Speech recognition apparatus and method using two opposite words

[0001] This application is based on and incorporates herein by reference Japanese Patent Application No. 2000-220564 filed on Jul. 21, 2000.

[0002] 1. Field of the Invention

[0003] The present invention relates to a speech recognition apparatus and method.

[0004] 2. Related Art

[0005] A conventional speech recognition apparatus and method used as a man-machine interface analyzes a speech signal received via a microphone and extracts characteristic parameters from the speech signal. The characteristic parameters are compared with reference patterns in a vocabulary stored in a dictionary section of the apparatus, and one or more words corresponding to the reference patterns which have high similarities with the extracted characteristic parameters are selected and outputted. The apparatus generally selects and outputs at least a word even if no reference pattern in the vocabulary has a high similarity with the characteristic parameters. In this case, the speech signal is often misrecognized. Further the speech signal is often misrecognized due to the background noise. If the result of the misrecognition is used for controlling an external device, the operation of the external device may be adversely and seriously affected.

[0006] For example, the speech recognition apparatus is used for a car navigation device which has a voice-control function for receiving a speech as an input control signal and outputting a speech as a query to a user or a reply to the user's query. If the user says ‘shukushou’ which means ‘reduce’ so that the map is reduced, it may be misrecognized as ‘jitaku’ which means the user's house. In this case, the apparatus discards the route to the destination and newly searches a route to the user's house, if the apparatus is configured to execute a command immediately in response to the input speech signal. In this way, the route to the destination is discarded against the user's will.

[0007] If the apparatus is configured to confirm the user's will, it queries ‘Your house will be set as the destination. OK?’, for example. If the user says ‘iie’ which means ‘NO’ as a reply to the query, the apparatus shifts to a wait state without searching a route to the user's house. However, the apparatus may misrecognize a user's reply as ‘hai’ which means ‘YES’, if the user says ‘shukushou’ again as a reply to the query by mistake. In this case, the route to the destination is also discarded.

[0008] In this way, the external device may operate against the user's will, if the speech signal from the user is misrecognized as a specific word such as ‘hai’ included in the vocabulary. Therefore, it is preferable that the speech recognition apparatus employs a technique for preventing the external device from performing an erroneous operation, which cannot be easily undone, against the user's will in response to the result of misrecognition.

[0009] The present invention has an object to provide a speech recognition apparatus which rarely misrecognizes an input speech signal as a specific word in the vocabulary. It is also an object to provide a record medium which stores a program and data for implementing a main function of the speech recognition apparatus in a computer system.

[0010] In a speech recognition apparatus according to the present invention, a pattern matching section performs pattern matching between each of reference patterns in a vocabulary and the extracted characteristic parameters of a speech signal. The vocabulary is stored beforehand and includes the reference patterns corresponding to words which should be identified. Further, a specific word in response to which the external device performs an operation which cannot be easily undone is selected from the vocabulary, and the similar sound group of the specific word is stored beforehand. The similar sound group includes the reference patterns corresponding to sounds which are similar to but different from that of a specific word. The pattern matching section selects as candidate words one or more words corresponding to the reference patterns which have high similarities with the characteristic parameters as the result of the pattern matching. If the candidate words include the specific word, the speech signal is rerecognized by using the similar sound group. That is, the pattern matching section performs pattern matching between each of the reference patterns in the similar sound group and the characteristic parameters of the speech signal. As a result of the rerecognition, the pattern matching section outputs a word other than the specific word if one of the reference patterns corresponding to the sounds similar to that of the specific word has a high similarity with the speech signal.

[0011] Preferably a program for implementing the function of the pattern matching section in a computer system and data including the reference patterns corresponding to the words and the sounds are provided as a record medium which stores the program and the data.

