JP2007206501A - Device for determining optimum speech recognition system, speech recognition device, parameter calculation device, information terminal device and computer program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To predict the accuracy of speech recognition without any test utterance. <P>SOLUTION: The optimum speech recognition system selection device for determining a system capable of performing the highest precise speech recognition among a plurality of speech recognition systems with respect to the speech including noise comprises: a feature amount calculating section 180 for calculating a predetermined sound feature amount vector from input speech; a transformation processing section 182 for transforming the sound feature amount vector to the feature amount vector of small dimensions by a transformation matrix; a prediction accuracy calculating section 184 for predicting accuracy of the speech recognition for each speech recognition system by using the feature amount vector after transformation and a multiple regression coefficient prepared beforehand; and a maximum value selecting section 185 for selecting the speech recognition system by which speech recognition can be performed with the highest precision. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a system for recognizing speech uttered in an environment where there is a possibility that noise of an unpredictable type exists, and in particular, using a plurality of speech recognition methods, speech recognition is performed accurately even under noise. The present invention relates to a speech recognition system capable of doing things, and a computer program therefor.

  In recent years, devices capable of recording voice information, such as voice recorders, answering machines, and mobile phones, have been widely used. However, since voice information requires a large storage capacity, it is desirable to convert it into text information by voice recognition and record it. If such a service can be provided, for example, in a mobile phone, information such as an address for a telephone directory can be input by voice and converted into text information by voice recognition and stored. By doing so, it is possible not only to reproduce the information as a sound but also to display it on the screen as characters, and to use the information in other devices.

  As one of such services, a service has been put into practical use in which voice input by a mobile phone is transmitted to a server and voice recognition is performed by the server. The speech recognition result is returned from the server to the mobile phone. In the cellular phone, various processes can be performed using the voice recognition result.

  In such a system using voice recognition, the accuracy of voice recognition becomes a problem. There are various speech recognition methods, and the accuracy varies depending on the speech recognition target.

  As such a speech recognition method, a speech recognition apparatus that improves speech recognition accuracy by selecting a speech recognition result having the best score from candidates of speech recognition results by a plurality of speech recognition methods has been proposed.

  FIG. 1 shows a block diagram of the server 30 of such a conventional speech recognition system. Referring to FIG. 1, server 30 uses first to third voice recognition units 50 to 54, and includes communication unit 40 for exchanging data with a mobile phone or the like, and input voice A frame forming unit 42 for converting a frame into a frame with a predetermined frame length and a predetermined shift length, and acoustic features used by the first to third speech recognition units 50 to 54 from the framed speech, respectively. And first to third feature quantity extraction units 44 to 48 for providing to the first to third voice recognition units 50 to 54.

  The first to third voice recognition units 50 to 54 perform voice recognition using different voice recognition methods. As the speech recognition, statistical speech recognition using an acoustic model and a language model has been mainly used recently, and the first to third speech recognition units 50 to 54 perform statistical speech recognition although the methods are different. Is. In statistical speech recognition, various speech recognition result candidates are generated with probabilities that they are considered to be correct speech recognition results. These are called the “score” of each candidate.

  The server 30 further transmits a selection unit 56 for selecting a candidate having the highest score among the outputs of the first to third speech recognition units 50 to 54 as a speech recognition result, and sends the selected speech recognition result from the selected speech recognition result. A transmission data creation unit 58 for creating credit data and transmitting the credit data to the mobile phone via the communication unit 40;

  The server 30 operates as follows. With reference to FIG. 1, the voice of the user of the mobile phone is given to communication unit 40 of server 30. The given voice is framed by the framing unit 42 with a predetermined frame length and a predetermined shift length. The framed voice is given to the first to third feature quantity extraction units 44 to 48.

  The first to third feature amount extraction units 44 to 48 extract acoustic feature amounts for the first to third speech recognition units 50 to 54 from the respective frames, and the first to third speech recognition units. 50-54. In this way, the first to third feature amount extraction units 44 to 48 to the first to third speech recognition units 50 to 54 are respectively provided with sequences of acoustic feature amounts extracted from the respective frames.

  The first to third speech recognition units 50 to 54 perform speech recognition on a given sequence of acoustic feature values, and output speech recognition result candidates as texts. At this time, a score is assigned to the candidate speech recognition result. It is considered that the higher the score, the more accurately the speech can be recognized with respect to the input speech. These speech recognition result candidates are given to the selection unit 56 together with the score.

  The selection unit 56 selects the one with the highest score from a plurality of speech recognition result candidates. The selected speech recognition result with the highest score is given to the transmission data creation unit 58. The transmission data creation unit 58 creates data for transmission suitable for transmission from the voice recognition result and gives the data to the communication unit 40. This transmission data is transmitted to a mobile phone or the like by the communication unit 40. A mobile phone or the like that has received the transmission data extracts the voice recognition result from the received transmission data and uses it.

  In such a method, speech recognition is performed by a plurality of speech recognition methods, and a result that seems to have the highest reliability among the results is adopted. Therefore, regardless of the type of input voice, the accuracy can be expected to be higher than when performing single voice recognition.

  On the other hand, another important problem in speech recognition is the problem of noise. In general, it is known that in speech recognition, if the speech information to be recognized contains noise, the recognition accuracy decreases. In particular, in a system in which voice is input via a mobile phone, it is considered that various kinds of noise are mixed in an utterance because there are various places where the voice is input. Therefore, a voice recognition system using a mobile phone needs a mechanism that can recognize voice input in an environment where such background noise exists with high accuracy.

  As described above, a system that employs a plurality of speech recognition methods cannot sufficiently cope with noise as it is.

  In order to cope with speech recognition when the type of background noise cannot be predicted as described above, there is a method using an acoustic model corresponding to various noise data. In this method, a plurality of learning data is created by superimposing a plurality of types of noise data on predetermined utterance data in advance, and by learning a plurality of acoustic models using this learning data, An acoustic model for speech recognition is created for each noise data, and in speech recognition, speech recognition is performed using the plurality of speech models, and the result with the best score is adopted.

  However, even in this method, a plurality of voice recognition devices must be operated simultaneously, and the load on the server is high. In addition, if voice recognition is performed using a plurality of acoustic models for each voice recognition method, the reliability of the result may be increased, but the load on the server will increase dramatically.

  As a method for accurately recognizing various noises without increasing the server load, there is a method using a test utterance. In this method, prior to speech recognition processing, a user of a speech recognition system utters a predetermined text, extracts a predetermined acoustic feature amount from speech information obtained from the utterance, and uses an acoustic model for each noise. Speech recognition is performed, and at the time of this speech recognition, since the content of the test utterance, that is, the correct answer of the speech recognition result is known in advance, each speech recognition device is caused to calculate a score of the speech recognition result that gives the correct answer. Then, it is considered that the acoustic model employed in the speech recognition apparatus that gave the highest score corresponds best to the background noise of the input speech, and this acoustic model is used in subsequent speech recognition.

