of phonemes “tuh-may-doe” in an utterance, it then determines, based on the vocabulary 344, that this sequence corresponds to the word “tomato.”

[0085] In some implementations, the STT processing module 330 uses approximate matching techniques to determine words in an utterance. Thus, for example, the STT processing module 330 can determine that the sequence of phonemes “duh-may-doe” corresponds to the word “tomato,” even if that particular sequence of phonemes is not one of the candidate pronunciations for that word.

[0086] The natural language processing module 332 (“natural language processor”) of the digital assistant takes the text string or text strings generated by the speech-to-text processing module 330 (also referred to as a sequence of words or tokens, or “token sequence”), and attempts to determine a domain of the token sequence and/or associate the token sequence with one or more “actionable intents” recognized by the digital assistant. An “actionable intent” represents a task that can be performed by the digital assistant, and has an associated task flow implemented in the task flow models 354. The associated task flow is a series of programmed actions and steps that the digital assistant takes in order to perform the task. The scope of a digital assistant's capabilities is dependent on the number and variety of task flows that have been implemented and stored in the task flow models 354, or in other words, on the number and variety of “actionable intents” that the digital assistant recognizes. The effectiveness of the digital assistant, however, is also dependent on the assistant's ability to infer the correct “actionable intent(s)” from the user request expressed in natural language.

[0087] In some implementations, in addition to the sequence of words or tokens obtained from the speech-to-text processing module 330, the natural language processor 332 also receives context information associated with the user request, e.g., from the I/O processing module 328. The natural language processor 332 optionally uses the context information to clarify, supplement, and/or further define the information contained in the token sequence received from the speech-to-text processing module 330. The context information includes, for example, user preferences, hardware and/or software states of the user device, sensor information collected before, during, or shortly after the user request, prior interactions (e.g., dialogue) between the digital assistant and the user, and the like. As described in this specification, context information is dynamic, and can change with time, location, content of the dialogue, and other factors.

[0088] In some implementations, the natural language processing is based on e.g., ontology 360. The ontology 360 is a hierarchical structure containing many nodes, each node representing either an “actionable intent” or a “property” relevant to one or more of the “actionable intents” or other “properties”. As noted above, an “actionable intent” represents a task that the digital assistant is capable of performing, i.e., it is “actionable” or can be acted on. A “property” represents a parameter associated with an actionable intent or a sub-aspect of another property. A linkage between an actionable intent node and a property node in the ontology 360 defines how a parameter represented by the property node pertains to the task represented by the actionable intent node.

[0089] In some implementations, the ontology 360 is made up of actionable intent nodes and property nodes. Within the ontology 360, each actionable intent node is linked to one or more property nodes either directly or through one or more intermediate property nodes. Similarly, each property node is linked to one or more actionable intent nodes either directly or through one or more intermediate property nodes. For example, as shown in FIG. 3C, the ontology 360 may include a “restaurant reservation” node (i.e., an actionable intent node). Property nodes “restaurant,” “date/time” (for the reservation), and “party size” are each directly linked to the actionable intent node (i.e., the “restaurant reservation” node). In addition, property nodes “cuisine,” “price range,” “phone number,” and “location” are sub-nodes of the property node “restaurant,” and are each linked to the “restaurant reservation” node (i.e., the actionable intent node) through the intermediate property node “restaurant.” For another example, as shown in FIG. 3C, the ontology 360 may also include a “set reminder” node (i.e., another actionable intent node). Property nodes “date/time” (for the setting the reminder) and “subject” (for the reminder) are each linked to the “set reminder” node. Since the property “date/time” is relevant to both the task of making a restaurant reservation and the task of setting a reminder, the property node “date/time” is linked to both the “restaurant reservation” node and the “set reminder” node in the ontology 360.

[0090] An actionable intent node, along with its linked concept nodes, may be described as a “domain.” In the present discussion, each domain is associated with a respective actionable intent, and refers to the group of nodes (and the relationships therebetween) associated with the particular actionable intent. For example, the ontology 360 shown in FIG. 3C includes an example of a restaurant reservation domain 362 and an example of a reminder domain 364 within the ontology 360. The restaurant reservation domain includes the actionable intent node “restaurant reservation,” property nodes “restaurant,” “date/time,” and “party size,” and sub-property nodes “cuisine,” “price range,” “phone number,” and “location.” The reminder domain 364 includes the actionable intent node “set reminder,” and property nodes “subject” and “date/time.” In some implementations, the ontology 360 is made up of many domains. Each domain may share one or more property nodes with one or more other domains. For example, the “date/time” property node may be associated with many different domains (e.g., a scheduling domain, a travel reservation domain, a movie ticket domain, etc.), in addition to the restaurant reservation domain 362 and the reminder domain 364.

[0091] While FIG. 3C illustrates two example domains within the ontology 360, other domains (or actionable intents) include, for example, “initiate a phone call,” “find directions,” “schedule a meeting,” “send a message,” and “provide an answer to a question,” read a list”, “providing navigation instructions,” “provide instructions for a task” and so on. A “send a message” domain is associated with a “send a message” actionable intent node, and may further include property nodes such as “recipient(s)”, “message type”, and “message body.” The property node “recipient” may be further defined, for example, by the sub-property nodes such as “recipient name” and “message address.”

[0092] In some implementations, the ontology 360 includes all the domains (and hence actionable intents) that the digital assistant is capable of understanding and acting upon. In some implementations, the ontology 360 may be modified, such as by adding or removing entire domains or nodes, or by modifying relationships between the nodes within the ontology 360.

