US 7,634,409 B2Grant
Dynamic speech sharpening
Issue Date:2009-12-15
•16 Claims
•4 Drawing Sheets
Abstract
An enhanced system for speech interpretation is provided. The system may include receiving a user verbalization and generating one or more preliminary interpretations of the verbalization by identifying one or more phonemes in the verbalization. An acoustic grammar may be used to map the phonemes to syllables or words, and the acoustic grammar may include one or more linking elements to reduce a search space associated with the grammar. The preliminary interpretations may be subject to various post-processing techniques to sharpen accuracy of the preliminary interpretation. A heuristic model may assign weights to various parameters based on a context, a user profile, or other domain knowledge. A probable interpretation may be identified based on a confidence score for each of a set of candidate interpretations generated by the heuristic model. The model may be augmented or updated based on various information associated with the interpretation of the verbalization.
Metadata
Assignee
- VoiceBox Technologies, Inc.
Inventors
- Robert A. Kennewick
- Min Ke
- Michael Tjalve
- Philippe Di Cristo
Application Information
Application Number:US 11/513,269
Filing Date:2006-08-31
Priority Date:2005-08-31
Art Unit:2626
Classifications
IPC:
G10L15/00
Patent Drawings (4 sheets)
Description
Cross Reference to Related Applications
[0001] This application claims the benefit of provisional U.S. Patent Application Ser. No. 60/712,412, entitled “Dynamic Speech Sharpening,” filed Aug. 31, 2005, and which is hereby incorporated by reference in its entirety.
[0002] This application is related to the following co-pending applications, which are hereby incorporated by reference in their entirety: “Systems and Methods for Responding to Natural Language Speech Utterance,” Ser. No. 10/452,147, filed Jun. 3, 2003, and published as US Patent Application Publication No. US 2004/0044516; “Mobile Systems and Methods for Responding to Natural Language Speech Utterance,” Ser. No. 10/618,633, filed Jul. 15, 2003, and published as US Patent Application Publication No. US 2004/0193420; “Systems and Methods for Responding to Natural Language Speech Utterance,” Ser. No. 11/197,504, filed Aug. 5, 2005; “System and Method of Supporting Adaptive Misrecognition in Conversational Speech,” Ser. No. 11/200,164, filed Aug. 10, 2005; and “Mobile Systems and Methods of Supporting Natural Language Human-Machine Interactions,” Ser. No. 11/212,693, filed Aug. 29, 2005.
Field of the Invention
[0003] The invention is related generally to automated speech interpretation, and in particular, to enhancing the accuracy and performance of speech interpretation engines.
Background of the Invention
[0004] The field of automated speech interpretation is in increasingly higher demand. One use of automated speech interpretation is to provide voice requests to electronic devices. This may enable a user to simply speak to an electronic device rather than manually inputting requests, or other information, through pressing buttons, uploading information, or by other request input methods. Controlling various electronic devices through speech may enable the user to use the electronic devices more efficiently.
[0005] However, existing technology in the field of automated speech interpretation, such as standard speech engines, automatic speech recognition (ASR), and other systems for interpreting speech, are unable to process a speech signal in an efficient manner, often constructing large grammars that include a large number of items, nodes, and transitions, which is a concern particularly for large-list recognition for embedded applications. If the grammar for an embedded application grows too much, it may not fit within the constrained space of an embedded application. With limited CPU power, response time and performance is easily affected due to the significant time needed to compile and load the grammar. Response time is further degraded because the speech engine has to parse through a large number of transition states to come up with a recognition result. Even when the speech engine is able recognize a word, the results are often unreliable because large grammars introduce greater risk of confusion between items as the size of the grammar increases. Existing techniques focus on reducing the size of a grammar tree by removing command variants or criteria items, but this approach strips functionality from the application.
[0006] In addition to the performance problems associated with speech recognition engines that employ large word grammars, existing speech processing engines are unable to interpret natural human speech with a suitable accuracy to sufficiently control some electronic devices. In particular, speech interpretation engines still have substantial problems with accuracy and interpreting words that are not defined in a predetermined vocabulary or grammar context. Poor quality microphones, extraneous noises, unclear or grammatically incorrect speech by the user, or an accent of the user may also cause shortcomings in accuracy, such as when a particular sound cannot be mapped to a word in the grammar.
[0007] In light of these and other problems, there is a need for enhanced automated speech interpretation that may interpret natural human speech with an augmented accuracy.
Summary of the Invention
[0008] According to one aspect of the invention, a system for enhancing automated speech interpretation is provided. The system may include a set of techniques for use in a speech-to-text engine to enhance accuracy and performance, for example, by reducing the search space of the speech engine. The problems with large-list recognition for embedded applications may also be improved by using phonetic dictation, which may recognize a phoneme string by disregarding the notion of words. The system may also use one or more post-processing techniques to sharpen an output of a preliminary speech interpretation made by a speech engine. The system may be modeled at least partially after one or more speech pattern recognition techniques used by humans, such as interpreting speech using words, word sequences, word combinations, word positions, context, phonetic similarities between two or more words, parts of speech, or other techniques.
[0009] In one implementation of the invention, the system may receive a verbalization made by a user, where a speech engine may receive the verbalization. The speech engine may output information relating to a plurality of preliminary interpretations of the verbalization, where the plurality of preliminary interpretations represent a set of best guesses at the user verbalization. According to one aspect of the invention, the performance of the speech engine may be improved by using phoneme recognition. Phoneme recognition may disregard the notion of words, instead interpreting a verbalization as a series of phonemes, which may provide out-of-vocabulary (OOV) capabilities, such as when a user misspeaks or an electronic capture devices drops part of a speech signal, or for large-list applications, such as city and street names or song titles, for example. Phoneme recognition may be based on any suitable acoustic grammar that maps a speech signal into a phonemic representation. For example, the English language may be broken down into a detailed grammar of the phonotactic rules of the English language. Portions of a word may be represented by a syllable, which may be further broken down into core components of an onset, a nucleus, and a coda, which may be further broken down into sub-categories. Various different acoustic grammars may be formed as trees with various branches representing many different syllables forming a speech signal.
[0010] According to another aspect of the invention, the performance of the speech engine and the phonemic recognition may be improved by pruning the search space used by the speech engine using a common phonetic marker. In one implementation, the acoustic grammar may be represented entirely by a loop of phonemes. In another implementation, the speech engine may reduce the search space by reducing the number of transitions in a grammar tree, thereby speeding up the process of compiling, loading, and executing the speech engine. For example, the phoneme loop may include a linking element between transitions. This may reduce the number of grammar transitions, such that grammar paths merge after a first transition and diverge after the linking element. In one implementation of the invention, a common acoustic element that is part of a speech signal may be used as the linking element. In one implementation of the invention, the acoustic element may be one that is very likely to be triggered even if it is unpronounced. For example, a schwa in the English language may be used as the linking element because schwa represents an unstressed, central vowel that is likely to be spoken even if unintended. Those skilled in the art will appreciate that acoustic models for different languages may use other frequently elided phonemes as linking elements to reduce the search space used by the speech engine.
