JP2010204175A - Voice learning device and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a voice learning device to learn sound models used for voice recognition. <P>SOLUTION: The voice learning device includes a sound model-with-identifier-generation section 24 which generates a series of sound models with identifier by merging each environmentally-dependent sound model for each of a plurality of utterance environment while an utterance environment identifier is added to each phoneme label of each environmentally-dependent sound model to identify each utterance environment; an utterance environment-parallel-voice recognition section 26 which performs voice recognition in parallel for the learning voice using each environmentally-dependent sound model with the utterance environment identifier being added for each of the plurality of utterance environment, and generates a recognition result; an transcription-with-identifier-creation section 28 which creates transcription of the generated recognition result while the utterance environment identifier is added; and a sound model identification-learning section 29 which performs identification learning of the sound model with identifier using the learning voice and the transcription with identifier. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to speech recognition using the statistic of the acoustic feature quantity of each phoneme as an acoustic model, and in particular, improves the accuracy of speech recognition when a plurality of utterance environments are mixed in one speech segment like a conversational speech. The present invention relates to a speech learning apparatus and a program for improving the learning effect of a plurality of acoustic models.

  In speech recognition using a statistical acoustic model, learning data is required to obtain a statistic of a feature amount that appears in each phoneme. This learning data is a pair of a voice and a transcription corresponding to the voice. By learning this acoustic model separately for each utterance environment such as a plurality of speakers and background sounds, improvement of speech recognition accuracy is expected. Hereinafter, such an acoustic model for each utterance environment is referred to as an utterance environment-dependent acoustic model.

  However, when learning or recognizing conversational speech, there may not be an appropriate silent section between recorded speech utterances, so each speech environment such as gender or multiple speakers It is difficult to divide audio into two. For such speech recognition, gender-dependent acoustic model is used to improve speech recognition accuracy in which speeches of male and female speakers are mixed (for example, Patent Document 1, Patent Document 1). Non-patent document 1).

  On the other hand, the maximum likelihood criterion (ML) that maximizes the likelihood of the learning speech is generally used as the criterion for learning the statistics of the acoustic model, but recently, based on the phoneme error minimization criterion (MPE). A learning method has been proposed (see, for example, Non-Patent Document 2).

  The learning based on MPE has been confirmed to improve the recognition accuracy even in the speech recognition in the above-described gender parallel speech recognition. In the learning based on the MPE, in addition to the speech used for learning the acoustic model and the transcription corresponding to the speech, a recognition result for evaluating the speech recognition error is required.

  It is desirable that this recognition result includes more possible recognition errors. Therefore, in order to efficiently evaluate recognition errors, an algorithm that performs MPE learning using a phoneme lattice obtained during a hypothesis search has also been proposed (for example, see Non-Patent Document 3).

  On the other hand, in order to learn a speech environment (speaker) dependent acoustic model, a technique of adapting a large amount of learned unspecified speaker models is effective (for example, see Non-Patent Document 3). This document also proposes an adaptation technique for linearly converting the parameters of the acoustic model based on the MPE standard. This linear transformation is called DLT (Discriminative Linear Transforms).

JP 2007-233149 A

Imai et al., Proceedings of the Spring Meeting of the Acoustical Society of Japan, 1-1-16, 2008 D. Povay, “Minimum phone error and L-smoothing for improved discriminative training”, in proc. ICASSP, 2002 L. Wang et al., Computer Speech and Language, 22, 2008, pp. 256-272

  In the above-described method, it is desirable to classify the speech segments for each speech environment and learn the acoustic model separately. However, in speech such as a conversation, speech in a plurality of speech environments is mixed in one speech segment. A technique for efficiently learning a plurality of utterance environment dependent acoustic models simultaneously from a speech segment in which a plurality of environments are mixed is required.