  In this method, speech recognition can be performed using an acoustic model that is most likely to obtain the best result in that environment regardless of the type of background noise. Further, after the acoustic model and the voice recognition method to be used are determined by the test utterance, only one voice recognition device operates for the user, so that an increase in server load can be avoided.

  However, this method has a problem that the user has to make a test utterance before using the speech recognition apparatus. That is, if it is necessary to perform test utterances before and after speech recognition, it is difficult to easily use the speech recognition apparatus.

  Further, according to the conventional technologies listed above, the user cannot know the accuracy of the voice recognition environment when the voice recognition is performed. For the user, it is not known whether or not a reliable result can be obtained using voice recognition, and there is a problem that the user feels uneasy about the use of voice recognition.

  Accordingly, one of the objects of the present invention is to provide an optimum speech recognition method selection device and a speech recognition device that can predict the accuracy of speech recognition without a prior test utterance.

  Another object of the present invention is to provide an information terminal device that can inform the user of the accuracy prediction result of speech recognition in an easy-to-understand manner.

  Furthermore, still another object of the present invention is to provide a speech recognition device, an optimal speech recognition method selection device, and a parameter calculation device that can reduce the amount of processing when performing speech recognition by a plurality of speech recognition methods.

  The optimum speech recognition method selection apparatus according to the first aspect of the present invention has the highest accuracy among a plurality of predetermined speech recognition methods for speech input in an environment where noise may exist. An optimum speech recognition method selection device for determining a method that is predicted to be capable of high speech recognition, and calculates a feature amount vector having a plurality of predetermined acoustic feature amounts as components from input speech A conversion means for converting the feature quantity vector into a feature quantity vector having a smaller number of dimensions, a feature quantity vector after conversion by the conversion means, and a plurality of voices. Accuracy prediction means for predicting the accuracy of speech recognition for each speech recognition method by a predetermined calculation between coefficient vectors for accuracy prediction prepared in advance for each recognition method; Based on the prediction calculated by the predicting means, and means for selecting the speech recognition system is capable of speech recognition with the highest accuracy.

  According to this optimum speech recognition method selection device, the accuracy of speech recognition is improved by using a feature vector obtained from the acoustic feature of the input speech, a pre-calculated transformation matrix, and a coefficient vector for accuracy prediction. Predict by speech recognition method. Then, the highest accuracy is selected from the predicted speech recognition accuracy, and the corresponding speech recognition method can be selected as the optimum speech recognition method for recognizing the input speech. There is no need to use test utterances for prediction. Accordingly, it is possible to provide an optimum speech recognition method selection device that can predict the accuracy of speech recognition without a prior test utterance.

  A speech recognition apparatus according to a second aspect of the present invention is a speech recognition apparatus for performing speech recognition using a plurality of speech recognition means, a communication means for exchanging data with other devices, and a communication Separating means for receiving voice information and method specifying information for specifying any of a plurality of voice recognition means from other devices via the means, and separating the voice information and the method specifying information, respectively A plurality of voice recognition means for performing voice recognition by a specific voice recognition method, and a method selection means for causing voice information to be recognized by selectively operating one of the plurality of voice recognition means according to the method specification information And means for returning the output of the voice recognition means selected by the method selection means and voice-recognizing the voice information to another apparatus via the communication means.

  According to this voice recognition device, voice information can be voice-recognized by selectively operating one of a plurality of voice recognition means in accordance with the received method specifying information. Therefore, unlike the prior art, it is not necessary to perform voice recognition by all voice recognition means. Therefore, it is possible to provide a speech recognition device that can reduce the amount of processing during speech recognition.

  The parameter calculation apparatus according to the third aspect of the present invention is a speech recognition method having the highest accuracy among a plurality of predetermined speech recognition methods for speech input in an environment where noise may exist. Is a parameter calculation device for calculating a parameter for accuracy prediction, used when determining a method that is predicted to be possible, and a plurality of types of noise data are added to speech data prepared in advance. A plurality of types of learning data and a plurality of types of speech recognition methods based on means for preparing a plurality of learning data obtained by superimposing each other, and results of speech recognition of the plurality of learning data by a plurality of speech recognition methods A speech recognition accuracy calculating means for calculating speech recognition accuracy for each of the combinations, and a predetermined acoustic feature amount of a first dimension number from each of a plurality of types of noise data Input based on the acoustic feature quantity calculation means for calculating the acoustic feature quantity vector, the acoustic feature quantity vector calculated by the acoustic feature quantity calculation means, and the speech recognition accuracy calculated by the speech recognition accuracy calculation means Parameter calculating means for calculating a parameter for predicting accuracy of speech recognition accuracy when the speech signal is speech-recognized by a plurality of speech recognition methods from a predetermined acoustic feature vector extracted from the speech signal including.

  According to this parameter calculation device, the speech recognition accuracy is calculated based on the result of speech recognition of a plurality of learning data prepared in advance by a plurality of speech recognition methods. A parameter for predicting accuracy of speech recognition accuracy when the speech signal is speech-recognized by a plurality of speech recognition methods is calculated from the calculated speech recognition accuracy and acoustic feature vector. Therefore, it is possible to provide a speech recognition device that can predict the accuracy of speech recognition using parameters obtained by the parameter calculation device without a prior speech test.

  The parameter calculation means performs a principal component analysis on the acoustic feature quantity vector obtained from each of the plurality of types of noise data, and uses the acoustic feature quantity vector as an acoustic having a second dimension number smaller than the first dimension number. Means for calculating a conversion function for converting into a feature vector, and an acoustic feature vector obtained from each of a plurality of types of noise using the conversion function as an acoustic feature vector of the second dimension number A set of parameters for accuracy prediction by multiple regression analysis using the speech recognition accuracy calculated by the speech recognition accuracy calculation means as an objective variable and each component of the acoustic feature vector of the second dimension as an explanatory variable. For calculating each of the plurality of voice recognition means.

  According to this parameter calculation device, the principal component analysis is performed on the acoustic feature quantity vector obtained from each of the plurality of types of noise data, and the acoustic feature quantity vector is converted into the second dimension number smaller than the first dimension number. A conversion function for converting to an acoustic feature vector is obtained. Further, the acoustic feature quantity vector obtained from each of the noise data is converted into an acoustic feature quantity vector of the second dimension number using this conversion function, and the speech recognition accuracy calculated by the speech recognition accuracy calculating means is used as the objective variable. Then, a parameter set for accuracy prediction is calculated for each of the plurality of speech recognition means by multiple regression analysis using each component of the acoustic feature vector of the second dimension number as an explanatory variable. Principal component analysis reduces the number of explanatory variables while preserving information, so that multiple regression analysis processing can be simplified and reliable multiple regression coefficients can be obtained. In addition, since the conversion function and the parameter obtained by the multiple regression analysis can be used at the time of predicting the speech recognition accuracy, it is possible to provide a parameter calculation device that can shorten the time required for prediction.