[0093] In some implementations, nodes associated with multiple related actionable intents may be clustered under a “super domain” in the ontology 360. For example, a “travel” super-domain may include a cluster of property nodes and actionable intent nodes related to travels. The actionable intent nodes related to travels may include “airline reservation,” “hotel reservation,” “car rental,” “get directions,” “find points of interest,” and so on. The actionable intent nodes under the same super domain (e.g., the “travels” super domain) may have many property nodes in common. For example, the actionable intent nodes for “airline reservation,” “hotel reservation,” “car rental,” “get directions,” “find points of interest” may share one or more of the property nodes “start location,” “destination,” “departure date/time,” “arrival date/time,” and “party size.”

[0094] In some implementations, each node in the ontology 360 is associated with a set of words and/or phrases that are relevant to the property or actionable intent represented by the node. The respective set of words and/or phrases associated with each node is the so-called “vocabulary” associated with the node. The respective set of words and/or phrases associated with each node can be stored in the vocabulary index 344 in association with the property or actionable intent represented by the node. For example, returning to FIG. 3B, the vocabulary associated with the node for the property of “restaurant” may include words such as “food,” “drinks,” “cuisine,” “hungry,” “eat,” “pizza,” “fast food,” “meal,” and so on. For another example, the vocabulary associated with the node for the actionable intent of “initiate a phone call” may include words and phrases such as “call,” “phone,” “dial,” “ring,” “call this number,” “make a call to,” and so on. The vocabulary index 344 optionally includes words and phrases in different languages.

[0095] The natural language processor 332 receives the token sequence (e.g., a text string) from the speech-to-text processing module 330, and determines what nodes are implicated by the words in the token sequence. In some implementations, if a word or phrase in the token sequence is found to be associated with one or more nodes in the ontology 360 (via the vocabulary index 344), the word or phrase will “trigger” or “activate” those nodes. Based on the quantity and/or relative importance of the activated nodes, the natural language processor 332 will select one of the actionable intents as the task that the user intended the digital assistant to perform. In some implementations, the domain that has the most “triggered” nodes is selected. In some implementations, the domain having the highest confidence score (e.g., based on the relative importance of its various triggered nodes) is selected. In some implementations, the domain is selected based on a combination of the number and the importance of the triggered nodes. In some implementations, additional factors are considered in selecting the node as well, such as whether the digital assistant has previously correctly interpreted a similar request from a user.

[0096] In some implementations, the natural language processor 332 determines a domain for a particular token sequence, but cannot initially determine (either at all or to a sufficient degree of confidence) a particular actionable intent for the token sequence. In some implementations, as described herein, the natural language processor 332 processes the token sequence again, for example, by relaxing its search and/or word-matching criteria, and/or by limiting its search to a particular domain or word list. Specifically, in some implementations, the natural language processor 332 relaxes a word-matching criteria relative to the initial processing (e.g., approximate string matching rather than exact string matching; phonetic matching instead of string matching, etc.). In some implementations, the natural Language processor 332 limits its subsequent processing of the token sequence to a particular list of words, such as a list of named entities that are associated with the identified domain (e.g., a list of known movie titles if the domain is a “movie” domain, a list of known restaurant names if the domain is a “restaurant reservation” domain, etc.) Further description of techniques for subsequent processing of a token sequence is provided herein, for example, with reference to FIG. 4.

[0097] In some implementations, the natural language processor 332 receives multiple token sequences (e.g., text strings) from the STT processor 330, and processes them to determine a domain and/or an actionable intent for each token sequence (or at least a subset of the token sequences from the STT processing module 330). In some implementations, each domain and/or actionable intent is associated with an intent deduction confidence score representing a confidence that the determined domain and/or actionable intent correctly reflects the intent represented by the token sequence. A further description of techniques for processing multiple token sequences is provided herein, for example, with reference to FIG. 5.

[0098] In some implementations, the digital assistant also stores names of specific entities in the vocabulary index 344, so that when one of these names is detected in the user request, the natural language processor 332 will be able to recognize that the name refers to a specific instance of a property or sub-property in the ontology. In some implementations, the names of specific entities are names of businesses, restaurants, people, movies, and the like. In some implementations, the digital assistant searches and identifies specific entity names from other data sources, such as the user's address book, a movies database, a musicians database, and/or a restaurant database. In some implementations, when the natural language processor 332 identifies that a word in the token sequence is a name of a specific entity (such as a name in the user's address book), that word is given additional significance in selecting the actionable intent within the ontology for the user request.

[0099] For example, when the words “Mr. Santo” are recognized from the user request, and the last name “Santo” is found in the vocabulary index 344 as one of the contacts in the user's contact list, then it is likely that the user request corresponds to a “send a message” or “initiate a phone call” domain. For another example, when the words “ABC Cafe” are found in the user request, and the term “ABC Cafe” is found in the vocabulary index 344 as the name of a particular restaurant in the user's city, then it is likely that the user request corresponds to a “restaurant reservation” domain.

[0100] User data 348 includes user-specific information, such as user-specific vocabulary, user preferences, user address, user's default and secondary languages, user's contact list, and other short-term or long-term information for each user. In some implementations, the natural language processor 332 uses the user-specific information to supplement the information contained in the user input to further define the user intent. For example, for a user request “invite my friends to my birthday party,” the natural language processor 332 is able to access user data 348 to determine who the “friends” are and when and where the “birthday party” would be held, rather than requiring the user to provide such information explicitly in his/her request.

[0101] Other details of searching an ontology based on a token string is described in U.S. Utility application Ser. No. 12/341,743 for “Method and Apparatus for Searching Using An Active Ontology,” filed Dec. 22, 2008, the entire disclosure of which is incorporated herein by reference.