  Moreover, in gender parallel speech recognition, it is possible to simultaneously identify an utterance environment, that is, identification of speaker gender, and search for a recognition hypothesis. At this time, a recognition error of the word due to the identification error of the speaker gender may occur. Therefore, in gender parallel speech recognition, a discriminative learning criterion that takes into account speaker discrimination errors is required.

  Therefore, an object of the present invention is to improve the recognition performance of speech recognition when searching for a plurality of acoustic models in parallel by improving the learning performance of a plurality of acoustic models adapted for each utterance environment. It is to provide a learning device and a program.

  The speech learning apparatus according to the present invention provides a speech environment identifier to an acoustic model and a phoneme label of a learned speech, and a plurality of acoustics adapted for each speech environment from speech segments that are automatically cut out in which a plurality of speech environments are mixed. Learn the model at the same time. That is, a plurality of speech environment dependent acoustic models are automatically identified and learned without individually learning the plurality of speech environment dependent acoustic models.

  That is, the speech learning device according to the present invention is a speech learning device that learns an acoustic model used for speech recognition, and each environment-dependent acoustic model for each of a plurality of speech environments is assigned to each phoneme label of each environment-dependent acoustic model. An acoustic model generation unit with an identifier that generates a series of acoustic models with identifiers by merging with a speech environment identifier for identifying each speech environment, and an environment for each of a plurality of speech environments with the speech environment identifier Using each of the dependent acoustic models, an utterance environment parallel speech recognition unit that performs speech recognition in parallel on the learning speech in which the plurality of utterance environments are mixed and generates a recognition result, and an utterance environment identifier in the generated recognition result Using the transcription unit with an identifier to create a transcript with a mark attached, and the learning speech and the transcription with the identifier, the acoustic model with the identifier is discriminated. An acoustic model identification learning unit that learns a plurality of acoustic models with identifiers that are adapted for each utterance environment from learning speech in which utterances of the plurality of utterance environments are mixed. It is characterized by.

  Further, in the speech learning device according to the present invention, the speech environment parallel speech recognition unit uses the environment-dependent acoustic model for each of the plurality of speech environments to which the speech environment identifier is attached to mix the plurality of speech environments. Speech recognition is performed in parallel on the learning speech, and an identifier is automatically attached to the recognition result to generate a recognition result with an utterance environment identifier.

  Further, in the speech learning device according to the present invention, the speech environment identifier is removed from the acoustic model with identifier for each speech environment generated by the acoustic model identification learning unit, and the plurality of speech environment dependencies after learning are removed. A post-learning environment-dependent acoustic model generation unit that generates an acoustic model is further provided.

  In the speech learning apparatus according to the present invention, the acoustic model identification learning unit learns a plurality of acoustic models using a speech environment identifier for each gender or speaker as a speech environment.

  In the speech learning device according to the present invention, the speech environment parallel speech recognition unit generates a hypothesis lattice with an identifier by assigning the speech environment identifier to the hypothesis lattice in speech recognition, and the acoustic model identification learning unit includes: Learning a hypothesis lattice with an identifier from the speech environment parallel speech recognition unit, and using the hypothesis lattice with the identifier, the learning speech, and the transcription with the identifier, utterances of the plurality of speech environments are mixed A plurality of acoustic models with identifiers adapted for each utterance environment are learned from speech.

  According to the present invention, for example, it is possible to perform speech learning collectively without individually learning the utterance environment-dependent acoustic model, and to improve the model accuracy of the post-learning utterance environment-dependent acoustic model as a learning result. it can. This has the further effect of facilitating the expansion of the width of the utterance environment-dependent acoustic model, and reduces recognition errors due to utterance environment identification errors, for example, even for speech recognition in which multiple speakers are mixed. As a result, the speaker can be accurately identified and recognized.