[0011] The speech engine may generate a plurality of preliminary interpretations representing a set of best guesses at the user verbalization. The preliminary interpretations may be stored in a matrix, array, or another form, and may be provided to an interpretation sharpening module to determine a probable interpretation of a verbalization made by a user by applying heuristic policies against the preliminary interpretation to identify dominant words and/or phrases. According to various aspects of the invention, the interpretation sharpening module may include a policy module that may manage and/or provide one or more policies that enable the sharpening module to generate a plurality of probable interpretations of the verbalization made by the user. For example, according to one aspect of the invention, the plurality of preliminary interpretations may be applied against one or more policies to generate a set of hypotheses as to a candidate recognition. Each hypothesis may be reanalyzed to generate an interpretation score that may relate to a likelihood of the probable interpretation being a correct interpretation of the verbalization, and the preliminary interpretation corresponding to the highest (or lowest) interpretation score may then be designated as a probable interpretation of the verbalization. The designated probable interpretation may be stored and used for augmenting the policies to improve accuracy.
[0012] According to one aspect of the invention, the policy module may include one or more agents that represent domains of knowledge. The agents may compete using a weighted model to revise a preliminary interpretation by determining context and intent. Relevant substitution of suspect words and phrases may be based on phonetic similarities or domain appropriateness. A domain agent may include one or more domain parameters for determining a probable interpretation from a preliminary interpretation. For example, domain parameters may include a policy vocabulary, a word position in the verbalization, a word combination, a sentence structure, or other parameters. A domain agent may include a parameter weighting scheme that may weight individual parameters according to one or more weighting factors, such as, a frequency of use, a difficulty to understand, or other factors.
[0013] According to one aspect of the invention, the domain agents may revise a preliminary interpretation into a probable interpretation using phonetic fuzzy matching (PFM). In one implementation of the invention, the speech engine may output a phoneme stream that is applied against a model of phoneme feature similarities, drawn from domain agents, to identify a closest phonetic match using a multi-pass method. Domain agents may be loaded and prioritized into an M-Tree, which accounts for the possibility of the speech engine dropping or adding phonemes. An M-Tree may be an index structure that resolves similarity queries between phonemes using a closest-distance metric based on relative weightings of phoneme misrecognition, phoneme addition, and phoneme deletion. The M-Tree may be updated using an adaptive misrecognition model. For example, information about a verbalization and its components, as well as a probability that the probable interpretation was correct, may be stored and used for adapting the policy module for the user.
[0014] In one implementation of the invention, the domain agents in the policy module may include one or more profile agents that may manage and/or provide one or more profile policies for revising a preliminary interpretation of a phoneme stream. For example, a profile agent may correspond to a user and may include one or more profile parameters tailored to the user. The profile agent may be used as a base policy to interpret any verbalizations made by the user. In other implementations, a profile agent may correspond to a particular language, a regional accent, or other profiles for interpreting a user verbalization. The profile agents may be augmented to enable the system to provide more accurate interpretations of verbalizations made by the user. The augmentation may include a user augmentation, such as providing additional vocabulary (e.g., names in an address book), one or more personalized pronunciations or other pronunciation information, or other user provided augmentations. The augmentation may also include a non-user provided augmentation, such as updates generated by a third party (e.g., a commercial administration and/or maintenance entity), or other non-user provided augmentations. The augmentation may be automated, such as adjusting a profile parameter-weighting scheme through an adaptive misrecognition model, as discussed above.
[0015] In another implementation of the invention, the domain agents in the policy module may include one or more context agents that may manage and/or provide one or more context policies for revising a preliminary interpretation of a phoneme stream. For example, a context agent may correspond to a context, such as song titles, city and street names, movie titles, finance, or other contexts. A context agent may include one or more context parameters that may be tailored to a verbalization context. The context policy may enhance an ability of the system related to interpreting verbalizations made by the user in the verbalization context corresponding to the context agent. The context agents may be augmented to enable the system to provide more accurate interpretations of verbalizations made in a verbalization context corresponding to the context agent. The augmentation may include a user provided augmentation, a non-user provided augmentation, an automated augmentation, or other augmentations. The augmentation may be automated, such as adjusting a profile parameter-weighting scheme through an adaptive misrecognition model, as discussed above.
[0016] According to various implementations of the invention, the policy module may determine which profile agents and/or which context agents to use through a set of heuristics provided in a context-tracking module. In one implementation, the context-tracking module may use phonetic fuzzy matching to track a series of verbalizations by the user to identify a verbalization context. The context-tracking module may utilize one or more M-Trees to track the series of verbalizations and determine a closest phonetic match. The context-tracking module may track one or more past verbalization contexts for the series of verbalizations, one or more current verbalization contexts for the series of verbalizations, and/or make predictions regarding one or more future verbalization contexts for the series of verbalizations. The policy module may utilize information about the verbalization context of the series of verbalizations generated by the context tracking module to manage and/or provide one or more profile and/or context agents.
[0017] According to one aspect of the invention, the system may include an interpretation history analysis module that may enable the system to augment one or more domain agents based on an analysis of past interpretations related to previously interpreted verbalizations. The augmentations enabled by the interpretation history analysis module may include a user augmentation, a third-party augmentation, an automated augmentation, or other augmentations. The interpretation history analysis module may include an information storage module that may store interpretation information related to past verbalizations, such as one or more preliminary interpretations associated with a past verbalization, one or more interpretation scores associated with a past verbalization, one or more probable interpretations associated with a past verbalization, whether or not a past verbalization was interpreted correctly, or other information. A frequency module may be included in the interpretation history module, and the frequency module may use some or all of the information stored in the information storage module to generate one or more frequencies related to one or more past verbalizations. For example, the frequency module may calculate a frequency of word usage, word combinations, phonetic homonyms, interpretation errors for a particular verbalization, or other frequencies.
[0018] The Information generated and/or stored by the interpretation history analysis module may be used to augment the profile and/or context agents in order to enhance the accuracy of subsequent interpretations. In some implementations, an adaptive misrecognition model may use one or more generated frequencies to augment one or more agents. For example, one or more parameters and/or weighting schemes of an agent or phonetic model may be augmented based on a frequency generated by the interpretation history analysis module. Other augmentations using information stored and/or generated by the interpretation history analysis module may be made, and the system may include a policy agent handler that may augment, update, remove, and/or provide one or more domain agents to the system. A domain agent may comprise a profile or context agent, and the policy agent handler may be controlled, directly or indirectly by a third party (e.g. a commercial entity, etc.). The policy agent handler may augment, update, remove, and/or provide domain agents to the system as part of a commercial agreement, such as a licensing agreement, a subscription agreement, a maintenance agreement, or other agreements.