[0102] In some implementations, once the natural language processor 332 identifies an actionable intent (or domain) based on the user request, the natural language processor 332 generates a structured query to represent the identified actionable intent. In some implementations, the structured query includes parameters for one or more nodes within the domain for the actionable intent, and at least some of the parameters are populated with the specific information and requirements specified in the user request. For example, the user may say “Make me a dinner reservation at a sushi place at 7.” In this case, the natural language processor 332 may be able to correctly identify the actionable intent to be “restaurant reservation” based on the user input. According to the ontology, a structured query for a “restaurant reservation” domain may include parameters such as {Cuisine}, {Time}, {Date}, {Party Size}, and the like. In some implementations, based on the information contained in the user's utterance, the natural language processor 332 generates a partial structured query for the restaurant reservation domain, where the partial structured query includes the parameters {Cuisine=“Sushi”} and {Time=“7 pm”}. However, in this example, the user's utterance contains insufficient information to complete the structured query associated with the domain. Therefore, other necessary parameters such as {Party Size} and {Date} are not specified in the structured query based on the information currently available. In some implementations, the natural language processor 332 populates some parameters of the structured query with received context information. For example, in some implementations, if the user requested a sushi restaurant “near me,” the natural language processor 332 populates a {location} parameter in the structured query with GPS coordinates from the user device 104.

[0103] In some implementations, the natural language processor 332 passes the structured query (including any completed parameters) to the task flow processing module 336 (“task flow processor”). The task flow processor 336 is configured to receive the structured query from the natural language processor 332, complete the structured query, if necessary, and perform the actions required to “complete” the user's ultimate request. In some implementations, the various procedures necessary to complete these tasks are provided in task flow models 354. In some implementations, the task flow models include procedures for obtaining additional information from the user, and task flows for performing actions associated with the actionable intent.

[0104] As described above, in order to complete a structured query, the task flow processor 336 may need to initiate additional dialogue with the user in order to obtain additional information, and/or disambiguate potentially ambiguous utterances. When such interactions are necessary, the task flow processor 336 invokes the dialogue processing module 334 (“dialogue processor 334”) to engage in a dialogue with the user. In some implementations, the dialogue processor 334 determines how (and/or when) to ask the user for the additional information, and receives and processes the user responses. The questions are provided to and answers are received from the users through the I/O processing module 328. In some implementations, the dialogue processor 334 presents dialogue output to the user via audio and/or visual output, and receives input from the user via spoken or physical (e.g., clicking) responses. Continuing with the example above, when the task flow processor 336 invokes the dialogue flow processor 334 to determine the “party size” and “date” information for the structured query associated with the domain “restaurant reservation,” the dialogue flow processor 335 generates questions such as “For how many people?” and “On which day?” to pass to the user. Once answers are received from the user, the dialogue flow processor 334 can then populate the structured query with the missing information, or pass the information to the task flow processor 336 to complete the missing information from the structured query.

[0105] In some cases, the task flow processor 336 may receive a structured query that has one or more ambiguous properties. For example, a structured query for the “send a message” domain may indicate that the intended recipient is “Bob,” and the user may have multiple contacts named “Bob.” The task flow processor 336 will request that the dialogue processor 334 disambiguate this property of the structured query. In turn, the dialogue processor 334 may ask the user “Which Bob?”, and display (or read) a list of contacts named “Bob” from which the user may choose.

[0106] Once the task flow processor 336 has completed the structured query for an actionable intent, the task flow processor 336 proceeds to perform the ultimate task associated with the actionable intent. Accordingly, the task flow processor 336 executes the steps and instructions in the task flow model according to the specific parameters contained in the structured query. For example, the task flow model for the actionable intent of “restaurant reservation” may include steps and instructions for contacting a restaurant and actually requesting a reservation for a particular party size at a particular time. For example, using a structured query such as: {restaurant reservation, restaurant=ABC Café, date=3/12/2012, time=7 pm, party size=5}, the task flow processor 336 may perform the steps of: (1) logging onto a server of the ABC Café or a restaurant reservation system such as OPENTABLE®, (2) entering the date, time, and party size information in a form on the website, (3) submitting the form, and (4) making a calendar entry for the reservation in the user's calendar.

[0107] In some implementations, the task flow processor 336 employs the assistance of a service processing module 338 (“service processor”) to complete a task requested in the user input or to provide an informational answer requested in the user input. For example, the service processor 338 can act on behalf of the task flow processor 336 to make a phone call, set a calendar entry, invoke a map search, invoke or interact with other user applications installed on the user device, and invoke or interact with third party services (e.g. a restaurant reservation portal, a social networking website, a banking portal, etc.). In some implementations, the protocols and application programming interfaces (API) required by each service can be specified by a respective service model among the services models 356. The service processor 338 accesses the appropriate service model for a service and generates requests for the service in accordance with the protocols and APIs required by the service according to the service model.

[0108] For example, if a restaurant has enabled an online reservation service, the restaurant can submit a service model specifying the necessary parameters for making a reservation and the APIs for communicating the values of the necessary parameter to the online reservation service. When requested by the task flow processor 336, the service processor 338 can establish a network connection with the online reservation service using the web address stored in the service model, and send the necessary parameters of the reservation (e.g., time, date, party size) to the online reservation interface in a format according to the API of the online reservation service.

[0109] In some implementations, the natural language processor 332, dialogue processor 334, and task flow processor 336 are used collectively and iteratively to infer and define the user's intent, obtain information to further clarify and refine the user intent, and finally generate a response (i.e., an output to the user, or the completion of a task) to fulfill the user's intent.

[0110] In some implementations, after all of the tasks needed to fulfill the user's request have been performed, the digital assistant 326 formulates a confirmation response, and sends the response back to the user through the I/O processing module 328. If the user request seeks an informational answer, the confirmation response presents the requested information to the user. In some implementations, the digital assistant also requests the user to indicate whether the user is satisfied with the response produced by the digital assistant 326.