It is a figure which shows the speech learning apparatus of one Example by this invention. It is a flowchart which shows operation | movement of the audio | voice learning apparatus of one Example by this invention. It is a figure which shows the example of the phoneme label used with the speech learning apparatus of one Example by this invention. It is the schematic of the speech recognition apparatus which performs speech recognition of unknown speech using the acoustic model learned with the speech learning apparatus of one Example by this invention. It is a figure which shows the outline | summary of the gender parallel speech recognition which is effective when the audio | voices of several speakers are mixed in one utterance area, such as conversation audio | voice, concerning the audio | voice learning apparatus of one Example by this invention. It is a figure which shows the example of the lattice obtained from the gender parallel speech recognition actually obtained based on the speech learning apparatus of one Example by this invention. (A) shows a label of a learning phoneme to which an utterance environment used for MPE-based identification learning is given, and (b) shows an example of a hypothesis lattice used for MPE-based identification learning.

  Hereinafter, a speech learning apparatus according to an embodiment of the present invention will be described.

[Device configuration]
FIG. 1 shows a speech learning apparatus according to an embodiment of the present invention. The speech learning apparatus 1 of the present embodiment is an apparatus that searches in parallel using a plurality of utterance environment-dependent acoustic models such as gender parallel speech recognition.

  The speech learning apparatus 1 according to the present embodiment includes a control unit 2 and a storage unit 3. The control unit 2 includes an acoustic model input unit 21, a voice input unit 22, a user interface unit 23, an acoustic model generation unit with identifier 24, an utterance environment parallel speech recognition unit 26, a recognition error correction unit 27, and an identifier. A supplementary transcription unit 28, an acoustic model identification learning unit 29, and a post-learning environment-dependent acoustic model generation unit 30 are provided. The storage unit 3 can appropriately store or read data necessary for processing under the control of the control unit 2. The control unit 2 that can be configured as a central processing unit (CPU) is stored in the storage unit 3, for example. Each function can be realized by executing a control program.

  The acoustic model input unit 21 attaches an identifier for identifying each utterance environment (for example, male voice, female voice, another male voice) to a plurality of types of environment-dependent acoustic models prepared in advance to each phoneme label. In this state, each environment-dependent acoustic model is supplied to the speech environment parallel speech recognition unit 26 and also supplied to the acoustic model generation unit 24 with an identifier. For example, a male voice acoustic model can be a first environment dependent acoustic model, a female voice acoustic model can be a second environment dependent acoustic model, and another male acoustic model can be a third environment dependent acoustic model, More environment-dependent acoustic models such as background sounds can be handled.

  Corresponding to the input of each environment-dependent acoustic model, the speech input unit 22 inputs learning speech in which a plurality of utterance environments are mixed, supplies the speech to the utterance environment parallel speech recognition unit 26, and acoustic model identification learning described later. This is also input to section 29.

  The user interface unit 23 is an interface that supplies correction information to the recognition error correction unit 27 in order to arbitrarily correct the identification result of the speech environment parallel speech recognition unit 26.

  The acoustic model generation unit with an identifier 24 outputs each utterance environment (for example, the above-described male voice, female voice, and another male voice) to each environment-dependent acoustic model supplied from the acoustic model input unit 21. An identifier for identification is merged with each phoneme label attached (FIG. 3 to be described later) to generate a series of acoustic models with identifiers. Note that the order of the environment-dependent acoustic models may be arbitrarily merged.

  The speech environment parallel speech recognition unit 26 has a function of recognizing a plurality of speech environments in parallel such as gender parallel speech recognition, for example, and a plurality of types of environment-dependent acoustic models supplied via the acoustic model input unit 21. The speech recognition is performed in parallel on the learning speech including a plurality of speech environments supplied via the speech input unit 22 with the speech environment identifier attached to each phoneme label of the learning speech (described later). Figure 8). Therefore, the acoustic model identification learning unit 29 learns a plurality of acoustic models with identifiers adapted for each utterance environment from learning speech in which utterances of the plurality of utterance environments are mixed. In order to automatically give each phoneme a speech environment identifier (for example, gender-specific identifier) of the speech environment (for example, gender for each speaker) of each word in the recognition result, as shown in FIG. From the speech environment parallel speech recognition unit 26, a recognition result with an identifier is obtained and a hypothesis lattice with an identifier is obtained. When the hypothesis lattice with an identifier is used for discrimination learning described later, it is possible to further reduce discrimination errors in the speech environment.