[0019] Other objects and advantages of the invention will be apparent to those skilled in the art based on the following detailed description and accompanying drawings.
Brief Description of the Drawings
[0020] FIG. 1 illustrates an exemplary system for enhancing automated speech interpretation according to one implementation of the invention.
[0021] FIG. 2 illustrates an exemplary grammar tree for enhancing the performance of a speech engine according to one implementation of the invention.
[0022] FIG. 3 illustrates an exemplary flow chart of a method for enhancing automated speech interpretation according to one implementation of the invention.
Detailed Description
[0023] FIG. 1 illustrates an exemplary system for enhancing automated speech interpretation according to one implementation of the invention. A speech-to-text processing engine 112 may receive a user verbalization, and speech engine 112 may generate one or more preliminary interpretations of the user verbalization. The preliminary interpretations may represent a set of best guesses as to the user verbalization arranged in any predetermined form or data structure, such as an array, a matrix, or other forms. In one implementation of the invention, speech engine 112 may generate the preliminary interpretations by performing phonetic dictation to recognize a stream of phonemes, instead of a stream of words. Phonemic recognition provides several benefits, particularly in the embedded space, such as offering out-of-vocabulary (OOV) capabilities, improving processing performance by reducing the size of a grammar, and eliminating the need to train Statistic Language Models (SLMs). Those skilled in the art will recognize other advantages of phonemic recognition.
[0024] Speech engine 112 may apply the phoneme stream against one or more acoustic grammars that reliably map a speech signal to a phonemic representation in order to generate the plurality of preliminary interpretations. Characteristics of a speech signal may be mapped to a phonemic representation to construct a suitable acoustic grammar, and various acoustic grammars may be included in speech engine 112 to generate one or more preliminary interpretations according to the various acoustic grammars. For example, the English language may be mapped into a detailed acoustic grammar representing the phonotactic rules of English, where words may be divided into syllables, which may further be divided into core components of an onset, a nucleus, and a coda, which may be further broken down into one or more sub-categories.
[0025] Once the phonotactic rules of a speech signal have been identified, a detailed acoustic grammar tree may be constructed that accounts for the nuances of the speech signal. The acoustic grammar may include a loop of phonemes, or the phoneme loop may include a linking element to reduce a size of a search space associated with the grammar. Using the English language as an example, the grammar tree may include various branches representing English language syllables. The speech engine may traverse one or more grammar trees to generate one or more preliminary interpretations of a phoneme stream as a series of syllables that map to a word or phrase. By using phonemic recognition rather than word recognition, the size of the grammar can be reduced, which reduces the amount of time required to compile, load, and execute speech interpretation. Moreover, because the grammar maintains a high level of phonotactic constraints and therefore a large number of syllables, speech engine 112 may be very precise in generating phonemic representations of human verbalizations.
[0026] An acoustic grammar used by speech engine 112 may be further optimized to reduce compile time, load time, and execution time by reducing the size of a search space associated with the acoustic grammar. Referring now to FIG. 2 , a traditional grammar tree 120 is compared to an exemplary grammar tree according to one aspect of the invention to demonstrate the performance enhancements of speech engine 112. In traditional speech processing engines, nodes in a grammar tree 210 tend to represent words, or large-list applications may be supported provided through a grammar tree 210 where the nodes represent items in the large-list. This requires the speech engine to parse through a large number of transition states to come up with a recognition result, which degrades response time. An example of this is seen in the following grammar structure:
[0027] “<street name> <city name>”→e.g., “NE 24th Street Bellevue”
[0028] In the above example, a large list of street names is followed by a large list of city names. Assuming three elements in the list of street names, and three elements in the list of city names, this results in twenty-one transitions, which may be represented by traditional grammar tree 210. Every end-node of the first list is followed by all entries in the second list, potentially leading to very large grammars because most real-world large-list applications are likely to include much more than three list items. For example, a city may have hundreds or thousands of street names, and there may be hundreds or thousands of city names. Moreover, every element in the second segment of traditional grammar tree 210 is repeated, once for each first segment, which introduces redundancy.
[0029] According to an aspect of the invention, the problems with traditional grammar trees may be resolved by using phonemic acoustic grammars instead of large-lists. The grammar may further be improved by including linking elements to reduce the number of transition states in the grammar. Thus, a grammar tree with a linking element 220 will merge after a first segment and then spread out again at a second segment, where the segments may represent a phoneme in an acoustic grammar, as discussed above. For example, assume a two-syllable word in an acoustic grammar consisting of three phonemes, which is able to reduce the number of transitions from twenty-one in a traditional grammar tree 210 to twelve in a grammar tree with a linking element 220. Two syllables and three phonemes are chosen to show the reduction in search space in a grammar tree with a linking element 220 as opposed to a corresponding traditional grammar tree 210, although a real-world acoustic grammar modeled after a language is likely to have a maximum of roughly fifty phonemes. Moreover, the search space may be further reduced by restricting available transitions based on phonotactic constraints for an acoustic model.
[0030] Using the approach described in FIG. 2 , adding a linking element to an acoustic grammar may reduce both grammar size and response time. Part of a speech signal may be mapped to the linking element in order to maintain the phonotactic rules of the acoustic grammar. The linking element may be an acoustic element that is likely to be triggered even if unpronounced. For example, a schwa represents an unstressed, central vowel in the English language (e.g., the first and last sound in the word “arena” is schwa). The phoneme schwa is an ideal linking element because of how it is represented in a frequency spectrum. That is, schwa is a brief sound and when a person opens their mouth to speak, there is a strong likelihood of passing through the frequencies of schwa even if unintended. Those skilled in the art will recognize that this approach may be extended to acoustic models of speech signals for other languages by using frequently elided phonemes as linking elements to reduce the search space of an acoustic grammar.
[0031] Referring again to FIG. 1 , speech engine 112 may receive a user verbalization and process the verbalization into a plurality of preliminary interpretations using the techniques described above. That is, the verbalization may be interpreted as a series of phonemes, and the series of phonemes may be mapped to one or more preliminary interpretations by traversing one or more acoustic grammars that are modeled after grammar 220 of FIG. 2 . The plurality of preliminary interpretations may take the form of words, parts of words, phrases, utterances, or a combination thereof, and the plurality of preliminary interpretations may be arranged as a matrix, an array, or in another form. The plurality of preliminary interpretations are then passed to a speech sharpening engine 110 for deducing a most probable interpretation of the verbalization.