[0111] The error detection module 339 detects errors in interactions between a user and the digital assistant. In some implementations, to detect errors, the error detection module 339 monitors interactions between a user and the digital assistant, and/or between a user and a user device. For example, the error detection module 339 monitors any of the following types of interactions, or a subset thereof: a user's speech inputs to the digital assistant (e.g., if a user says “you got that wrong” or “you are pronouncing that wrong”), button presses (e.g., if a user selects a lock-screen or “home” button (or any other affordance) to cancel an action), movements of the device (e.g., shaking the device, setting the device down in a certain orientation, such as screen-down), termination of actions or suggested actions on the user device (e.g., cancelling a telephone call, email, text message, etc. after the digital assistant initiates or suggests it), initiation of an action shortly after a digital assistant fails to successfully infer an intent or adequately respond to a user, etc. In some implementations, the error detection module 339 monitors other types of interactions to detect errors as well.

[0112] In order to detect such errors, in some implementations, the error detection module 339 communicates with or otherwise receives information from various modules and components of the digital assistant system 300 and/or the user device 104, such as the I/O processing module 328 (and/or the I/O devices 316), the STT processing module 330, natural language processing module 332, the dialogue flow processing module 334, the task flow processing module 336, the service processing module 338, the phone module 260, the sensor processing module 258, the I/O subsystem 240, and/or any of the sensors or I/O devices associated therewith.

[0113] More details on the digital assistant can be found in the U.S. Utility application Ser. No. 12/987,982, entitled “Intelligent Automated Assistant”, filed Jan. 10, 2011, U.S. Utility Application No. 61/493,201, entitled “Generating and Processing Data Items That Represent Tasks to Perform”, filed Jun. 3, 2011, the entire disclosures of which are incorporated herein by reference.

[0114] In most scenarios, when the digital assistant receives a user input from a user, the digital assistant attempts to provide an appropriate response to the user input with as little delay as possible. For example, suppose the user requests certain information (e.g., current traffic information) by providing a speech input (e.g., “How does the traffic look right now?”). Right after the digital assistant receives and processes the speech input, the digital assistant optionally provides a speech output (e.g., “Looking up traffic information . . . ”) acknowledging receipt of the user request. After the digital assistant obtains the requested information in response to the user request, the digital assistant proceeds to provide the requested information to the user without further delay. For example, in response to the user's traffic information request, the digital assistant may provide a series of one or more discrete speech outputs separated by brief pauses (e.g., “There are 2 accidents on the road. <Pause> One accident is on 101 north bound near Whipple Avenue. <Pause> And a second accident is on 85 north near 280.”), immediately after the speech outputs are generated.

[0115] For the purpose of this specification, the initial acknowledgement of the user request and the series of one or more discrete speech outputs provided in response to the user request are all considered sub-responses of a complete response to the user request. In other words, the digital assistant initiates an information provision process for the user request upon receipt of the user request, and during the information provision process, the digital assistant prepares and provides each sub-response of the complete response to the user request without requiring further prompts from the user.

[0116] Sometimes, additional information or clarification (e.g., route information) is required before the requested information can be obtained. In such scenarios, the digital assistant outputs a question (e.g., “Where are you going?”) to the user asking for the additional information or clarification. In some implementations, the question provided by the digital assistant is considered a complete response to the user request because the digital assistant will not take further actions or provide any additional response to the user request until a new input is received from the user. In some implementations, once the user provides the additional information or clarification, the digital assistant initiates a new information provision process for a “new” user request established based on the original user request and the additional user input.

[0117] In some implementations, the digital assistant initiates a new information provision process upon receipt of each new user input, and each existing information provision process terminates either (1) when all of the sub-responses of a complete response to the user request have been provided to the user or (2) when the digital assistant provides a request for additional information or clarification to the user regarding a previous user request that started the existing information provision process.

[0118] In general, after a user request for information or performance of a task is received by the digital assistant, it is desirable that the digital assistant provides a response (e.g., either an output containing the requested information, an acknowledgement of a requested task, or an output to request a clarification) as promptly as possible. Real-time responsiveness of the digital assistant is one of the key factors in evaluating performance of the digital assistant. In such cases, a response is prepared as quickly as possible, and a default delivery time for the response is a time immediately after the response is prepared.

[0119] Sometimes, however, after an initial sub-response provided immediately after receipt of the user input, the digital assistant provides the remaining one or more sub-responses one at a time over an extended period of time. In some implementations, the information provision process for a user request is stretched out over an extended period of time that is longer than the sum of the time required to provide each sub-response individually. For example, in some implementations, short pauses (i.e., brief periods of silence) are inserted between an adjacent pair of sub-responses (e.g., a pair of consecutive speech outputs) when they are delivered to the user through an audio-output channel.

[0120] In some implementations, a sub-response is held in abeyance after it is prepared and is delivered only when a predetermined condition has been met. In some implementations, the predetermined condition is met when a predetermined trigger time has been reached according to a system clock and/or when a predetermined trigger event has occurred. For example, if the user says to the digital assistant “set me a timer for 5 minutes,” the digital assistant initiates an information provision process upon receipt of the user request. During the information provision process, the digital assistant provides a first sub-response (e.g., “OK, timer started.”) right away, and does not provide a second and final sub-response (e.g., “OK, five minutes are up”) until 5 minutes later. In such cases, the default delivery time for the first sub-response is a time immediately after the first sub-response is prepared, and the default delivery time for the second, final sub-response is a time immediately after the occurrence of the trigger event (e.g., the elapse of 5 minutes from the start of the timer). The information provision process is terminated when the digital assistant finishes providing the final sub-response to the user. In various implementations, the second sub-response is prepared any time (e.g., right after the first sub-response is prepared, or until shortly before the default delivery time for the second sub-response) before the default delivery time for the second sub-response.