[0018] (First Embodiment)

[0019] Referring to FIG. 1, a car navigation equipment is connected to a speech recognition apparatus 18 used as a man-machine interface for the car navigation equipment 1. The car navigation equipment 1 includes a position detecting unit 2, a map data input unit 3, control switches 4, a control circuit 5, an external memory 6, a display unit 7, a remote-control signal sensor 8, a dialog control section 9, speech synthesis section 10, and a loudspeaker 11. The control circuit 5 mainly comprises a microcomputer, and connected to the position detecting unit 2, the map data input unit 3, and the control switches 4. The external memory 6, the display unit 7, the remote-control signal sensor 8, and the dialog control section 9 are connected to the control circuit 5.

[0020] The position detecting unit 2 includes a geomagnetic sensor 12, a gyro sensor 13, a speed sensor 14 and a GPS receiver 15. The GPS receiver 15 detects the position of a vehicle based on the radio wave received from the GPS satellite. Each of the sensors 12-15 is placed in an appropriate position within the vehicle. The control circuit 5 receives signals from the sensors 12-15 and interpolates them, because each of the sensors 12-15 has an error of a different type. The control circuit 5 detects the present position, traveling direction, speed, mileage of the vehicle accurately based on the received signals.

[0021] The map data input unit 3 is formed with a drive unit for retrieving data from a storage medium which stores various data such as road map data, map matching data for correcting the detected present position. A mass storage medium such as CD-ROM, DVD-ROM may be employed as the storage medium. The road map data includes the shapes, widths, names of roads, and various buildings or facilities on the roads and their phone numbers. Further it includes the names and geographical features of places, and data used for displaying the road map on the display unit 7.

[0022] The control switches 4 include mechanical switches which are used by a user (driver) for inputting various commands. The user can designate the destination or select the road map displayed on the display unit 7 by using control switches 4. Some of the control switches 4 may include touch switches on the screen of the display unit 7. Further a remote-control terminal (remote controller) 16 which has the same function as the control switches 4 is provided. A control signal from the remote controller 16 is detected by the remote-control signal sensor 8.

[0023] On the screen of the display unit 7, the road map is displayed with various scales. Further pointers which indicate the present position or traveling direction of the vehicle are superimposed on the displayed road map. A recommended route to the destination may be also superimposed on the displayed road map. Further a form for prompting the user to input the destination and the like or various massages and information are also displayed on the screen.

[0024] The road map is retrieved by the map data input unit 3. The control circuit 5 directs the display unit 7 to display the retrieved road map on the screen, and further to superimpose the pointers which indicate the present position and the traveling direction of the vehicle based on the output from the position detecting unit 2. For indicating the present position of the vehicle accurately on the road map, the control circuit 5 performs map matching. Further the control circuit 5 directs the display unit 7 to change (the scale of) the map displayed on the screen.

[0025] The control circuit 5 performs an automatic route search and route guidance. In the automatic route search, the control circuit 5 automatically calculates a recommended route from the present position of the vehicle to a destination inputted by the user. In the route guidance, the control circuit 5 directs the display unit 7 to display the recommended route on the screen as described above. Further the control circuit 5 may direct the speech synthesis section 10 to generate an audio guide such as ‘turn left at the intersection 200 meters ahead’ via the dialog control section 9 during the route guidance. The generated audio guide is outputted via the loudspeaker 11.

[0026] The speech recognition apparatus 18 is connected to the dialog control section 9 so that the user's safety and the user-friendliness of the car navigation equipment 1 are improved. The user (driver) can input a command (destination search command, reroute command or map scale change command) or the destination (user's house or registered point) by voice while keeping his/her eyes on the road ahead. The speech recognition apparatus 18 receives a speech signal, which represents a command or the destination, from the user via a microphone 17, and recognizes it. The result of the recognition is provided for the dialog control section 9. The microphone 17 is placed in an appropriate position within the vehicle so that it receives the user's voice well. For example, the microphone 17 is attached to the top surface of the steering column cover or to the sun visor over the driver's seat.

[0027] Referring to FIG. 2, the speech recognition apparatus 18 includes an acoustic analysis section 19, a feature extraction section 20, and a pattern matching section 21.