[0032] According to various aspects of the invention, speech sharpening engine 110 may include an interpretation sharpening module 116, a policy module 114, an interpretation history analysis module 118, and a policy agent handler 120. The plurality of preliminary interpretations may be received by interpretation sharpening module 116, which forwards the preliminary interpretations to policy module 114 for further processing. Policy module 114 may include one or more context agents 126, one or more profile agents 128, and a context tracking module 130 that collectively revise the plurality of preliminary interpretations into a set of hypotheses that represent candidate recognitions of the verbalization. Policy module 114 may assign each hypothesis an interpretation score, and interpretation sharpening module 116 may designate the hypothesis with the highest (or lowest) interpretation score as a probable interpretation.
[0033] According to one aspect of the invention, policy module 114 may include one or more context agents 126. Context agents 126 may represent domains of knowledge corresponding to a given context, such as song titles, city and street names, finance, movies, or other contexts. Context agents 126 may use context objects and associated dynamic languages to represent a corresponding context. Policy module 114 may also include one or more profile agents 128. Profile agents 128 may represent domains of knowledge corresponding to a given profile, such as a specific user, language, accent, or other profiles. Profile agents 128 may use profile objects and dynamic languages to represent a corresponding profile. Dynamic languages for context agents 126 or profile agents 128 may specify vocabularies, word combinations, phrases, sentence structures, criteria, and priority weightings for any given context or profile, respectively. The priority weightings may weight individual parameters according to one or more weighting factors, such as assigning a weight according to a frequency of use, a difficulty to understand, or other factors. Policy module 114 may also include a context-tracking module 130. Context tracking module 130 may track a verbalization context of a consecutive series of verbalizations. Context tracking module 130 may utilize one or more conversation trees to track the series of verbalizations. Context tracking sub-module 214 may track one or more past or current verbalization contexts of the series of verbalizations, and/or may make predictions regarding one or more future verbalization contexts of the series of verbalizations. Policy module 114 may utilize information about the verbalization context, generated by context tracking module 130, to generate one or more sharpened interpretations and corresponding interpretation scores.
[0034] In some implementations, policy module 114 may use context tracking module 130 to apply objects from one or more context agents 126 and/or profile agents 128 to the preliminary interpretations provided by speech engine 112. The various agents may compete with each other using a set of heuristics in a phonetic fuzzy matcher, where an intent or context of the user may be identified based on the set of heuristics about how a request may be phrased in a given domain. A closest phonetic match may be identified for suspect words and/or phrases among the plurality of preliminary interpretations.
[0035] The phonetic fuzzy matcher may include an M-Tree that is populated with context objects, profile objects, and/or dynamic language data from one or more of context agents 126 and/or profile agents 128. M-Trees are known to those skilled in the art. The M-Tree may assign relative priority weights to the context objects, profile objects, and/or dynamic language data in order to account for the possibility of misrecognized phonemes, extraneous phonemes, or erroneously deleted phonemes. A closest distance metric associated with the M-Tree may be used given the relative weightings of phoneme misrecognition, phoneme addition, and phoneme deletion for various contexts and/or profiles.
[0036] According to one aspect of the invention, one or more passes may be taken over the plurality of preliminary interpretations to identify dominant words and/or phrases among the plurality of preliminary interpretations. Using the M-Tree weighted model, one or more candidate interpretations may be made based on relevant substitution of suspect words and/or phrases based on phonetic similarities and/or domain appropriateness. For example, if a set of dominant words appear to be a movie name, a candidate interpretation will substitute the relevant words and/or phrases to generate a candidate interpretation about movies. After a set of candidate interpretations have been generated, the candidate interpretations are analyzed using the M-Tree weighted model. With the relevant domains constrained by the candidate interpretations, a confidence or interpretation score may be assigned to each candidate interpretation, with the interpretation score representing a likelihood that a particular candidate interpretation is a correct interpretation of the verbalization. The candidate interpretations may then be returned to interpretation sharpening module 116, and interpretation sharpening module 116 may select a candidate interpretation with a highest (or lowest) interpretation score as a probable interpretation of the verbalization.
[0037] According to various implementations of the invention, speech sharpening engine 110 may include an interpretation history analysis module 118. Interpretation history analysis module 118 may include an information storage module 122 a frequency module 124. Information storage module 122 may store information related to verbalizations, including components of verbalizations, preliminary interpretations, dominant words and/or phrases, candidate interpretations, probable interpretations, and/or interpretation scores associated with verbalizations, as well as whether or not a verbalization was interpreted correctly, or other information. Interpretation history analysis module 118 may also include a frequency module 124. Frequency module 124 may use some or all of the information stored in information storage module 122 to generate one or more frequencies related to one or more past verbalizations. For example, frequency module 124 may calculate a word usage frequency, a word combination frequency, a frequency related to a set of verbalizations that are phonetically similar but have distinct meanings, an interpretation error frequency for a particular verbalization, or other frequencies.
[0038] Information stored and/or generated by interpretation history analysis module 118 may be used to augment speech sharpening engine 110. In some implementations, the information may be used to adjust various weights used in phonetic models, such as context agents 126 or profile agents 128, as well as adapting the relative weights in the M-Tree in context tracking module 130 to enhance accuracy for subsequent verbalizations. In another implementation, the stored information may be sent to a third-party or commercial entity for analyzing the data and developing new domain agents or further improving the accuracy of speech sharpening engine 110. For example, one or more parameters and/or weighting schemes of an agent may be augmented based on a frequency generated by interpretation history analysis module 118. Other augmentations related to information stored on and/or generated by interpretation history analysis module 118 may be made. Speech sharpening engine 110 may also include a policy agent handler 120 that may augment, update, remove, and/or provide one or more domain agents to policy module 114. A domain agent may include one or more new, modified, or updated context agents 126 and/or profile agents 128. Policy agent handler 120 may also augment or update the M-Tree in context tracking module 130 to adjustments in priority weighting schemes or phonetic models. Policy agent handler 120 may be controlled, directly or indirectly, by a third party, such as a commercial entity, and domain agents may be augmented, updated, removed, and/or provided by policy agent handler 120 as part of a commercial agreement, licensing agreement, subscription agreement, maintenance agreement, or other agreement.
[0039] Referring to FIG. 3 , a flow chart demonstrating an exemplary method for enhancing the performance and accuracy of speech interpretation is provided. The method may begin by receiving a user verbalization at an operation 312. The received user verbalization may be electronically captured at operation 312, such as by a microphone or other electronic audio capture device. The electronically captured verbalization may be provided to a speech interpretation engine, such as speech engine 112 in FIG. 1 .