[0121] FIG. 4 illustrates a technique for processing a token sequence using a natural language processor, according to some implementations. In some implementations, the technique described with reference to FIG. 4 uses and/or relates to methods 700 and 800 for determining a user intent from a text string, described below with reference to FIGS. 7-8.

[0122] FIG. 4 illustrates a speech input 400, corresponding to the speech input “show times for Argo,” undergoing speech-to-text (STT) processing. (In some implementations, the STT processing is performed by the STT processing module 330.) The results 404 of the STT processing include a plurality of candidate text strings 406 each associated with a speech recognition confidence score 408. The speech recognition score represents a confidence that the candidate text string is a correct transcription of the speech input. As shown in the results 404, the first candidate text string is associated with a 90% recognition score, even though it is incorrect (e.g., because the STT processor did not recognize the significance of the word “Argo” as a movie title), and has given the second candidate text string a lower recognition score (66%), even though it is the correct transcription. If the STT processor simply provided its best-guess text string, it is possible that subsequent natural language processing would be unable to determine the user's intent.

[0123] As shown in FIG. 4, however, each candidate text string 406 in the results 404 (which may be only a subset of the results of the STT processing) are provided to a natural language processor for natural language processing. (In some implementations, the natural language processor is the natural language processing module 332.) The candidate text strings 406 are processed by the natural language processor to determine a respective candidate domain 414 for each respective candidate text string 406. (E.g., the natural language processor determines a respective intent for each respective candidate text string, independent of the other text strings.)

[0124] The natural language processor also determines an intent deduction confidence score 416 for each of the candidate text strings 406. The intent deduction confidence scores 416 represent a confidence of the natural language processor in the selection of that domain for that text string. For example, the text string “shoe ties for arco” may implicate a “shopping” domain because of the words “shoe” and “ties,” but the intent deduction confidence score is low (e.g., 5%) because the natural language processor cannot identify any particular task from the text string. (If, however, the text string included additional words like “where to buy” or “get prices for,” the intent deduction confidence score for that text string would be higher.) In some implementations, the intent deduction confidence scores are determined independently from each other.

[0125] In some implementations, the intent deduction confidence scores 416 are based, at least partially, on the amount and/or quality of word matches between a text string and a domain in the ontology (e.g., the absolute number or the percentage of words in the text string that are also associated with the domain). For example, if most of the words in a text string are associated with a single domain, the intent deduction confidence score for that domain would he high. As a specific example, most of the words and/or phrases in the text string “show times for the movie Argo” would be found in a movie domain (e.g., “show times,” “movie,” “Argo”), so the confidence score for this domain would be high.

[0126] In some implementations, the intent deduction confidence scores 416 are based, at least partially, on whether the text string includes one or more named entities associated with the domain (e.g., business names, movie titles, restaurant names, contact names, etc.), as discussed below.

[0127] In some implementations, the particular domain determined for a text string does not affect the intent deduction confidence score (i.e., the intent deduction confidence scores do not reflect any preference or ranking for particular domains). For example, the fact that the domain for “shoe ties for arco” relates to a shopping domain does not affect the intent deduction confidence score. Rather, the (low) score is based on the fact that there are only two words that match the shopping domain, and that no particular task within the shopping domain can be determined from the text string.

[0128] Returning to FIG. 4, the natural language processor determines that the first candidate text string “show times for our go” corresponds to a “movie” domain, for example, because the words “show times” match words that are particularly salient to the “movie” domain. However, because it cannot determine the movie for which the user is requesting show times (e.g., because “our go” does not match any known movies), the intent deduction confidence score is relatively low (40%). On the other hand, the second candidate text string “show times for Argo” also corresponds to a movie domain, but has a high intent deduction confidence score (99%), for example, because the natural language processor recognized the word “Argo” as a known movie name.

[0129] Accordingly, the natural language processor can select the domain and/or task determined from the second candidate text string, even though it was ranked lower by the STT processor, because it has the highest intent deduction score 416. In some implementations, the natural language processor selects the domain and/or task determined from the candidate text string with the highest intent deduction score.

[0130] In some implementations, the natural language processor selects the domain and/or task based on a combination of the intent deduction score 416 and the recognition score 408. For example, in some implementations, the natural language processor produces an average score 418 representing an arithmetic mean value of a respective recognition score and a respective intent deduction score 416. In some implementations, the scores are combined using different mathematical operations and/or algorithms (e.g., a weighted average, a sum, etc.).

[0131] The scales and/or value ranges used to illustrate recognition and intent deduction scores in FIG. 4 are merely exemplary, any appropriate scale and/or value range may be used instead or in addition to those shown above.

[0132] FIG. 5 illustrates a technique for processing a token sequence using a natural language processor, according to some implementations. In some implementations, the technique described with reference to FIG. 5 uses and/or relates to a method 600 for determining a user intent from a text string, described below with reference to FIG. 6.

[0133] FIG. 5 illustrates a speech input 500, corresponding to the speech input “show times for Argo,” undergoing speech-to-text (STT) processing. (In some implementations, the STT processing is performed by the STT processing module 330.) The result 504 of the STT processing includes one or more text strings 504, including, for example, “show times for our go.” In some implementations, the one or more text strings 504 correspond to one or more best-guess transcriptions of the speech input 500 by the STT processor. In some implementations, the result 504 is the top best-guess transcription of the speech input 500 by the STT processor.