[0028] The acoustic analysis section 19 receives the analog speech signal from the microphone 17, and samples it at a sampling frequency of 12 kHz into digital signals. The digital signals are divided into frames so that the frames have a fixed length and overlap each other. For each of the frames, the Fourier transform of the frame is computed, so that the acoustic spectrum of the speech signal is provided. Further the square of the amplitude of the acoustic spectrum is computed, so that the short-time acoustic power spectrum is obtained. The inverse Fourier transform of the power spectrum is computed, so that auto-correlation coefficients are provided. The LPC (linear predictive coding) analysis is performed using the auto-correlation coefficients, so that LPC coefficients are obtained. The LPC coefficients are sequentially provided for the feature extraction section 20.

[0029] The feature extraction section 20 computes LPC cepstrum coefficients for each of the frames based on the LPC coefficients received from the acoustic analysis section 19. The LPC cepstrum coefficients are sequentially provided for the pattern matching section 21 as characteristic parameters.

[0030] The pattern matching section 21 performs pattern matching between each of reference patterns in a vocabulary stored in the dictionary section and time-series data of the LPC cepstrum coefficients. That is, the time-series data is divided into segments by using hidden Markov models and the similarity (i.e., likelihood ratio) between each of the reference patterns and each of the segments is computed. The dictionary section is a part of a fixed memory of the speech recognition apparatus. Each of the reference patterns is a time-series of LPC cepstrum coefficients which are computed beforehand and correspond to one of words which should be identified.

[0031] The pattern matching section 21 selects as candidate words one or more words corresponding to the reference patterns which have the high similarities with the LPC cepstrum coefficients. In the case of a conventional speech recognition apparatus, the candidate words are provided for the dialog control section 9 as the result of the recognition.

[0032] However, the present speech recognition apparatus 18 rerecognizes the speech signal as follows, if the speech signal has been received in some situations and the candidate words include a specific word. The specific word is a critical word for the car navigation equipment 1. That is, the car navigation equipment 1 performs an operation that cannot be easily undone if it receives the specific word. The speech recognition apparatus 18 generates and stores the similar sound group of the specific word beforehand. The similar sound group includes reference patterns corresponding to sounds which are different from but similar to that of the specific word. The rerecognition of the speech signal is performed by using the similar sound group of the specific word.

[0033] The apparatus 18 further generates reference patterns corresponding to sounds similar to that of a second specific word beforehand, if necessary. The second specific word is a word which means the opposite to the specific word. The generated reference patterns are added to the similar sound group.

[0034] The rerecognition of the speech signal will be explained by means of example. When the car navigation equipment 1 receives a voice-command such as ‘reroute’, ‘my house’, ‘go home’ or ‘change the destination’, it outputs a voice message such as ‘Your home will be set as the destination. OK?’ for confirming the user's will. In this way, the user is prompted to input ‘YES’ or ‘NO’ by voice as a reply to the voice-message. In this case, the speech recognition apparatus 18 selects the word ‘YES’ as the above specific word and the similar sound group of the word ‘YES’ is generated. This similar sound group includes reference patterns corresponding to sounds /j//*//s/, /j//e//*/ and the like. The symbol ‘*’ stands for an arbitrary vowel or a consonant. Further, the word ‘NO’ is selected as the second specific word and reference patterns corresponding to sounds similar to that of the word ‘NO’ is also generated and added to the similar sound group. The reference patterns corresponding to sounds /au/, /uu/, and the like are added to the similar sound group in this case. It is preferable that the similar sound group should include the reference patterns corresponding to sounds similar to that of the second specific word, when the speech signal is received in an either-or situation like this case.

[0035] The pattern matching section 21 performs pattern matching between each of the reference patterns in the similar sound group and the LPC cepstrum coefficients of the speech signal, when the candidate words include the specific word ‘YES’. If one of the reference patterns in the similar sound group has a high similarity with the LPC cepstrum coefficients, the pattern matching section 21 outputs the second specific word ‘NO’ as the result of recognition.