[0040] The speech interpretation may then generate one or more preliminary interpretations of the received verbalization at an operation 314. According to one implementation of the invention, the plurality of preliminary interpretations may be generated using phonetic dictation, grammar trees with linking elements, or any combination thereof to improve performance and enhance accuracy. Phonetic dictation and reducing a search space of a grammar tree by including linking elements is discussed in greater detail above. The preliminary interpretations may be arranged in any predetermined form, such as an array, a matrix, or other forms.
[0041] In an operation 320, the preliminary interpretations may be provided to a speech sharpening engine. The speech sharpening engine may take one or more passes over the plurality of preliminary interpretations to identify dominant words and/or phrases in operation 320. This information may then be used to generate one or more candidate interpretations. The candidate interpretations may be based on various domain agents, such as context agents and/or profile agents, which may be organized as a weighted domain model, such as an M-Tree. For example, if a set of dominant words sound like a movie name, apply policies operation 320 may generate a candidate interpretation that substitutes relevant words and/or phrases based on a domain agent populated with movie titles. Additional passes may be made over the candidate interpretations, which may be constrained by domain information associated with the candidate interpretations, to thereby generate a confidence score or interpretation score for each candidate interpretation. The interpretation score may represent a likelihood that a particular candidate interpretation is a correct interpretation of the verbalization received in operation 312. The operation of apply policies 320 is described in greater detail above in reference to FIG. 1 .
[0042] The candidate interpretations and corresponding interpretation scores may then be analyzed to determine a probable interpretation in an operation 322. In one implementation of the invention, a candidate interpretation with a highest (or lowest) score may be designated as a probable interpretation. The probable interpretation may then be output in an operation 324, such as for use in a voice-activated vehicular navigation system, a voice-controlled server or desktop computer, or other electronic device that can be controlled using voice commands.
[0043] Information relating to the verbalization and the interpretations of the verbalization may be provided in a store interpretation operation 325. Store interpretation operation 324 may store interpretation information related to verbalizations, such as components of verbalizations, preliminary interpretations, dominant words and/or phrases, candidate interpretations, probable interpretations, and/or interpretation scores associated with verbalizations, as well as whether or not a verbalization was interpreted correctly, or other information. In some implementations of the invention, some or all of the interpretation information stored at store interpretation operation 324 may be used to determine one or more frequencies at a determine frequencies operation 326. The frequencies calculated at determine frequencies operation 326 may include one or more frequencies related to past verbalizations, such as, a word usage frequency, a word combination frequency, a frequency related to a set of verbalizations that are phonetically similar but have distinct meanings, an interpretation error frequency for a particular verbalization, or other frequencies. Determine frequencies operation 326 may be performed by interpretation history analysis module 118.
[0044] In various implementations, a decision may be made whether to augment a speech sharpening engine in an augmentation decision operation 328. The decision concerning system augmentation may be based at least in part on information generated at determine frequencies block 326, such as one or more frequencies, or other information. If it is decided that no augmentation is needed, no further action is taken until another verbalization is captured, and the method ends. In some instances, decision operation 328 may determine that augmentation may be made and control passes to an augment system operation 330. Augment system operation 330 may include making an augmentation to a speech sharpening engine. For example, one or more domain agents may be augmented to reflect probabilities of an interpretation being a correct interpretation of a verbalization, to update a user profile, or other augmentation. Dynamic languages associated with context agents and/or profile agents may be augmented, or parameters weights may be augmented to enhance accuracy when interpreting subsequent verbalizations. For example, an adaptive misrecognition technique may adjust the various weights in a phonetic model or update similarity weights for regional accents, or other augmentations may be made. In parallel to augment system operation 330, new agent policies may be received in an operation 332. For example, a third party or commercial entity may redesign or modify various domain agents, new domain agents may be developed and installed as plug-ins, domain agents that are unreliable may be removed, or other augmentations or modifications may be made. Thus, the method continually refines the domain agents and the weighting of various parameters in order to refine the accuracy of the speech sharpening engine for subsequent verbalizations.
[0045] The above disclosure has been described in terms of specific exemplary aspects, implementations, and embodiments of the invention. However, those skilled in the art will recognize various changes and modifications that may be made without departing from the scope and spirit of the invention. For example, references throughout the specification to “one implementation,” “one aspect,” “an implementation,” or “an aspect” may indicate that a particular feature, structure, or characteristic is included in at least one implementation. However, the particular features, structures, or characteristics may be combined in any suitable manner in one or more implementations. Therefore, the specification and drawings are to be regarded as exemplary only, and the scope of the invention is to be determined solely by the appended claims.
Claims
What is claimed is:
1. A method for providing out-of-vocabulary interpretation capabilities and for tolerating noise when interpreting natural language speech utterances, the method comprising:
receiving an utterance from a user;
recognizing a stream of phonemes contained in the utterance on an electronic device;
mapping the recognized stream of phonemes to an acoustic grammar that phonemically represents one or more syllables, the recognized stream of phonemes mapped to a series of one or more of the phonemically represented syllables; and
generating at least one interpretation of the utterance, wherein the generated interpretation includes the series of syllables mapped to the recognized stream of phonemes.
2. The method of claim 1, the acoustic grammar phonemically representing the one or more syllables in accordance with acoustic elements of an acoustic speech model, wherein each syllable is represented by acoustic elements for an onset, a nucleus, and a coda.
3. The method of claim 2, the acoustic grammar including transitions between the acoustic elements, wherein the transitions are constrained according to phonotactic rules of the acoustic speech model.
4. The method of claim 3, the acoustic elements including at least an unstressed central vowel and a plurality of phonemic elements associated with the acoustic speech model, wherein the acoustic grammar uses the unstressed central vowel as a linking element between sequential phonemic elements.
5. The method of claim 4, the unstressed central vowel including a schwa acoustic element.
6. The method of claim 1, further comprising:
generating a plurality of candidate interpretations of the utterance, wherein each candidate interpretation includes a series of words or phrases corresponding to the series of syllables mapped to the recognized stream of phonemes;
assigning a score to each of the plurality of candidate interpretations; and
selecting a candidate interpretation having a highest assigned score as being a probable interpretation of the utterance.
7. The method of claim 6, wherein generating the plurality of candidate interpretations of the utterance includes:
identifying a series of words or phrases possibly corresponding to the series of syllables mapped to the recognized stream of phonemes; and
providing the identified series of words or phrases to a plurality of domain agents, each of the plurality of domain agents generating a candidate interpretation by:
identifying any suspect words or phrases among the identified series of words or phrases;
identifying one or more closest phonetic matches for the identified suspect words or phrases using a closest-distance metric associated with an M-Tree; and
substituting the identified closest phonetic matches for the identified suspect words or phrases to generate the candidate interpretation.