[0134] The one or more text strings 504 are provided to a natural language processor for initial natural language processing. (In some implementations, the natural language processor is the natural language processing module 332.) As shown in FIG. 5, the initial natural language processing results in a determination of a domain 514 for the one or more text strings, but is unable to determine a specific task 516 that the user is requesting. For example, the natural language processor may be able to determine the domain because of the high likelihood that the terms “show times for” relate to a movie domain, but unable to determine a specific task because it cannot identify any movie that matches the words “our go.” This error may be caused, for example, because the natural language processor performs a word-matching process to match words in the text string to words in an ontology (e.g., in order to determine a domain and/or an actionable intent for the text string), and the word-matching process did not return any salient results for the words “our go,” or did not find any words in the text string that matched a known movie title (e.g., because “our go” did not match any results in a list of known movie titles).

[0135] Accordingly, as shown in FIG. 5, the results of the initial natural language processing are provided to the natural language processor to undergo supplemental natural language processing. In some implementations, the supplemental natural Language processing uses a different word-matching process than the initial natural language processing. For example, in some implementations, the supplemental natural language processing uses a more relaxed string matching criteria than the first processing, such as an approximate string match or a phonetic match (as opposed to an exact match criteria of the initial natural Language processing). Thus, the natural language processor is more likely to determine that the words “our go” are similar enough to the movie title “Argo” that the speech input 500 most likely included the latter.

[0136] In some implementations, for the supplemental natural language processing, the natural language processor limits its word-matching to the domain 514 identified by the initial natural language processing (and/or word lists that are specific to that domain). For example, natural language processor will search for words within the movie domain and/or words lists associated with the movie domain that match (or are substantially similar to) words in the text string 504, rather than re-searching multiple domains of the ontology. Because the search space is limited to a single domain, the natural language processor can apply the relaxed criteria without returning too many potentially incorrect and/or irrelevant results (as may occur if the relaxed criteria were applied to a search of the entire ontology).

[0137] In some implementations, the natural language processor only attempts to match a portion of the text string 504. For example, the natural language processor may determine that the words “our go” are in a position in the text string that would typically include a movie title, and process only those words in a supplemental natural language processing phase.

[0138] As shown in FIG. 5, the results 512 of the supplemental natural language processing 512 include a task “retrieve show times for Argo,” indicating that the supplemental processing successfully identified a task for the text string.

[0139] As noted above, in some implementations, the results 504 of the STT processing include only one text string, such as a text string corresponding to the best-guess transcription by an STT processor. However, in some implementations, the results 504 include multiple text strings. In some implementations, the initial natural language processing includes one or more of the techniques described above with respect to FIG. 4. For example, in some implementations, each respective text string of the results 504 is processed to determine a respective domain for the text strings. In some implementations, each domain determination is associated with a speech recognition score and/or an intent deduction confidence score, as discussed above. In some implementations, the natural language processor selects one of the candidate text strings for which to perform supplemental natural language processing (e.g., the candidate text string with the highest intent deduction confidence score or average score). Applying supplemental natural language processing to a single candidate text string is discussed above.

[0140] In some implementations, however, the natural language processor selects more than one candidate text string for which to perform supplemental natural language processing. In some implementations, all of the candidate text strings are selected. In some implementations, a subset of the candidate text strings is selected. For example, in some implementations, the natural language processor selects: the n-best candidate text strings (e.g., as judged by the intent deduction confidence score or average score); the n-best candidate text strings having a common domain (e.g., the n-best candidate text strings for which a “movie” domain was determined); or all (or some) of the text strings that correspond to a domain having the most candidate text strings (e.g., if the “movie” domain is the most commonly determined domain among the candidate text strings, all of the text strings for which the “movie” domain was determined are selected).

[0141] As shown in FIG. 5, a “movie” domain is determined for the text strings “show times for our go” and “show times for are go.” Thus, the natural Language processor performs supplemental natural language processing for these two text strings, and ignores the others (e.g., based on their lower speech recognition and/or intent deduction scores, etc.).

[0142] As described above, the supplemental natural language processing applies relaxed string matching criteria (e.g., approximate string matching instead of exact string matching). Applying supplemental natural language processing to multiple candidate text strings increases the likelihood that a satisfactory domain will be identified, especially where none of the candidate text strings accurately reflect the speech input. For example, the text strings “show times for our go” and “show times for are go” each include a different incorrect transcription of the movie title “Argo.” The supplemental natural language processing may be able to determine a match to the correct movie title from one of these but not the other. For this specific example, the words “are go” may be a closer phonetic and approximate string match to the movie title “Argo” than “our go.” Accordingly, applying supplemental natural language processing to multiple text strings increases the chances that a domain and/or an actionable intent will be determined.

[0143] In some implementations, if the natural language processor cannot determine a domain and/or an actionable intent even after the techniques in FIGS. 4 and/or 5 are applied, the digital assistant requests additional information from the user, such as by engaging in a dialog with the user, as described above. Continuing the examples from above, the digital assistant will ask a user “what movie are you searching for?”, and may ask the user to type the movie title in order to bypass the speech-to-text process and ensure that the natural language processor has an accurate text string to process. Techniques for engaging in a dialog with a user (e.g., to acquire additional information necessary to complete a task) are described above.

[0144] FIG. 6 is a flow diagram of an exemplary method 600 implemented by a digital assistant for determining a user intent from a text string. In some implementations, the method 600 is performed at an electronic device with one or more processors and memory storing one or more programs for execution by the one or more processors. For example, in some implementations, the method 600 is performed at the user device 104 and/or the server system 108. In some implementations, the method 600 is performed by the digital assistant system 300 (FIG. 3A), which, as noted above, may be implemented on a standalone computer system (e.g., either the user device 104 or the server system 108) or distributed across multiple computers (e.g., the user device 104, the server system 108, and/or additional or alternative devices or systems). While the following discussion describes the method 600 as being performed by a digital assistant (e.g., the digital assistant system 300), the method is not limited to performance by any particular device or combination of devices. Moreover, the individual steps of the method may be distributed among the one or more computers, systems, or devices in any appropriate manner.