  An information terminal device according to a fourth aspect of the present invention is an information terminal device having a microphone and a communication function, and features a plurality of predetermined acoustic feature quantities as components from sound input via the microphone. A means for calculating a quantity vector, a conversion means for converting a feature quantity vector into a feature quantity vector having a smaller number of dimensions using a pre-calculated transformation matrix, and a feature quantity vector converted by the conversion means, An accuracy predicting means for predicting the accuracy of speech recognition for speech for each speech recognition method by a predetermined calculation between coefficient vectors for accuracy prediction prepared in advance for each of a plurality of speech recognition methods; , Means for selecting a speech recognition method capable of speech recognition with the highest accuracy calculated by the accuracy prediction means, and sound input via a microphone And means for transmitting the information for identifying the voice recognition method selected by the means for selecting to a predetermined destination via the communication function, and voice and voice recognition method from the predetermined destination Processing execution means for receiving a voice recognition result sent back with respect to the specified information via a communication function and executing a predetermined process on the voice recognition result;

  According to this information terminal device, by a predetermined calculation between a feature vector of speech input via a microphone and a coefficient vector for accuracy prediction prepared in advance for each of a plurality of speech recognition methods, The accuracy of speech recognition for speech is predicted for each speech recognition method. A speech recognition method capable of performing speech recognition with the highest accuracy among the predicted speech recognition accuracy is selected and transmitted to the speech recognition apparatus that is the transmission destination together with the speech. Based on the transmitted information, speech recognition can be performed by the most appropriate speech recognition method. It is not necessary to perform speech recognition using other speech recognition methods. Therefore, it is possible to provide a speech recognition apparatus that can reduce the processing required when using speech recognition by a plurality of speech recognition methods.

  Preferably, the information terminal device selects one of a plurality of types of notification methods according to the highest accuracy specified by the means for selecting and notifies the user of the information terminal of the highest accuracy level. Includes level notification means.

  According to this information terminal device, the user can know the level indicating the accuracy that can be expected for voice recognition. Not only the accuracy of speech recognition but the level of accuracy is notified, so that the user can easily know whether speech recognition is available. Therefore, it is possible to provide an information terminal device that can easily inform the user of the accuracy prediction result of voice recognition.

  More preferably, the information terminal device further includes a display device, and the level notification means selects and displays one of a plurality of types of level display symbols on the display device according to the highest accuracy specified by the means for selecting. Level display means for doing this.

  According to this information terminal device, one of a plurality of types of level display symbols is displayed on the display device in accordance with the specified highest accuracy. Since the accuracy that can be expected for speech recognition is informed in an easy-to-understand form called symbols, the user can immediately understand the accuracy of speech recognition. Therefore, it is possible to provide an information terminal device that can notify the user of the accuracy prediction result of speech recognition in a form that can be easily understood immediately.

  When executed by a computer, the computer program according to the fifth aspect of the present invention causes the computer to operate as any of the above-described devices. Therefore, any of the above effects can be brought about.

[Constitution]
FIG. 2 shows a block diagram of the speech recognition system 60 according to the present embodiment. Referring to FIG. 2, voice recognition system 60 includes a server 76 that can perform voice recognition using a plurality of voice recognition methods. The server 76 has a function of receiving information specifying the voice recognition method from the outside together with the voice signal, performing voice recognition using the specified voice recognition method, and returning the result to the voice transmission source. Details of the configuration of the server 76 will be described later.

  The speech recognition system 60 further predicts speech recognition accuracy when recognizing input speech using a plurality of speech recognition methods of the server 76 using parameters calculated in advance, and achieves the highest accuracy speech recognition method. Includes a mobile phone 70 that transmits information for identifying to the server 76 together with voice, and a parameter calculation device 72 for calculating in advance parameters for use in voice recognition accuracy determination in the mobile phone 70. These parameters are once stored in the storage medium 74 and then copied to each of a large number of mobile phones to be produced.

  In this speech recognition system, the parameter calculation device 72 uses parameters that are used for accuracy prediction in the mobile phone 70 by using a large number of learning data composed of speech data in which one type of each of a large number of types of noise data is superimposed. Is calculated. Specifically, the following processing is performed.

  First, speech recognition is performed on each of the large number of learning data using each of a plurality of speech recognition methods performed by the server 76, and the recognition accuracy thereof is calculated. Further, feature quantity vectors composed of acoustic feature quantities for accuracy prediction are extracted from each of the noise data superimposed on the learning data, and principal component analysis is performed on these acoustic feature quantities. A transformation matrix for reducing the number of dimensions of the feature vector is calculated by principal component analysis. Using the calculated transformation matrix, the feature quantity vector obtained from the noise data is transformed into a feature quantity vector having a smaller number of dimensions. Then, from the converted feature vector and the accuracy of the speech recognition result previously performed on the learning data, multiple regression for predicting the accuracy of the speech recognition result by linear summation of the components of the converted feature vector A set of coefficients is calculated for each speech recognition method.

  The transformation matrix thus calculated and the set of multiple regression coefficients obtained for each voice recognition method are parameters calculated by the parameter calculation device 72 and are used for accuracy prediction by the mobile phone 70. In the present embodiment, the set of multiple regression coefficients includes an average of prediction accuracy and a weight for each element of the converted feature vector. In this specification, a set of multiple regression coefficients related to one speech recognition method is referred to as a multiple regression coefficient vector.

  Here, the multiple regression analysis is one of statistical methods for creating a prediction formula that is easy to calculate and has few errors by selecting a plurality of appropriate variables. In the regression analysis to quantitatively analyze how much the dependent variable can be explained by the explanatory variable by applying an equation between the dependent variable (objective variable) and the explanatory variable (independent variable) of the continuous scale, there are two explanatory variables. It refers to the above. In the present embodiment, the dependent variable is the accuracy of speech recognition, and the explanatory variable is a component of the feature vector after conversion. The accuracy can be calculated for each speech recognition method by prior speech recognition. Therefore, a multiple regression coefficient vector is also obtained for each speech recognition method.

In this embodiment, it is assumed that there are J speech recognition methods used by the server 76. The multiple regression coefficient vectors obtained for the _{first to} Jth speech recognition methods are represented by B1 to _BJ , respectively. Further, it is assumed that there are M types of noise data prepared at the time of learning. Therefore, M types of learning data obtained by superimposing these noise data on the speech data for learning are also obtained. One recognition accuracy of one speech recognition method is calculated for one learning data. If there are M types of learning data and J types of speech recognition methods, the number of recognition accuracy obtained during learning is M × J types. This M × J type of recognition accuracy is used as a target when calculating the dependent variable in the multiple regression analysis.