  The recognition error correction unit 27 optionally corrects the identification result of the utterance environment parallel speech recognition unit 26 according to the correction information supplied from the user interface unit 23, and transcribes the corrected recognition result with an identifier with an identifier. Supplied to the unit 28.

  Based on the recognition result with identifier supplied (corrected) supplied from the recognition error correction unit 27, the transcription unit with identifier 28 creates a transcription with the utterance environment identifier attached to the generated recognition result.

  The acoustic model identification learning unit 29 discriminates and learns an acoustic model with an identifier using the learning speech supplied via the speech input unit 22 and the transcription with identifier supplied via the transcription unit with identifier 28. . For this discriminative learning, discriminative learning using a phoneme error minimization criterion (MPE) is effective as will be described later. The acoustic model identification learning unit 29 uses a hypothetical lattice with an identifier (for example, a lattice in which male and female recognition hypotheses are mixed) obtained by the speech environment parallel speech recognition unit 26, and performs identification learning in consideration of speaker identification. Is preferably performed. Thereby, the acoustic model can be optimized for parallel speech recognition.

  The post-learning environment-dependent acoustic model generation unit 30 removes the utterance environment identifier from the acoustic model with the identifier for each utterance environment obtained from the acoustic model identification learning unit 29, and the plurality of utterance environments after learning. Generate a dependent acoustic model. For example, a first environment-dependent acoustic model, a second environment-dependent acoustic model, and a third environment-dependent acoustic model after learning are generated and transmitted. The generated utterance environment-dependent acoustic model after learning can also be stored in the storage unit 3.

  Next, the operation of the speech learning apparatus according to the embodiment of the present invention will be described in detail. FIG. 2 is a flowchart showing the operation of the speech learning apparatus according to an embodiment of the present invention.

[Device operation]
In step S1, the acoustic model input unit 21 inputs an identifier for identifying each utterance environment with respect to a plurality of types of environment-dependent acoustic models prepared in advance in a state where each phoneme label is attached. Corresponding to the input of each environment-dependent acoustic model, a learning speech in which a plurality of utterance environments are mixed is input.

  In step S2, the acoustic model generation unit with identifier 24 merges each input environment-dependent acoustic model with an identifier for identifying each utterance environment attached to each phoneme label, and adds a series of identifiers. Generate an acoustic model. Further, learning in which a plurality of utterance environments input through the voice input unit 22 are mixed using a plurality of types of environment-dependent acoustic models supplied through the acoustic model input unit 21 by the utterance environment parallel voice recognition unit 26. Speech recognition is performed in parallel on speech, and a recognition result with an identifier is generated.

  FIG. 3 is a diagram illustrating an example of a phoneme label used in the speech learning apparatus according to an embodiment of the present invention. As shown in FIG. 3, the state (circle) can be represented by a transition (arrow) connecting the state (circle) as shown in the center of the HMM 3. Also, as shown in the right of FIG. 3, a conventional phoneme label can be expressed as “m + a” for “phone” of a phoneme, for example, “m + a” for “mass” of continuous phonemes. ”,“ M−a + s ”,“ m−s + u ”, and“ su ”. On the other hand, as shown on the left side of FIG. 3, the phoneme label with an identifier for identifying the speech environment of the present embodiment is used to identify, for example, men and women who are two speech environments. “M_” and “F_” are attached respectively. For example, a phoneme label with an identifier can be represented as “M_m + a”, “M_m−a + s”, “M_m−s + u”, and “M_s−u” for male phonemes.

  When learning an acoustic model, it is desirable that the speech to be recognized and the utterance extraction criteria of the learning data are the same. For this reason, in the speech learning apparatus 1 of the present embodiment, it is possible to use the result of recognizing the learning speech by gender parallel speech recognition, correct the recognition error if necessary, and create appropriate learning data.