[0145] The digital assistant provides a text string corresponding to a speech input, wherein the text string comprises a first portion and a second portion (603). In some implementations, the digital assistant generates the text string with a speech-to-text processor (e.g., the STT processing module 330) (602). In some implementations, the digital assistant receives the speech input (601) (e.g., via the microphone 230 of the user device 104).

[0146] In some implementations, the first portion of the text string precedes the second portion of the text string, while in other implementations, the second portion precedes the first. In some implementations, the first and the second portions of the text string comprise a plurality of words.

[0147] The digital assistant determines a domain of the text string using natural language processing by applying a first word-matching process to at least the first portion of the text string (604). In some implementations, the first word-matching process requires exact word matches (e.g., all characters must match exactly). in some implementations, determining the domain of the text string includes processing the text string using a natural language processor that includes an ontology, where the natural language processor (e.g., the natural language processing module 332) determines what nodes in the ontology are implicated by the words in the text string, as described above with reference to FIGS. 3A-3C.

[0148] The digital assistant determines whether the second portion of the text string matches at least one word of a set of words associated with the domain by applying a second word-matching process to the second portion of the text string (606). In some implementations, word-matching criteria applied by the second word-matching process are more relaxed than word-matching criteria applied by the first word-matching process (608). For example, in some implementations, the first word-matching process requires exact matches between words (as described above), and the second word-matching process does not require exact matches between words.

[0149] In some implementations, the second word-matching process applies approximate string matching techniques (610). Continuing the example from FIGS. 4-5, the second word-matching process can determine that the words “are go” likely match the word “Argo,” because these text strings are very similar to each other.

[0150] In some implementations, the second word-matching process comprises determining an edit distance between a word of the second portion of the text string and a word of the set of words associated with the domain (612). In some implementations, the edit distance is a Levenshtein distance.

[0151] In some implementations, the second word-matching process applies phonetic matching techniques (614). Specifically, the natural language processing module 332 can determine whether words in a text string sound like any of the words associated with an ontology, rather than whether words in the text string are spelled the same as (or similar to) the words associated with the ontology. In some implementations, the natural language processing module 332 determines a set of phonemes for one or more words of the text string, and attempts to match the phonemes (e.g., using exact or approximate matching techniques) to words associated with the ontology. Accordingly, continuing the example from FIGS. 4-5, the second word-matching process can determine that the words “are go” likely match the word “Argo” because they can be mapped to the same or substantially the same phonemes (e.g., “ar-go”).

[0152] Upon determining that the second portion of the text string matches at least one word of the set of words, the digital assistant determines a user intent from the text string based at least in part on the domain and the at least one word of the set of words (615). In some implementations, determining the user intent includes determining an actionable intent, i.e., a task that the digital assistant is capable of performing. Examples of tasks that the digital assistant can perform are described above. For example, in some implementations, the digital assistant can make (and/or facilitate the making of) a restaurant reservation, set a reminder, provide driving directions to a user, etc. Continuing the example from FIGS. 4-5, the digital assistant can determine the user's intent as a request to receive show times for the movie “Argo.”

[0153] In some implementations, the digital assistant performs one or more actions to satisfy the user intent (616). Continuing once again the example from FIGS. 4-5, the digital assistant can retrieve show times for the movie “Argo,” and present them to the user (e.g., on a display device, or via speech output). Other examples of actions that the digital assistant can perform to satisfy a user intent are described above.

[0154] FIG. 7 is a flow diagram of an exemplary method 700 implemented by a digital assistant for determining a user intent from a text string. In some implementations, the method 700 is performed at an electronic device with one or more processors and memory storing one or more programs for execution by the one or more processors. For example, in some implementations, the method 700 is performed at the user device 104 or the server system 108. In some implementations, the method 700 is performed by the digital assistant system 300 (FIG. 3A), which, as noted above, may be implemented on a standalone computer system (e.g., either the user device 104 or the server system 108) or distributed across multiple computers (e.g., the user device 104, the server system 108, and/or additional or alternative devices or systems). While the following discussion describes the method 700 as being performed by a digital assistant (e.g., the digital assistant system 300), the method is not limited to performance by any particular device or combination of devices. Moreover, the individual steps of the method may be distributed among the one or more computers, systems, or devices in any appropriate manner.

[0155] The digital assistant provides a plurality of candidate text strings for a speech input, wherein the candidate text strings are generated by a speech recognition process (703). In some implementations, the speech recognition process is performed by the STT processing module 330. In some implementations, the digital assistant generates the candidate text strings with a speech-to-text processor (e.g., the STT processing module 330) (702). In some implementations, the digital assistant receives the speech input (701) (e.g., via the microphone 230 of the user device 104).

[0156] In some implementations, the plurality of candidate text strings are associated with a plurality of speech recognition confidence scores (704). Speech recognition confidence scores represent a confidence that a candidate text string is a correct transcription of the speech input (e.g., as determined by the STT processing module 330). In some implementations, the speech recognition confidence score is based, at least partially, on a statistical model associated with one or both of an acoustic model and a language model applied by the STT processing module 330.

[0157] In some implementations, the plurality of candidate text strings correspond to the n-best text strings generated by the speech recognition process, as determined by the speech recognition confidence scores of the text strings generated by the speech recognition process. In some implementations, the plurality of candidate text strings correspond to the 5 best, 10 best, or any other appropriate number of text strings.

[0158] In some implementations, the plurality of candidate text strings are text strings generated by the speech recognition process that satisfy a predetermined speech recognition confidence threshold. For example, in some implementations, the plurality of candidate text strings includes any candidate text strings that have a speech recognition confidence score greater than 50%, 75%, 90%, or any other appropriate threshold. The speech recognition confidence scores are represented here as percentages, but may be represented by values of any appropriate metric, scale, and/or value range.