8. The method of claim 7, wherein identifying the series of words or phrases possibly corresponding to the series of syllables includes using one or more of a context associated the utterance, criteria identified in the utterance, a history of previous utterances, user profile information, or domain specific information.
9. A system for providing out-of-vocabulary interpretation capabilities and for tolerating noise when interpreting natural language speech utterances, the system comprising:
at least one input device that receives an utterance from a user and generates an electronic signal corresponding to the utterance; and
a speech interpretation engine that receives the electronic signal corresponding to the utterance, the speech interpretation engine operable to:
recognize a stream of phonemes contained in the utterance;
map the recognized stream of phonemes to an acoustic grammar that phonemically represents one or more syllables, the recognized stream of phonemes mapped to a series of one or more of the phonemically represented syllables; and
generate at least one interpretation of the utterance, wherein the generated interpretation includes the series of syllables mapped to the recognized stream of phonemes.
10. The system of claim 9, the acoustic grammar phonemically representing the one or more syllables in accordance with acoustic elements of an acoustic speech model, wherein each syllable is represented by acoustic elements for an onset, a nucleus, and a coda.
11. The system of claim 10, the acoustic grammar including transitions between the acoustic elements, wherein the transitions are constrained according to phonotactic rules of the acoustic speech model.
12. The system of claim 11, the acoustic elements including at least an unstressed central vowel and a plurality of phonemic elements associated with the acoustic speech model, wherein the acoustic grammar uses the unstressed central vowel as a linking element between sequential phonemic elements.
13. The system of claim 12, the unstressed central vowel including a schwa acoustic element.
14. The system of claim 9, further comprising a sharpening engine that receives the generated interpretation of the utterance from the speech interpretation engine, the sharpening engine operable to:
generate a plurality of candidate interpretations of the utterance, wherein each candidate interpretation includes a series of words or phrases corresponding to the series of syllables mapped to the recognized stream of phonemes;
assign a score to each of the plurality of candidate interpretations; and
select a candidate interpretation having a highest assigned score as being a probable interpretation of the utterance.
15. The system of claim 14, the sharpening engine operable to generate the plurality of candidate interpretations of the utterance by:
identifying a series of words or phrases possibly corresponding to the series of syllables mapped to the recognized stream of phonemes; and
providing the identified series of words or phrases to a plurality of domain agents, each of the plurality of domain agents generating a candidate interpretation by:
identifying any suspect words or phrases among the identified series of words or phrases;
identifying one or more closest phonetic matches for the identified suspect words or phrases using a closest-distance metric associated with an M-Tree; and
substituting the identified closest phonetic matches for the identified suspect words or phrases to generate the candidate interpretation.
16. The system of claim 15, the sharpening engine operable to identify the series of words or phrases possibly corresponding to the series of syllables using one or more of a context associated the utterance, criteria identified in the utterance, a history of previous utterances, user profile information, or domain specific information.
Patent Citations (173)
| Patent | Date | Inventor | Cited By |
|---|---|---|---|
| US4430669(A) | 1984-02-01 | Cheung | Applicant |
| US5155743(A) | 1992-10-01 | Jacobs | Applicant |
| US5274560(A) | 1993-12-01 | LaRue | Applicant |
| US5377350(A) | 1994-12-01 | Skinner | Applicant |
| US5386556(A) | 1995-01-01 | Hedin et al. | Applicant |
| US5424947(A) | 1995-06-01 | Nagao et al. | Applicant |
| US5471318(A) | 1995-11-01 | Ahuja et al. | Applicant |
| US5475733(A) | 1995-12-01 | Eisdorfer et al. | Applicant |
| US5499289(A) | 1996-03-01 | Bruno et al. | Applicant |
| US5500920(A) | 1996-03-01 | Kupiec | Applicant |
| US5517560(A) | 1996-05-01 | Greenspan | Applicant |
| US5533108(A) | 1996-07-01 | Harris et al. | Applicant |
| US5537436(A) | 1996-07-01 | Bottoms et al. | Applicant |
| US5539744(A) | 1996-07-01 | Chu et al. | Applicant |
| US5557667(A) | 1996-09-01 | Bruno et al. | Applicant |
| US5563937(A) | 1996-10-01 | Bruno et al. | Applicant |
| US5590039(A) | 1996-12-01 | Ikeda et al. | Applicant |
| US5633922(A) | 1997-05-01 | August et al. | Applicant |
| US5675629(A) | 1997-10-01 | Raffel et al. | Applicant |
| US5696965(A) | 1997-12-01 | Dedrick | Applicant |
| US5708422(A) | 1998-01-01 | Blonder et al. | Applicant |
| US5721938(A) | 1998-02-01 | Stuckey | Applicant |
| US5722084(A) | 1998-02-01 | Chakrin et al. | Applicant |
| US5742763(A) | 1998-04-01 | Jones | Applicant |
| US5752052(A) | 1998-05-01 | Richardson et al. | Applicant |
| US5754784(A) | 1998-05-01 | Garland et al. | Applicant |
| US5761631(A) | 1998-06-01 | Nasukawa | Applicant |
| US5774859(A) | 1998-06-01 | Houser et al. | Applicant |