[0036] In this way, the present speech recognition apparatus 18 recognizes the speech signal received in the situation of the above example as the word ‘YES’ only when the time-series of the LPC cepstrum coefficients is very close to the reference pattern of the word ‘YES’. If the input speech signal is similar to but different from the specific word ‘YES’ or similar to but different from the second specific word ‘NO’, the word ‘NO’ would be outputted because the speech signal would have a high similarity with one of the reference patterns in the similar sound group. That is, the feature space of an input speech signal that is recognized as the word ‘YES’ is narrow and the feature space of an input speech signal that is recognized as the word ‘NO’ is broad as shown in FIG. 3A. In FIG. 3A, the horizontal axis represents, for example, the value of zeroth-degree term of the LPC cepstrum coefficients and the vertical axis represents the value of first-degree term of the LPC cepstrum coefficients. Further mark ◯ represents the reference pattern corresponding to the word ‘YES’, and marks x represent the reference pattern corresponding to the word ‘NO’ and the reference patterns in the similar sound group. Thus the word ‘NO’ would be outputted as the result of recognition, if the candidate words include the word ‘YES’ but the level of confidence that the speech signal actually represent the word ‘YES’.

[0037] In contrast to this, in the case of a conventional speech recognition apparatus, the characteristic space of an input speech signal that is recognized as the word ‘YES’ is relatively broad as shown in FIG. 3B. In FIG. 3B, the horizontal axis represents, for example, the value of zeroth-degree term of the LPC cepstrum coefficients and the vertical axis represents the value of first-degree term of the LPC cepstrum coefficients. Further mark ◯ represents the reference pattern corresponding to the word ‘YES’, and marks x represent the reference pattern corresponding to the word ‘NO’. Therefore, the speech signal that is similar to but different from the word ‘YES’ would be misrecognized as the word ‘YES’.

[0038] The above effect of the present speech recognition apparatus 18 was verified through experimentation in the configuration shown in FIG. 2. The result of the experimentation is shown in FIG. 4. When input speech signals were not rerecognized by using the similar sound group, the probability that the speech signals which were different from the word ‘YES’ or the word ‘NO’ were misrecognized as the word ‘YES’ was relatively high (prior art: 61.9%). In contrast to this, such a probability was much lower (embodiment: 8.7%) when input speech signals were rerecognized by using the similar sound group. Further, the recognition rate of the word ‘NO’ was slightly higher (embodiment: 98.2%). The recognition rate of the word ‘YES’ was the same in both cases.

[0039] Accordingly, the car navigation equipment 1 is prevented from performing an operation that cannot be easily undone against the user's will, if an input speech signal received from the user in the situation of the above example is recognized by the present speech recognition apparatus.

[0040] In the present embodiment, Japanese corresponding words ‘hai and ‘iie’ may be selected as the specific word and the second specific word instead of ‘YES’ and ‘NO’. Further, German corresponding words ‘ja’ and ‘nein’, French corresponding words “oui’ and ‘non’, Dutch corresponding words ‘ja’ and ‘nee’, Spanish corresponding words ‘sí’ and ‘no’, Italian corresponding words ‘sí’ and ‘no’, or Portuguese corresponding words ‘sim’ and ‘não’ may be selected as the specific word and the second specific word.

[0041] (Second Embodiment)

[0042] Referring to FIG. 5, a speech recognition apparatus 18 includes an acoustic analysis section 19, a feature extraction section 20, and a pattern matching section 21′ similarly to the first embodiment. The speech recognition apparatus 18 receives an input speech signal from a microphone 17 and recognizes it. The result of the recognition is outputted to a dialog control section 9 of a car navigation equipment 1. When the speech signal is recognized, the acoustic analysis section 19 and the feature extraction section 20 operate similarly to the first embodiment. However, the pattern matching section 21′ performs pattern matching by using only ordinary dictionary data, and does not perform pattern matching by using the similar sound group. The ordinary dictionary data is stored in the dictionary section of the speech recognition apparatus 18 and includes reference patterns corresponding to words which should be identified. The pattern matching section 21′ generates candidate words as the result of the pattern matching between the time-series data of LPC cepstrum coefficients received from the feature extraction section 20 and each of the reference patterns.