  In addition, since the speech learning apparatus 1 according to the present embodiment knows which phoneme network each word of the recognition result has passed, it automatically assigns the sex of the man and woman as the speech environment identifier. FIG. 3 shows a phoneme label with an identifier for learning in the case where utterances of male voice (M_) and female voice (F_) are mixed. Furthermore, in the acoustic model (Triphone HMM) and HMM, Similar to the learning phoneme label, the utterance environment identifier is given to the definition of each shared state and transition probability, the male voice HMM and the female voice HMM are merged, and the learning label with the identifier is used at one time. Adaptive learning can be performed collectively.

  Referring to FIG. 2 again, in step S3, the utterance environment parallel speech recognition unit 26 optionally has a hypothesis with an utterance environment identifier in order to further reduce recognition errors caused by parallel processing of multiple utterance environments. Generate a lattice.

  In step S4, if the operator of the apparatus refers to the recognition result of the utterance environment parallel speech recognition unit 26 and determines that correction is required, the utterance is made via the user interface unit 23 in step S5. The identification result of the environment parallel speech recognition unit 26 can be arbitrarily modified.

  In step S6, the transcription unit with identifier 28 creates a transcript of the recognition result with identifier based on the recognition result with identifier supplied (corrected) from the recognition error correction unit 27.

  In step S7, the learning speech supplied via the speech input unit 22 by the acoustic model identification learning unit 29, the transcription with identifier supplied through the transcription unit with identifier 28, and speech environment parallel speech recognition Using the hypothesis lattice with identifier obtained by the unit 26, the acoustic model with identifier is identified and learned.

  In step S8, the speech environment identifier is removed from the learned speech environment dependent acoustic model with identifier generated by the acoustic model identification learning unit 29 and stored in the storage unit 3.

  As described above, when learning the acoustic model, a transcript corresponding to the learning speech with, for example, a gender utterance environment identifier is generated. To create a transcript with an identifier, the result of recognition in parallel using a learning male and female acoustic model is used. Although the transcription can be easily created by using the parallel recognition result, it is also possible to transcribe all the utterance contents by hand. Since the recognition result shows the gender of the speaker of each word in the recognition result, a gender identifier can be automatically assigned to each phoneme. In addition, the speech learning apparatus 1 according to the present embodiment corrects the error in the recognition result with the identifier as necessary to create a transcript with the identifier.

  Furthermore, in the speech learning apparatus 1 of the present embodiment, an acoustic model with an identifier that is a learning base is created by automatically assigning a gender-specific identifier to each phoneme of the male and female acoustic models. .

  Furthermore, the speech learning apparatus 1 according to the present embodiment can discriminate and learn an acoustic model with an identifier using a learning speech and a created transcript with an identifier, and a mixture of male and female recognition hypotheses obtained by parallel speech recognition. Discriminative learning is performed in consideration of speaker identification. Thereby, the acoustic model can be optimized for parallel speech recognition.

  Moreover, the speech learning apparatus 1 according to the present embodiment generates an acoustic model of a phoneme label obtained by removing the identifier from the acoustic model with identifier after learning the acoustic model with identifier. Using this learned male and female acoustic model, unknown speech can be recognized in parallel.

  For example, FIG. 4 shows a schematic diagram of a speech recognition device that performs speech recognition of unknown speech using an acoustic model learned by a speech learning device according to an embodiment of the present invention. Similar to the speech learning device 1, the speech recognition device 101 includes a control unit 102 and a storage unit 103. The control unit 102 includes an acoustic model input unit 121, a speech input unit 122, and an utterance environment parallel speech recognition unit (gender parallel speech recognition unit) 126. The acoustic model input unit 21 and the speech input unit 22 described above, respectively. If the speech recognition device 101 and the speech learning device 1 are configured as a single device, the control unit 102 and the storage unit 103 are respectively connected to the speech environment parallel speech recognition unit 26. It can also be configured as the control unit 2 and the storage unit 3 described above.