[0159] The digital assistant selects at least a subset of the plurality of candidate text strings, the subset including at least two candidate text strings (705). In some implementations, the digital assistant selects the top 2, 3, 4, or 5 (or any appropriate number) of the plurality of candidate text strings.

[0160] The digital assistant determines a plurality of user intents from the selected subset of the plurality of text strings (706). For example, the digital assistant determines a respective user intent for each respective one of the text strings selected at step (705). In some implementations, determining the user intent includes determining an actionable intent, i.e., a task that the digital assistant is capable of performing. Examples of actions that the digital assistant can perform are described above. For example, in some implementations, the digital assistant can make (and/or facilitate the making of) a restaurant reservation, set a reminder, provide driving directions to a user, etc.

[0161] In some implementations, the plurality of user intents are associated with a plurality of intent deduction confidence scores (707). Intent deduction confidence scores are discussed above in reference to FIG. 4.

[0162] The digital assistant selects a user intent from the plurality of user intents (708). In some implementations, the digital assistant selects the user intent based on the plurality of intent deduction confidence scores. For example, in some implementations, the digital assistant selects the user intent having the highest intent deduction confidence score as the selected user intent (710).

[0163] In some implementations, the digital assistant selects the user intent based on the speech recognition confidence scores and the intent deduction confidence scores (712).

[0164] In some implementations, the digital assistant determines a plurality of composite confidence scores for the plurality of user intents (714).

[0165] In some implementations, a composite confidence score is any of the following: a combination of a respective speech recognition confidence score and a respective intent deduction confidence score; a sum of a respective speech recognition confidence score and a respective intent deduction confidence score; an average of a respective speech recognition confidence score and a respective intent deduction confidence score; or a weighted average of a respective speech recognition confidence score and a respective intent deduction confidence score.

[0166] In some implementations, the user intent selected at step (708) is the user intent with the highest composite confidence score (715).

[0167] In some implementations, the digital assistant performs one or more actions to satisfy the user intent (716). Examples of actions that the digital assistant can perform to satisfy a user intent are described above.

[0168] FIG. 8 is a flow diagram of an exemplary method 800 implemented by a digital assistant for determining a user intent from a text string. In some implementations, the method 800 is performed at an electronic device with one or more processors and memory storing one or more programs for execution by the one or more processors. For example, in some implementations, the method 800 is performed at the user device 104 or the server system 108. In some implementations, the method 700 is performed by the digital assistant system 300 (FIG. 3A), which, as noted above, may be implemented on a standalone computer system (e.g., either the user device 104 or the server system 108) or distributed across multiple computers (e.g., the user device 104, the server system 108, and/or additional or alternative devices or systems). While the following discussion describes the method 800 as being performed by a digital assistant (e.g., the digital assistant system 300), the method is not limited to performance by any particular device or combination of devices. Moreover, the individual steps of the method may be distributed among the one or more computers, systems, or devices in any appropriate manner.

[0169] The digital assistant provides a plurality of text strings corresponding to a speech input, wherein the plurality of text strings is associated with a plurality of speech recognition confidence scores and a plurality of user intents, and wherein the plurality of user intents is associated with a plurality of intent deduction confidence scores (804). In some implementations, the digital assistant generates the candidate text strings with a speech-to-text processor (e.g., the STT processing module 330) (802). In some implementations, the digital assistant receives the speech input (801) (e.g., via the microphone 230 of the user device 104).

[0170] In some implementations, the text strings and the plurality of speech recognition confidence scores are generated by the STT processing module 330. As described above, speech recognition confidence scores represent a confidence that a candidate text string is a correct transcription of the speech input (e.g., as determined by the STT processing module 330). In some implementations, the speech recognition confidence score is based, at least partially, on a statistical model associated with one or both of an acoustic model and a language model applied by the STT processing module 330.

[0171] The digital assistant selects one of the plurality of user intents based on one or both of the speech recognition confidence scores and the intent deduction confidence scores (804). In some implementations, selecting one of the plurality of user intents includes selecting the user intent with the highest composite confidence score (806).

[0172] In some implementations, the digital assistant determines a plurality of composite confidence scores for the plurality of user intents (808). As described above, in some implementations, a composite confidence score is any of the following: a combination of a respective speech recognition confidence score and a respective intent deduction confidence score; a sum of a respective speech recognition confidence score and a respective intent deduction confidence score; an average of a respective speech recognition confidence score and a respective intent deduction confidence score; or a weighted average of a respective speech recognition confidence score and a respective intent deduction confidence score.

[0173] In some implementations, the digital assistant performs one or more actions to satisfy the user intent (810). Examples of actions that the digital assistant can perform to satisfy a user intent are described above.

[0174] The operations described above with reference to FIGS. 6-8 are, optionally, implemented by components depicted in FIG. 2 and/or FIG. 3. Similarly, it would be clear to a person having ordinary skill in the art how other processes can be implemented based on the components depicted in FIG. 2 and/or FIG. 3.

[0175] It should be understood that the particular order in which the operations have been described above is merely exemplary and is not intended to indicate that the described order is the only order in which the operations could be performed. One of ordinary skill in the art would recognize various ways to reorder the operations described herein.

[0176] The foregoing description, for purpose of explanation, has been described with reference to specific implementations. However, the illustrative discussions above are not intended to be exhaustive or to limit the invention to the precise forms disclosed. Many modifications and variations are possible in view of the above teachings. The implementations were chosen and described in order to best explain the principles of the invention and its practical applications, to thereby enable others skilled in the art to best utilize the invention and various implementations with various modifications as are suited to the particular use contemplated.