  FIG. 3 shows a block diagram of the mobile phone 70. Referring to FIG. 3, a mobile phone 70 includes a microphone 80 for converting input sound into a sound signal, and a frame for converting the converted sound signal into a frame with a predetermined frame length and a predetermined shift length. , A parameter storage unit 82 for storing parameters used in predicting speech recognition accuracy, a framed speech signal, and parameters stored in the parameter storage unit 82, The speech recognition accuracy of each speech recognition method is predicted, and as a result, the speech recognition method estimated to be the speech recognition accuracy with the highest prediction accuracy is determined, and the input speech is predicted based on the determination result A speech recognition method determination unit 84 for determining a display symbol for representing accuracy, and a speech recognition result text provided from the server 76 is converted into image information. A text display processing unit 98, a text display generated by the text display processing unit 98, a symbol indicating the predicted speech recognition accuracy determined by the speech recognition method determination unit 84, and other information on a screen or the like. And a display unit 86 for displaying.

  The cellular phone 70 further generates a transmission data creation unit 88 for creating data for transmission to be transmitted to the server 76 from the voice signal given from the framing unit 81 and the voice recognition accuracy result predicted by the voice recognition method determination unit 84. And a communication unit 90 for transmitting and receiving data to and from the server 76. In the following description, it is assumed that a voice input is a command for the mobile phone 70.

  The mobile phone 70 further includes a command processing unit 92 for causing the mobile phone 70 to perform a predetermined operation based on the data received by the communication unit 90, and a voice for processing a normal voice call signal received by the communication unit 90. A processing unit 94 and a speaker 96 for converting the output of the audio processing unit 94 into audio are included.

  FIG. 4 shows a block diagram of the speech recognition method determination unit 84. Referring to FIG. 4, voice recognition method determination unit 84 uses the parameters stored in parameter storage unit 82 to convert the voice signal framed by framing unit 81 using a plurality of voice recognition methods of server 76. An accuracy prediction unit 102 for predicting the accuracy of speech recognition when performing speech recognition, and a table showing the relationship between the accuracy and the symbol used when displaying a symbol representing the predicted accuracy on a screen of a mobile phone or the like Display symbol table storage unit 104 for storing a display symbol, and a display symbol determination unit for determining a display symbol by comparing the prediction accuracy in the accuracy prediction unit 102 with the table stored in the display symbol table storage unit 104 106.

  FIG. 5 shows an example of the display symbol table. Referring to FIG. 5, the prediction accuracy is shown in the left frame of the table, and the display symbol is shown in the right frame. The display symbol includes entries 120-126.

  In FIG. 5, for example, if the predicted accuracy of speech recognition is 95% or more, it is determined that the speech recognition is very suitable for speech recognition, and ◎ is displayed on the screen of the mobile phone (entry 120). If the accuracy of speech recognition is 90% or more and less than 95%, it is determined that the speech recognition is suitable, and a circle is displayed on the screen of the mobile phone (entry 122). If the accuracy of voice recognition is not less than 75% and less than 90%, it is determined that the voice recognition is not suitable, and Δ is displayed on the screen of the mobile phone (entry 124). If the accuracy of voice recognition is less than 75%, it is determined that the voice recognition is not suitable, and x is displayed on the screen of the mobile phone (entry 126).

  By viewing these symbols displayed on the screen of the mobile phone, the user can easily know the prediction accuracy of the current speech recognition.

FIG. 6 shows a block diagram of the accuracy prediction unit 102. With reference to FIG. 6, the accuracy prediction unit 102 calculates speech feature amounts X _{1 to} X _N for accuracy prediction from the speech signal framed by the framing unit 81, and calculates these N feature amounts. a feature quantity calculation unit 180 for outputting a feature vector as a component, using the transformation matrix 186 stored in the parameter storage unit 82, feature vector (X ₁ ~X _N) the feature vector (X ₁ Conversion processing unit 182 for converting to “˜X _i ”).

  The number of dimensions i of the feature vector after conversion is smaller than the number of dimensions N before conversion. However, this feature vector is transformed by the transformation matrix obtained by the principal component analysis described above. Therefore, it has enough information especially related to voice conversion.

The accuracy prediction unit 102 further performs speech recognition using the converted feature vector (X ₁ ′ to X _i ′) and the multiple regression coefficient vector 188 obtained by multiple regression analysis stored in the parameter storage unit 82. A prediction accuracy calculation unit 184 using multiple regression coefficients for calculating a predicted value of accuracy for each speech recognition method, and a speech recognition method that determines the accuracy of obtaining the maximum value among the calculated prediction accuracy and gives the accuracy. And a maximum value selection unit 185 for selecting as a voice recognition method in the server 76.

FIG. 7 shows a block diagram of the prediction accuracy calculation unit 184 using multiple regression coefficients. Referring to FIG. 7, the prediction accuracy calculation unit 184 using multiple regression coefficients calculates multiple regression coefficient vectors for predicting the accuracy of the first to Jth speech recognition methods from the parameters stored in the parameter storage unit 82. First to J-th multiple regression coefficient extraction unit 190 for extraction, multiple regression coefficient vector (B _{1 to} B _J ) extracted by multiple regression coefficient extraction unit 190 and feature quantity converted by conversion processing unit 182 First to J-th accuracy calculation units 192 to 198 for calculating prediction accuracy Z _{1 to} Z _J of the _first to J-th speech recognition methods using vectors (X ₁ ′ to X _i ′), Including.

  FIG. 8 shows a block diagram of the parameter calculation device 72. Referring to FIG. 8, parameter calculation device 72 (see FIG. 2) includes first to M-th noise storage units 130 to 134 for storing first to M-th noises, and learning learning prepared in advance. Utterance data storage unit 136 for storing utterance data for the purpose, and correct data storage unit 142 for storing data in which the content of the utterance data is converted into text.

  The parameter calculation device 72 further uses the data stored in the first to Mth noise storage units 130 to 134, the utterance data storage unit 136, and the correct data storage unit 142 to perform voice recognition using a plurality of voice recognition methods. Based on the accuracy calculation unit 138 based on learning data for calculating the accuracy of each, the accuracy calculated by the accuracy calculation unit 138, and the noise stored in the first to Mth noise storage units 130-134 And a parameter calculation unit 140 for calculating parameters used by the prediction unit 102 (see FIG. 6) by principal component analysis and multiple regression analysis.

  FIG. 9 shows a block diagram of the accuracy calculation unit 138 based on learning data. Referring to FIG. 9, the accuracy calculation unit 138 based on learning data learns by superimposing noise stored in the first to Mth noise storage units 130 to 134 on the utterance data stored in the utterance data storage unit 136. It includes a sound synthesizer 150 for creating data and a first learning data storage unit 152 for storing data obtained by superimposing first noise on speech data. The learning data accuracy calculation unit 138 further includes second to Mth learning data storage units 154 to 156 for storing second to Mth learning data obtained in the same manner.