[0043] As shown in FIG. 5, the speech recognition apparatus 18 further includes a success probability estimation section 22 disclosed in Japanese Patent Application No. H11-359076 for estimating the level of confidence that the input speech signal actually represents a candidate word. When the candidate words include a specific word which is a critical word for the car navigation equipment 1, the pattern matching section 21 provides characteristic values concerning the recognition of the specific word for the success probability estimation section 22. The characteristic values provided by the pattern matching section 21′ represent the likelihood ratio of the specific word and a connected-syllable model or the variance of duration of syllables in the specific word. The success probability estimation section 22 receives the characteristic values concerning the recognition of a candidate word, which is the specific word in this case, and estimates as the absolute level of confidence the probability that the input speech signal actually represents the candidate word based on the received characteristic values.

[0044] The characteristic values may be provided by the acoustic analysis section 19 instead of the pattern matching section 21′ as shown by the broken line in FIG. 5. In this case, values which represent the degree of success in the acoustic analysis (i.e., the noise level or LPC residuals) are provided as the characteristic values. Alternatively the characteristic values may be provided by the feature extraction section 20 instead of the pattern matching section 21′ as shown by the broken line in FIG. 5. In this case, values which represent LPC cepstrum coefficients are provided as the characteristic values.

[0045] The absolute level of confidence estimated by the success probability estimation section is provided for the pattern matching section 21′ in the form of the probability that the input speech signal actually represents the specific word. The level of confidence can be grasped intuitively and quantitatively, if it is provided in the form of the success probability in this way. The pattern matching section 21′ outputs a candidate word other than the specific word as the result of recognition, if the received absolute level of confidence is equal to or lower than the predetermined reference level. The reference level is experimentally determined beforehand.

[0046] As described above, the words corresponding to the reference patterns which have the high similarities with the LPC cepstrum coefficients are selected as the candidate words. Therefore the possibility that the input speech signal actually represents the candidate word is relatively high, but not always absolutely high. Accordingly, in the present speech recognition apparatus, the pattern matching section 21′ determines based on the absolute level of confidence provided by the success probability estimation section 22 whether it outputs as the result of recognition the specific word which is one of the candidate word.

[0047] In this way, the specific word is outputted as the result of recognition only when the level of confidence that the input speech signal actually represents the specific word is absolutely high. That is, the possibility that the input speech signal which does not actually represent the specific word is misrecognized as the specific word is very low. The specific word is a critical word for the car navigation equipment 1 as described above. Accordingly, the car navigation equipment 1 is prevented from performing an operation which cannot be easily undone against the user's will.

[0048] (Modifications)

[0049] In the first embodiment, a program for implementing the function of the pattern matching section 21 in a computer system and data such as the reference patterns and the similar sound group may be provided as a detachable record medium which stores the program and the data. Various record media such as CD-ROM, DVD-ROM, non-volatile memory (i.e., EEPROM), a floppy disk, a magnetic optical disk, a hard disk may be employed as the record medium for storing the program and the data.

[0050] In the second embodiment, a program for implementing the function of the pattern matching section 21′ and the success probability estimation section 22 in a computer system and data such as the reference patterns may be also provided as a detachable record medium which stores the program and the data. Further a confidence level estimation section which provides the absolute level of confidence in the form other than the success probability may be employed instead of the success probability estimation section 22.

[0051] The speech recognition apparatus 18 according to the present invention may be used for an external device other than the car navigation equipment 1, for example, a personal mobile navigation device or other popular electrical appliances.