| US5794050(A) | 1998-08-01 | Dahlgren et al. | Applicant |
| US5799276(A) | 1998-08-01 | Komissarchik et al. | Applicant |
| US5802510(A) | 1998-09-01 | Jones | Applicant |
| US5832221(A) | 1998-11-01 | Jones | Applicant |
| US5878385(A) | 1999-03-01 | Bralich et al. | Applicant |
| US5878386(A) | 1999-03-01 | Coughlin | Applicant |
| US5892813(A) | 1999-04-01 | Morin et al. | Applicant |
| US5895466(A) | 1999-04-01 | Goldberg et al. | Applicant |
| US5902347(A) | 1999-05-01 | Backman et al. | Applicant |
| US5911120(A) | 1999-06-01 | Jarett et al. | Applicant |
| US5918222(A) | 1999-06-01 | Fukui et al. | Applicant |
| US5926784(A) | 1999-07-01 | Richardson et al. | Applicant |
| US5933822(A) | 1999-08-01 | Braden-Harder et al. | Applicant |
| US5953393(A) | 1999-09-01 | Culbreth et al. | Applicant |
| US5963894(A) | 1999-10-01 | Richardson et al. | Applicant |
| US5963940(A) | 1999-10-01 | Liddy et al. | Applicant |
| US5987404(A) | 1999-11-01 | Della Pietra et al. | Applicant |
| US5991721(A) | 1999-11-01 | Asano et al. | Applicant |
| US5995119(A) | 1999-11-01 | Cosatto et al. | Applicant |
| US6009382(A) | 1999-12-01 | Martino et al. | Applicant |
| US6014559(A) | 2000-01-01 | Amin | Applicant |
| US6021384(A) | 2000-02-01 | Gorin et al. | Applicant |
| US6044347(A) | 2000-03-01 | Abella et al. | Applicant |
| US6049602(A) | 2000-04-01 | Foladare et al. | Applicant |
| US6058187(A) | 2000-05-01 | Chen | Applicant |
| US6078886(A) | 2000-06-01 | Dragosh et al. | Applicant |
| US6081774(A) | 2000-06-01 | De Hita et al. | Applicant |
| US6101241(A) | 2000-08-01 | Boyce et al. | Applicant |
| US6119087(A) | 2000-09-01 | Kuhn et al. | Applicant |
| US6134235(A) | 2000-10-01 | Goldman et al. | Applicant |
| US6144667(A) | 2000-11-01 | Doshi et al. | Applicant |
| US6160883(A) | 2000-12-01 | Jackson et al. | Applicant |
| US6173279(B1) | 2001-01-01 | Levin et al. | Applicant |
| US6175858(B1) | 2001-01-01 | Bulfer et al. | Applicant |
| US6185535(B1) | 2001-02-01 | Hedin et al. | Applicant |
| US6192110(B1) | 2001-02-01 | Abella et al. | Applicant |
| US6192338(B1) | 2001-02-01 | Haszto et al. | Applicant |
| US6195634(B1) | 2001-02-01 | Dudemaine et al. | Applicant |
| US6195651(B1) | 2001-02-01 | Handel et al. | Applicant |
| US6208972(B1) | 2001-03-01 | Grant et al. | Applicant |
| US6219346(B1) | 2001-04-01 | Maxemchuk | Applicant |
| US6219643(B1) | 2001-04-01 | Cohen et al. | Applicant |
| US6233556(B1) | 2001-05-01 | Teunen et al. | Applicant |
| US6233559(B1) | 2001-05-01 | Balakrishnan | Applicant |
| US6246981(B1) | 2001-06-01 | Papineni et al. | Applicant |
| US6272455(B1) | 2001-08-01 | Hoshen et al. | Applicant |
| US6292767(B1) | 2001-09-01 | Jackson et al. | Applicant |
| US6314402(B1) | 2001-11-01 | Monaco et al. | Applicant |
| US6366886(B1) | 2002-04-01 | Dragosh et al. | Applicant |
| US6385646(B1) | 2002-05-01 | Brown et al. | Applicant |
| US6393428(B1) | 2002-05-01 | Miller et al. | Applicant |
| US6404878(B1) | 2002-06-01 | Jackson et al. | Applicant |
| US6408272(B1) | 2002-06-01 | White et al. | Applicant |
| US6411810(B1) | 2002-06-01 | Maxemchuk | Applicant |
| US6415257(B1) | 2002-07-01 | Junqua et al. | Applicant |
| US6418210(B1) | 2002-07-01 | Sayko | Applicant |
| US6420975(B1) | 2002-07-01 | DeLine et al. | Applicant |
| US6430285(B1) | 2002-08-01 | Bauer et al. | Applicant |
| US6434523(B1) | 2002-08-01 | Monaco | Applicant |
| US6434524(B1) | 2002-08-01 | Weber | Applicant |
| US6442522(B1) | 2002-08-01 | Carberry et al. | Applicant |
| US6446114(B1) | 2002-09-01 | Bulfer et al. | Applicant |
| US6453153(B1) | 2002-09-01 | Bowker et al. | Applicant |
| US6453292(B2) | 2002-09-01 | Ramaswamy et al. | Applicant |
| US6456711(B1) | 2002-09-01 | Cheung et al. | Applicant |
| US6466654(B1) | 2002-10-01 | Cooper et al. | Applicant |
| US6498797(B1) | 2002-12-01 | Anerousis et al. | Applicant |
| US6499013(B1) | 2002-12-01 | Weber | Applicant |
| US6501833(B2) | 2002-12-01 | Phillips et al. | Applicant |
| US6501834(B1) | 2002-12-01 | Milewski et al. | Applicant |
| US6513006(B2) | 2003-01-01 | Howard et al. | Applicant |
| US6523061(B1) | 2003-02-01 | Halverson et al. | Applicant |
| US6532444(B1) | 2003-03-01 | Weber | Applicant |
| US6539348(B1) | 2003-03-01 | Bond et al. | Applicant |
| US6553372(B1) | 2003-04-01 | Brassell et al. | Applicant |
| US6556973(B1) | 2003-04-01 | Lewin | Applicant |
| US6560576(B1) | 2003-05-01 | Cohen et al. | Applicant |
| US6567778(B1) | 2003-05-01 | Chao Chang et al. | Applicant |
| US6567797(B1) | 2003-05-01 | Schuetze et al. | Applicant |
| US6570555(B1) | 2003-05-01 | Prevost et al. | Applicant |
| US6570964(B1) | 2003-05-01 | Murveit et al. | Applicant |
| US6574597(B1) | 2003-06-01 | Mohri et al. | Applicant |
| US6574624(B1) | 2003-06-01 | Johnson et al. | Applicant |
| US6594257(B1) | 2003-07-01 | Doshi et al. | Applicant |
| US6598018(B1) | 2003-07-01 | Junqua | Applicant |
| US6604077(B2) | 2003-08-01 | Dragosh et al. | Applicant |
| US6611692(B2) | 2003-08-01 | Raffel et al. | Applicant |
| US6614773(B1) | 2003-09-01 | Maxemchuk | Applicant |
| US6615172(B1) | 2003-09-01 | Bennett et al. | Applicant |
| US6629066(B1) | 2003-09-01 | Jackson et al. | Applicant |
| US6631346(B1) | 2003-10-01 | Karaorman et al. | Applicant |
| US6643620(B1) | 2003-11-01 | Contolini et al. | Applicant |
| US6650747(B1) | 2003-11-01 | Bala et al. | Applicant |
| US6691151(B1) | 2004-02-01 | Cheyer et al. | Applicant |
| US6721001(B1) | 2004-04-01 | Berstis | Applicant |