  The accuracy calculation unit 138 based on learning data further includes first to Jth data for recognizing voices stored in the first to Mth learning data storage units 152 to 156 using the first to Jth different methods, respectively. Voice recognition units 158 to 162.

The accuracy calculation unit 138 based on learning data further obtains J × M speech recognition results obtained for the M learning data by the first to Jth speech recognition units 158 to 162, and the J × M Learning for determining whether or not each speech recognition result matches the correct answer data stored in the correct answer data storage unit 142 and calculating the accuracy vector (Y _{1 to} Y _J ) of the speech recognition result by each method An accuracy calculation unit 166 using data is included.

  In FIG. 10, the first to Jth speech recognition units 158 to 162 perform the first learning data to the Mth learning data stored in the first to Mth learning data storage units 152 to 156. The table | surface about the relationship between the 1st-Jth speech recognition system and the precision calculated as a result is shown.

Referring to FIG. 10, the leftmost frame indicates the type of learning data, and the uppermost frame indicates the first to Jth speech recognition methods. In FIG. 10, for example, the accuracy when speech recognition is performed on the first learning data by the first speech recognition method is y ₁₁ %, and speech recognition is performed by the second speech recognition method. Is represented as y ₁₂ %.

  In this way, voice recognition is performed on all of the first to Mth learning data by the first to Jth voice recognition methods.

Of the accuracy vectors (Y _{1 to} Y _J ), which are the vectors representing the accuracy calculated by the accuracy calculation unit 166 based on the learning data, the accuracy vector Y ₁ is (y ₁₁ , y ₂₁ , y ₃₁ ,..., Y _M1 ). The accuracy vector Y _{2 and} below are similarly expressed, and the accuracy vector Y _J is expressed by (y _1J , y _2J , y _3J ..., Y _MJ ).

FIG. 11 shows a block diagram of the parameter calculation unit 140. Referring to FIG. 11, parameter calculation unit 140 extracts a feature amount vector (X _{1 to} X _N ) of noise given from first to Mth noise storage units 130 to 134. 170 for converting the extracted feature vector (X _{1 to} X _N ) into a feature vector (X ₁ ′ to X _i ′) in which the number of dimensions of the feature vector data is reduced by principal component analysis. A principal component analysis unit 172 for obtaining a transformation matrix and giving it to the parameter storage medium 74.

The parameter calculation unit 140 further converts the feature vector (X _{1 to} X _N ) into a feature vector (X ₁ ′ to X _i ′) having a reduced number of dimensions using the transformation matrix obtained by the principal component analysis unit 172. Using the conversion processing unit 174 for conversion, the feature vector (X ₁ ′ to X _i ′), and a plurality of accuracy vectors obtained by the accuracy calculation unit 166 (see FIG. 9) based on learning data, speech recognition is performed. A multiple regression analysis unit 176 for calculating a multiple regression coefficient vector for each method.

  FIG. 12 shows a block diagram of the server 76. Referring to FIG. 12, server 76 separates data transmitted from mobile phone 70 into data specifying voice and a voice input method, and communication unit 210 for exchanging data with mobile phone 70. And a first to a first for recognizing voices by the first to Jth schemes respectively. J speech recognition units 216 to 222 and a switch 214 for switching the connection to one of the first to Jth speech recognition units 216 to 222 based on the data for specifying the voice input method.

  The server 76 further includes a switch 224 for switching connection so as to select an output from a voice recognition unit that performs voice recognition using the specified voice recognition method based on data for specifying the voice recognition method, A transmission data generation unit 226 for generating transmission data for transmitting the recognized data to the mobile phone 70 via the communication unit 210.

  Although not shown, only the speech recognition unit of the specified method operates among the first to Jth speech recognition units 216 to 222 by the data specifying the speech recognition method separated by the separation unit 212, and the other The voice recognition unit does not work.

[Operation]
The apparatus according to the present embodiment operates as follows. The operation will be described with reference to FIGS. First, the parameter used when obtaining the speech recognition accuracy must be calculated by the parameter calculation device 72 in advance.

  In the parameter calculation device 72, first, each of various noises stored in the first to Mth noise storage units 130 to 134 (see FIG. 8) in the utterance data stored in the utterance data storage unit is converted into the sound synthesis unit 150 ( (See FIG. 9).

  The sound data (learning data) obtained by the sound synthesis unit 150 is stored in the first to Mth learning data storage units 152 to 156, respectively. The total M types of learning data stored in the first to Mth learning data storage units 152 to 156 are given to the first speech recognition unit 158. The given M types of learning data are voice-recognized using the first method. Similarly, the M types of superposition learning data are also provided to the second speech recognition unit 160, and speech recognition is performed using the second method. In this way, M kinds of superposition learning data are given to a total of J speech recognition units up to the J-th speech recognition unit 162, and each speech recognition is performed. The result of this speech recognition is output as text.

The data output as text is given to the accuracy calculation unit 166 using learning data. The accuracy calculation unit 166 based on learning data uses the correct answer data stored in the correct answer data storage unit 142 in advance, and the speech recognition accuracy of M × J types of data recognized by the first to Jth methods is high. It is calculated for each method and each learning data. The accuracy of the speech recognition result calculated by the accuracy calculation unit 166 based on the learning data is given to the parameter calculation unit 140 as an accuracy vector (Y _{1 to} Y _J ).

The parameter calculation unit 140 operates as follows. Referring to FIG. 11, first, noise data stored in first to Mth noise storage units 130 to 134 is given to feature amount extraction unit 170. The feature quantity extraction unit 170 extracts acoustic feature quantity vectors (X _{1 to} X _N ) of noise data. Next, the principal component analysis unit 172 performs principal component analysis on the feature quantity vectors (X _{1 to} X _N ) of the noise data to obtain a matrix for feature quantity conversion. The obtained matrix is given to the conversion processing unit 174 and the parameter storage medium 74.

In the conversion processing unit 174, the feature vector (X ₁ ′ to X _i ′) is obtained by converting the feature vector (X _{1 to} X _N ) using the matrix obtained by the principal component analysis. The converted feature vector (X ₁ ′ to X _i ′) is given to the multiple regression analysis unit 176.

The multiple regression analysis unit 176 uses the converted feature vector (X ₁ ′ to X _i ′) and the plurality of accuracy vectors (Y _{1 to} Y _J ) calculated by the accuracy calculation unit 138 to perform speech recognition. Multiple regression coefficient vectors (B _{1 to} B _J ) are calculated for each. The calculated multiple regression coefficient vectors (B _{1 to} B _J ) are stored in the parameter storage medium 74. Here, the multiple regression coefficient vector obtained for the k-th speech recognition method is represented as B _k = (b _k0 , b _k1 ,..., B _ki ) (1 ≦ k ≦ J). . Among these elements, b _k0 is an average value of accuracy, and b _k1 ,..., B _ki are weights for the first to i-th acoustic feature amounts, respectively.

  The parameter calculated by the parameter calculation device 72 (see FIG. 2) and stored in the parameter storage medium 74 is copied to the parameter storage unit 82 (see FIG. 3) of the mobile phone 70, and the mobile phone 70 uses this parameter. Is performed. Referring to FIG. 3, first, voice is input to mobile phone 70. The input voice is converted into a voice signal by the microphone 80. The audio signal is framed by the framing unit 81 with a predetermined frame length and a predetermined shift length.

Next, feature amount vectors (X _{1 to} X _N ) are calculated from the speech signal framed by the feature amount calculation unit 180 (see FIG. 6). The calculated feature vector (X _{1 to} X _N ) is converted into a feature vector (X ₁ ′ to X _i ′) using a conversion matrix obtained by principal component analysis stored in the parameter storage unit 82. Is done. The converted feature vector (X ₁ ′ to X _i ′) is given to the first to Jth accuracy calculation units 192 to 198 (see FIG. 7).

Multiple regression coefficient vectors B _{1 to} B _J are extracted from the parameter storage unit 82 by the multiple regression coefficient extraction unit 190 (see FIG. 7). The extracted multiple regression coefficient vector B ₁ is given to the first accuracy calculator 192. The multiple regression coefficient vector B ₂ is given to the second accuracy calculation unit 194. Thereafter, the same operation is performed for each of the multiple regression coefficient vectors, and the multiple regression coefficient vector B _J is given to the Jth accuracy calculation unit 198.

The first accuracy calculation unit 192 calculates the following from the given feature vector (X ₁ ′ to X _i ′) and the multiple regression coefficient vector B ₁ = (b ₁₀ , b ₁₁ ,..., B _1i ). The predicted speech recognition accuracy Z ₁ is calculated by the following formula.

第２の精度算出部１９４でも同様に、与えられた特徴量ベクトル（Ｘ₁’〜Ｘ_i’）と重回帰係数ベクトルＢ₂＝（ｂ₂₀、ｂ₂₁、・・・、ｂ_2i）とから、予測音声認識精度Ｚ₂が次の式より算出される。

Similarly, in the second accuracy calculation unit 194, from the given feature vector (X ₁ ′ to X _i ′) and the multiple regression coefficient vector B ₂ = (b ₂₀ , b ₂₁ ,..., B _2i ) The predicted speech recognition accuracy Z ₂ is calculated from the following equation.

以下、同様にして特徴量ベクトルと重回帰係数ベクトルとから予測音声認識精度が算出されるという処理が、全ての精度算出部で行なわれる。算出された合計Ｊ個の予測音声認識精度は、最大値選択部１８５に与えられる。

In the same manner, the process of calculating the predicted speech recognition accuracy from the feature vector and the multiple regression coefficient vector is performed in all accuracy calculation units. The calculated total J predicted speech recognition accuracies are given to the maximum value selection unit 185.

  The maximum value selection unit 185 (see FIG. 6) selects the one having the highest prediction accuracy from the given J predicted speech recognition accuracy. The maximum value of the selected prediction accuracy is given to the display symbol determination unit 106. Further, both the information for identifying the voice recognition method giving the maximum prediction accuracy and the voice signal converted from the voice by the microphone 80 are given to the transmission data creating unit 88. In the transmission data creation unit 88 (see FIG. 3), the given data is converted into transmission data. The transmission data is transmitted to the server 76 via the communication unit 90.

  Referring to FIG. 12, the data transmitted to server 76 is provided to separation unit 212 via communication unit 210. The separation unit 212 separates data for specifying a speech recognition method and data related to speech from the given data. Of the separated data, data specifying the voice recognition method is given to the switch 214 and the switch 224.

  When the switch 214 is supplied with data specifying a voice recognition method, the switch 214 switches the connection so that an input is given to a voice recognition unit that uses the method.

  When the data specifying the voice recognition method is also given to the switch 224, the switch 224 gives the output of the voice recognition unit that uses the specified voice recognition method to the transmission data generation unit 226. In the mobile phone 70 (FIG. 3), voice recognition is performed by a voice recognition unit using a method predicted to have the maximum accuracy, and the result is transmitted to the mobile phone 70. The transmitted voice recognition result is given to the text display processing unit 98 via the communication unit 90. The given speech recognition result text is converted into image information and given to the display unit 86.

  Receiving this transmission data, the communication unit 90 gives the received transmission data to the command processing unit 92. Here, the command processing unit 92 interprets the given text data as a command and performs corresponding processing.

  Since the operations of the speech processing unit 94 and the speaker 96 are not directly related to the operations of the speech recognition apparatus according to this embodiment, the description of the operations is omitted.

  According to this speech recognition apparatus, it is possible to specify a speech recognition method that is estimated to provide the best speech recognition accuracy by using parameters and input speech obtained from learning data prepared in advance. The parameters for this estimation are calculated from the acoustic feature vector obtained from the noise so as to predict the speech recognition accuracy for the learning data in which the noise is superimposed on the speech data. Therefore, noise data alone is more suitable for this prediction rather than speech data. Therefore, if speech recognition is performed by this speech recognition method, speech recognition can be performed with high accuracy without a test utterance. For example, during the first period of voice input, it is considered that there is no utterance and only background noise. An optimal speech recognition method can be estimated from the noise portion.

[Realization by computer]
The server 76 of this embodiment is substantially realized by computer hardware, a program executed by the computer hardware, and data stored in the computer hardware. FIG. 13 shows the external appearance of the computer system 330, and FIG. 14 shows the internal configuration of the computer system 330.

  Referring to FIG. 13, the computer system 330 includes a computer 340 having an FD (flexible disk) drive 352 and a CD-ROM (compact disk read only memory) drive 350, a keyboard 346, a mouse 348, and a monitor 342. including.

  14, in addition to the FD drive 352 and the CD-ROM drive 350, the computer 340 includes a CPU (Central Processing Unit) 356 and a bus 366 connected to the CPU 356, the FD drive 352, and the CD-ROM drive 350. And a read only memory (ROM) 358 for storing a boot-up program and the like, and a random access memory (RAM) 360 connected to the bus 366 for storing a program command, a system program, work data, and the like.

  Although not shown here, the computer 340 may further include a network adapter board that provides a connection to a local area network (LAN).

  A computer program for causing the computer system 330 to operate as the server 76 is stored in the CD-ROM 362 or FD 364 inserted in the CD-ROM drive 350 or FD drive 352 and further transferred to the hard disk 354. Alternatively, the program may be transmitted to the computer 340 through a network (not shown) and stored in the hard disk 354. The program is loaded into the RAM 360 when executed. The program may be loaded directly into the RAM 360 from the CD-ROM 362, from the FD 364, or via a network.

  This program includes a plurality of instructions that cause the computer 340 to operate as the server 76 of this embodiment. Some of the basic functions required to perform this operation are provided by operating system (OS) or third party programs running on the computer 340 or various toolkit modules installed on the computer 340. Therefore, this program does not necessarily include all functions necessary to realize the system and method of this embodiment. If this program includes only an instruction for executing the operation as the server 76 described above by calling an appropriate function or “tool” in a controlled manner so as to obtain a desired result, Good. The operation of computer system 330 is well known and will not be repeated here.