| US6721706(B1) | 2004-04-01 | Strubbe et al. | Applicant |
| US6735592(B1) | 2004-05-01 | Neumann et al. | Applicant |
| US6741931(B1) | 2004-05-01 | Kohut et al. | Applicant |
| US6742021(B1) | 2004-05-01 | Halverson et al. | Applicant |
| US6757718(B1) | 2004-06-01 | Halverson et al. | Applicant |
| US6795808(B1) | 2004-09-01 | Strubbe et al. | Applicant |
| US6901366(B1) | 2005-05-01 | Kuhn et al. | Applicant |
| US6937977(B2) | 2005-08-01 | Gerson | Applicant |
| US6944594(B2) | 2005-09-01 | Busayapongchai et al. | Applicant |
| US6980092(B2) | 2005-12-01 | Turnbull et al. | Applicant |
| US6990513(B2) | 2006-01-01 | Belfiore et al. | Applicant |
| US6996531(B2) | 2006-02-01 | Korall et al. | Applicant |
| US7020609(B2) | 2006-03-01 | Thrift et al. | Applicant |
| US7027975(B1) | 2006-04-01 | Pazandak et al. | Applicant |
| US7062488(B1) | 2006-06-01 | Reisman | Applicant |
| US7092928(B1) | 2006-08-01 | Elad et al. | Applicant |
| US7127400(B2) | 2006-10-01 | Koch | Applicant |
| US7137126(B1) | 2006-11-01 | Coffman et al. | Applicant |
| US7146319(B2) | 2006-12-01 | Hunt | Examiner |
| US7289606(B2) | 2007-10-01 | Sibal etal. | Applicant |
| US7376645(B2) | 2008-05-01 | Bernard | Applicant |
| US7398209(B2) | 2008-07-01 | Kennewick et al. | Applicant |
| US7424431(B2) | 2008-09-01 | Greene et al. | Applicant |
| US7493559(B1) | 2009-02-01 | Wolff et al. | Applicant |
| US2001/0041980(A1) | 2001-11-01 | Howard et al. | Applicant |
| US2002/0049805(A1) | 2002-04-01 | Yamada et al. | Applicant |
| US2002/0065568(A1) | 2002-05-01 | Silfvast et al. | Applicant |
| US2002/0082911(A1) | 2002-06-01 | Dunn et al. | Applicant |
| US2002/0120609(A1) | 2002-08-01 | Lang et al. | Applicant |
| US2002/0124050(A1) | 2002-09-01 | Middeljans | Applicant |
| US2002/0143535(A1) | 2002-10-01 | Kist et al. | Applicant |
| US2002/0188602(A1) | 2002-12-01 | Stubler et al. | Applicant |
| US2002/0198714(A1) | 2002-12-01 | Zhou | Applicant |
| US2003/0014261(A1) | 2003-01-01 | Kageyama | Applicant |
| US2003/0112267(A1) | 2003-06-01 | Belrose | Applicant |
| US2003/0182132(A1) | 2003-09-01 | Niemoeller | Applicant |
| US2004/0025115(A1) | 2004-02-01 | Sienel et al. | Applicant |
| US2004/0044516(A1) | 2004-03-01 | Kennewick et al. | Applicant |
| US2004/0166832(A1) | 2004-08-01 | Portman et al. | Applicant |
| US2004/0193420(A1) | 2004-09-01 | Kennewick et al. | Applicant |
| US2004/0199375(A1) | 2004-10-01 | Ehsani et al. | Applicant |
| US2005/0021334(A1) | 2005-01-01 | Iwahashi | Applicant |
| US2005/0021826(A1) | 2005-01-01 | Kumar | Applicant |
| US2005/0033574(A1) | 2005-02-01 | Kim et al. | Applicant |
| US2005/0114116(A1) | 2005-05-01 | Fiedler | Applicant |
| US2007/0033005(A1) | 2007-02-01 | Cristo et al. | Applicant |
| US2007/0038436(A1) | 2007-02-01 | Cristo et al. | Applicant |
| US2007/0050191(A1) | 2007-03-01 | Weider et al. | Applicant |
| US2008/0091406(A1) | 2008-04-01 | Baldwin et al. | Applicant |
| WO1/78065 | 2001-10-01 | Applicant |
Non-Patent Literature (14)
- Chai et al., “MIND: A Semantics-Based Multimodal Interpretation Framework for Conversational System”, Proceedings of the International CLASS Workshop on Natural, Intelligent and Effective Interaction in Multimodal Dialogue Systems, Jun. 2002, pp. 37-46.Applicant
- Cheyer et al., “Multimodal Maps: An Agent-Based Approach”, International Conference on Cooperative Multimodal Communication (CMC/95), May 24-26, 1995, pp. 111-121.Applicant
- Elio et al., “On Abstract Task Models and Conversation Policies” in Workshop on Specifying and Implementing Conversation Policies, Autonomous Agents '99, Seattle, 1999, 10 pages.Applicant
- Turunen, “Adaptive Interaction Methods in Speech User Interfaces”, Conference on Human Factors in Computing Systems, Seattle, Washington, 2001, pp. 91-92.Applicant
- Reuters, “IBM to Enable Honda Drivers to Talk to Cars”, Charles Schwab & Co., Inc., Jul. 28, 2002, 1 page.Applicant
- Lin, Bor-shen, et al., “A Distributed Architecture for Cooperative Spoken Dialogue Agents with Coherent Dialogue State and History”, ASRU'99, 1999, 4 pages.Applicant
- Kuhn, Thomas, et al., “Hybrid In-Car Speech Recognition for Mobile Multimedia Applications”, Vehicular Technology Conference, IEEE, Jul. 1999, pp. 2009-2013.Applicant
- Belvin, Robert, et al., “Development of the HRL Route Navigation Dialogue System”, Proceedings of the First International Conference on Human Language Technology Research, San Diego, 2001, pp. 1-5.Applicant
- Lind, R., et al., “The Network Vehicle—A Glimpse into the Future of Mobile Multi-Media”, IEEE Aerosp. Electron. Systems Magazine, vol. 14, No. 9, Sep. 1999, pp. 27-32.Applicant
- Zhao, Yilin, “Telematics: Safe and Fun Driving”, IEEE Intelligent Systems, vol. 17, Issue 1, 2002, pp. 10-14.Applicant
- El Meliani at al., “A Syllabic-Filler-Based Continuous Speech Recognizer for Unlimited Vocabulary”, Canadian Conference on Electrical and Computer Engineering, vol. 2, Sep. 5-8, 1995, pp. 1007-1010.Applicant
- Mao, Mark Z., “Automatic Training Set Segmentation for Multi-Pass Speech Recognition”, Department of Electrical Engineering, Stanford University, Ca, copyright 2005, IEEE, pp. 1-685 to 1-688.Applicant
- Vanhoucke, Vincent, “Confidence Scoring and Rejection Using Multi-Pass Speech Recognition”, Nuance Communications, Menlo Park, CA, [no date], 4 pages. 2005.Applicant
- Weng, Fuliang, et al., “Efficient Lattice Representation and Generation”, Speech Technology and Research Laboratory, SRI International, Menlo Park, CA, [no date], 4 pages. 1998.Applicant