  Further, although components such as a microphone and a speaker are necessary, the hardware of the mobile phone 70 is substantially the same as that of the computer 340.

[Evaluation experiment]
FIG. 15 shows the signal recognition accuracy actually obtained by the SN (Signal-to-Noise) ratio, the speech recognition method SS (Spectrum Subtraction), and the parameters obtained by the principal component analysis and the multiple regression analysis described above. The relationship between the predicted accuracy and the correlation is shown in a graph. If the correlation is 1, the speech recognition accuracy actually obtained and the prediction accuracy completely match. Here, if the correlation is 0.8 or more, it is considered that the prediction accuracy is reliable.

  Referring to FIG. 15, the horizontal axis represents the SN ratio and the vertical axis represents the correlation. Legend 260 indicates the ideal number of dimensions, and legend 262 indicates the number of dimensions determined in advance. The ideal number of dimensions refers to the number of dimensions of the feature vector that has the best prediction accuracy of speech recognition accuracy as a result of predicting speech recognition accuracy using feature vectors of various dimensions. .

  From this graph, it can be seen that the smaller the SN ratio, the higher the correlation between the predicted value and the actual measurement value in both the ideal dimension number 260 and the predetermined dimension number 262.

  The correlation is lowest when the S / N ratio is 20 dB. However, even in this case, the correlation is 0.85 or more for both the ideal dimension number 260 and the predetermined dimension number 262. Therefore, it can be said that the predicted value by the above method is reliable in the speech recognition by the system SS.

  FIG. 16 shows the SN ratio, the speech recognition accuracy actually obtained by the speech recognition method ADSR (Advanced Distributed Speech Recognition Front End), and the accuracy predicted using the parameters obtained by the principal component analysis and the multiple regression analysis described above. The relationship with the correlation between is shown in a graph.

  Referring to FIG. 16, the horizontal axis represents the SN ratio and the vertical axis represents the correlation. A legend 230 indicates the ideal number of dimensions, and a legend 232 indicates a predetermined number of dimensions.

  Also from this graph, as in the graph shown in FIG. 15, the correlation between the ideal dimension number 230 and the predetermined dimension number 232 tends to increase as the S / N ratio decreases. Further, when the SN ratio is 10 dB or less, the correlation is 0.8 or more, and thus it can be said that the prediction accuracy is reliable to some extent.

  FIG. 17 is a graph showing the relationship between the SN ratio and the residual standard deviation of the prediction error in the speech recognition method SS. Referring to FIG. 17, the horizontal axis represents the SN ratio and the vertical axis represents the residual standard deviation of the prediction error. Legend 240 indicates the ideal number of dimensions, and legend 242 indicates the number of dimensions determined in advance.

  According to this graph, the residual standard deviation of the prediction error becomes the largest when the SN ratio is 5 dB for both the ideal dimension number 240 and the predetermined dimension number 242. Further, the residual standard deviation of the prediction error at this time is about 6.5% in any number of dimensions.

  Thus, even the highest residual prediction error is about 6.5%, so it can be said that the prediction accuracy for SS has a certain degree of reliability.

  FIG. 18 is a graph showing the relationship between the S / N ratio and the residual standard deviation of the prediction error in the speech recognition method ADSR. Referring to FIG. 18, the horizontal axis represents the SN ratio and the vertical axis represents the residual standard deviation of the prediction error. The legend 250 indicates the ideal number of dimensions, and the legend 252 indicates the predetermined number of dimensions.

  According to this graph, when the SN ratio is 0 dB and −5 dB, the residual standard deviation of the prediction error takes a large value. Its value is about 8%.

  Since the highest prediction error is about 8% in this way, it can be said that the prediction accuracy by ADSR has a certain degree of reliability.

  FIG. 19 is a graph showing the correlation between the actual accuracy of speech recognition by the system SS and the prediction accuracy. This graph uses data when the SN ratio is 5 dB and the correlation is 0.921. Referring to FIG. 19, the horizontal axis represents the prediction accuracy, and the vertical axis represents the actual accuracy. The closer to the straight line drawn on the graph, the closer the prediction accuracy value and the actual accuracy value are. From this graph, it can be seen that there is little variation in prediction accuracy and actual accuracy. Therefore, it is considered that the reliability of the prediction accuracy is high for the speech recognition system SS.

  FIG. 20 is a graph showing the correlation between the actual accuracy of speech recognition by the method ADSR and the prediction accuracy. This graph uses data when the SN ratio is 15 dB and the correlation is 0.749. Referring to FIG. 20, the horizontal axis represents prediction accuracy, and the vertical axis represents actual accuracy. The closer to the straight line drawn on the graph, the closer the prediction accuracy to the actual accuracy. In this graph, although there is some variation, it can be seen that the variation between the prediction accuracy and the actual accuracy is relatively small. Therefore, it can be said that the reliability of the prediction accuracy is high even for the speech recognition method ADSR.

  As described above, in the system according to the present embodiment, a parameter for predicting accuracy for each of a plurality of speech recognition methods is obtained from a plurality of noise data and speech data in the presence of various noises. In actual speech recognition, the accuracy of each method is predicted from the input speech signal using these parameters, and the method with the highest accuracy is selected as the speech recognition method.

  Since it is not necessary to actually perform voice recognition by another voice recognition method, the load on the server can be reduced. Further, since the predicted voice recognition accuracy is displayed in a level that is easy to understand, it is easy for the user to know whether or not voice recognition can be used. Therefore, unreliable speech recognition is not performed, and speech recognition can be easily performed in an environment suitable for speech recognition.

  In the present embodiment, a system composed of three parts, a mobile phone 70, a parameter calculation device 72, and a server 76 operates in order to solve the problem of the present invention. However, this can also be realized by a single system.

  In the present embodiment, accuracy prediction or the like is performed by a mobile phone, but this is not always necessary. For example, an embodiment in which accuracy prediction or the like is performed by another small terminal can be adopted.

  The embodiment disclosed herein is merely an example, and the present invention is not limited to the above-described embodiment. The scope of the present invention is indicated by each claim in the claims after taking into account the description of the detailed description of the invention, and all modifications within the meaning and scope equivalent to the wording described therein are intended. Including.

It is a block diagram which shows the structure of the server 30 of the speech recognition apparatus by a prior art using the prior test utterance. It is a block diagram which shows the structure of the speech recognition system 60 which concerns on this Embodiment. 3 is a block diagram showing a configuration of a mobile phone 70. FIG. 4 is a block diagram showing a configuration of a speech recognition method determination unit 84. FIG. It is a table | surface which shows an example of a display symbol table. 3 is a block diagram illustrating a configuration of an accuracy prediction unit 102. FIG. It is a block diagram which shows the structure of the prediction accuracy calculation part 184 by a multiple regression coefficient. 3 is a block diagram showing a configuration of a parameter calculation device 72. FIG. It is a block diagram which shows the structure of the precision calculation part 138 by learning data. It is a table | surface shown about the relationship between the 1st-Jth speech recognition system performed with respect to 1st learning data-Mth learning data, and the precision calculated as a result. 3 is a block diagram showing a configuration of a parameter calculation unit 140. FIG. 3 is a block diagram showing a configuration of a server 76. FIG. 1 is an external view of a computer system that implements a speech recognition system 60 according to an embodiment of the present invention. It is a block diagram of the computer shown in FIG. A graph showing the relationship between the S / N ratio and the accuracy of speech recognition actually obtained by the speech recognition method SS and the accuracy predicted using the parameters obtained by the principal component analysis and multiple regression analysis described above. is there. A graph showing the relationship between the signal-to-noise ratio and the accuracy of speech recognition actually obtained by the speech recognition method ADSR and the accuracy predicted using the parameters obtained by the principal component analysis and multiple regression analysis described above. is there. It is a graph showing the relationship between SN ratio and the residual standard deviation of the prediction error in the speech recognition method SS. It is a graph showing the relationship between SN ratio and the residual standard deviation of the prediction error in the speech recognition method ADSR. It is a graph showing the correlation with the actual precision and prediction precision of the speech recognition by system SS. It is a graph showing the correlation with the actual precision and prediction precision of the speech recognition by system ADSR.

Explanation of symbols

70 Cellular Phone 72 Parameter Calculation Device 74 Parameter Storage Medium 76 Server 84 Speech Recognition Method Determination Unit 102 Accuracy Prediction Unit 106 Display Symbol Determination Unit 180 Feature Amount Calculation Unit 182 Conversion Processing Unit 184 Prediction Accuracy Calculation Unit 185 Using Multiple Regression Coefficients Maximum Value Selection Units 192, 194, 196, and 198 Accuracy calculation unit 138 Accuracy calculation unit 140 based on learning data Parameter calculation unit 150 Sound synthesis units 152 to 156 Learning data storage units 158 to 162 Speech recognition unit 166 Accuracy calculation unit 170 based on learning data Extraction unit 172 Principal component analysis unit 174 Conversion processing unit 176 Multiple regression analysis unit 212 Separation unit 214 and 224 Switches 216 to 222 Voice recognition unit

Claims (6)

To determine the method that is predicted to be the most accurate speech recognition method among a plurality of predetermined speech recognition methods for speech input in an environment where noise may exist An optimal speech recognition method selection device,
Means for calculating a feature quantity vector having a predetermined plurality of types of acoustic feature quantities as components from input speech;
Conversion means for converting the feature quantity vector into a feature quantity vector having a smaller number of dimensions using a pre-calculated transformation matrix;
The accuracy of speech recognition for the speech is improved by a predetermined calculation between the feature vector after the conversion by the conversion means and the coefficient vector for accuracy prediction prepared in advance for each of the plurality of speech recognition methods. Accuracy prediction means for predicting each speech recognition method;
An optimum speech recognition method selection apparatus including means for selecting a speech recognition method capable of performing speech recognition with the highest accuracy based on the prediction calculated by the accuracy prediction means. A speech recognition apparatus for performing speech recognition using a plurality of speech recognition means,
A communication means for exchanging data with other devices;
For receiving voice information and method specifying information for specifying any of the plurality of voice recognition units from the other device via the communication unit, and for separating the voice information and the method specifying information Separating means;
A plurality of speech recognition means each for performing speech recognition by a specific speech recognition method;
In accordance with the method specifying information, a method selection unit for selectively operating one of the plurality of voice recognition units to perform voice recognition of the voice information;
A voice recognition apparatus including means for returning the output of the voice recognition means selected by the method selection means and voice-recognizing the voice information to the other apparatus via the communication means. When determining a method that is predicted to be the most accurate voice recognition method among a plurality of predetermined voice recognition methods for speech input in an environment where noise may exist A parameter calculation device for calculating a parameter used for accuracy prediction used,
Means for preparing a plurality of learning data obtained by superimposing a plurality of types of noise data on the utterance data prepared in advance;
Based on a result of speech recognition of the plurality of learning data by the plurality of speech recognition methods, for calculating speech recognition accuracy for each of the combination of the plurality of types of learning data and the plurality of types of speech recognition methods Speech recognition accuracy calculating means;
An acoustic feature quantity calculating means for calculating an acoustic feature quantity vector comprising a predetermined acoustic feature quantity of a first dimension from each of the plurality of types of noise data;
The predetermined acoustic feature vector extracted from the input speech signal based on the acoustic feature vector calculated by the acoustic feature calculator and the speech recognition accuracy calculated by the speech recognition accuracy calculator. Parameter calculation means for calculating parameters for predicting the accuracy of speech recognition accuracy when the speech signal is speech-recognized by the plurality of speech recognition methods. The parameter calculation means includes
A principal component analysis is performed on the acoustic feature quantity vector obtained from each of the plurality of types of noise data, and the acoustic feature quantity vector is obtained as an acoustic feature having a second dimension number smaller than the first dimension number. Means for calculating a conversion function for conversion to a quantity vector;
The voice recognition calculated by the voice recognition accuracy calculation means by converting the acoustic feature quantity vector obtained from each of the plurality of types of noise into the second dimension number of the acoustic feature quantity vector using the conversion function. By means of multiple regression analysis with accuracy as an objective variable and each component of the acoustic feature vector of the second dimension as an explanatory variable, a set of parameters for accuracy prediction is assigned to each of the plurality of speech recognition means. The parameter calculation device according to claim 3, further comprising means for calculating. An information terminal device having a microphone and a communication function,
Means for calculating a feature quantity vector having a predetermined plurality of types of acoustic feature quantities as components from speech input via the microphone;
Conversion means for converting the feature quantity vector into a feature quantity vector having a smaller number of dimensions using a pre-calculated transformation matrix;
The accuracy of speech recognition for the speech is improved by a predetermined calculation between the feature vector after the conversion by the conversion means and the coefficient vector for accuracy prediction prepared in advance for each of the plurality of speech recognition methods. Accuracy prediction means for predicting each speech recognition method;
Means for selecting a voice recognition method capable of voice recognition with the highest accuracy calculated by the accuracy prediction means;
Means for transmitting the voice input via the microphone and information identifying the voice recognition method selected by the means for selecting to a predetermined destination via the communication function;
A voice recognition result returned from the predetermined transmission destination to the voice and information specifying the voice recognition method is received via the communication function, and a predetermined process is performed on the voice recognition result. An information terminal device including processing execution means for executing.   A computer program that, when executed by a computer, causes the computer to operate as the device according to any one of claims 1 to 5.

Images