  The acoustic model input unit 121 inputs each learned environment-dependent acoustic model stored in the storage unit 3 and supplies it to the speech environment parallel speech recognition unit (gender parallel speech recognition unit) 126.

  Corresponding to the input of each environment-dependent acoustic model, the speech input unit 122 inputs unknown speech in which a plurality of utterance environments are mixed, and supplies the utterance environment parallel speech recognition unit (gender parallel speech recognition unit) 126.

  The utterance environment parallel speech recognition unit (gender parallel speech recognition unit) 126 has a function of recognizing a plurality of utterance environments in parallel, such as gender parallel speech recognition, and is supplied via the acoustic model input unit 121. Using a kind of environment-dependent acoustic model, speech recognition is performed in parallel on a learning speech in which a plurality of speech environments supplied via the speech input unit 122 are mixed, and a recognition result is obtained.

  Next, the speech environment parallel speech recognition unit (gender parallel speech recognition unit) 126 will be briefly described as an example of the speech environment parallel speech recognition unit 26. For details of gender parallel speech recognition, see, for example, Patent Document 1 and Non-Patent Document 1.

[Gender parallel speech recognition]
FIG. 5 shows an outline of gender parallel speech recognition that is effective when voices of a plurality of speakers are mixed in one utterance section, such as a dialogue voice. In gender parallel speech recognition, phoneme networks of word pronunciation dictionaries linked to gender-dependent acoustic models are parallelized, and searches are performed while allowing gender attributes to be replaced at word boundaries.

  The speech environment parallel speech recognition unit (gender parallel speech recognition unit) 126 includes a speech detection / gender change control unit 126a. When the recognition starts, the utterance detection / gender change control unit 126a performs gender parallel phoneme recognition that allows transition between men and women and also uses the same pruning method, and uses the cumulative phoneme likelihood to quickly detect the beginning and end of the utterance. The speaker attribute change time is identified based on the result.

  As shown in Fig. 5, gender parallel large vocabulary continuous speech recognition is possible, and cumulative phoneme likelihood is used. Then, the word string of the recognition result is output.

  Specifically, when speech recognition is started (S12), phoneme environment-dependent sound learned from male speaker speech in various acoustic environments using the feature vectors of the input speech as cepstrum, short-time power, and dynamic features thereof. A phoneme network that forms words as shown in FIG. 5 is configured from a model (Triphone) and a female acoustic model learned in the same manner using a word bigram. Here, in the phoneme network of the female utterance environment acoustic model, a word bigram (S14a) is constructed between silences (S13a, S15a) using the time of the utterance start and utterance ends, and the utterance environment for men. In a phoneme network of an acoustic model, a word bigram (S14b) is constructed between silences (S13b, S15b) using the time of utterance start and end of utterance, enabling transition between men and women, Output (S16).

  Further, according to the speech environment parallel speech recognition unit (gender parallel speech recognition unit) 126, attribute information can be assigned to each word of the recognition result (hypothesis).

  On the other hand, since the speech environment identifier is attached to the phoneme label in the recognition result at the time of learning in the speech learning apparatus 1 of the present embodiment, the maximum likelihood word hypothesis sequence and each word pass through any phoneme network. It is possible not only to obtain a speaker attribute indicating whether it has been recognized, but also to obtain a search path that remains without being pruned as a lattice with an identifier. This lattice with identifier is used for acoustic model discrimination learning. For example, FIG. 6 shows an example of a lattice obtained from gender parallel speech recognition actually obtained, according to the speech learning apparatus of one embodiment of the present invention. FIG. 6 also shows hypothesized phonemes and speaker attributes indicating which gender phoneme network has passed through each branch of the arc connecting each numbered node, and is prefixed with each phoneme label. M_ ”and“ F_ ”are utterance environment identifiers indicating the sexes of men and women, respectively, and can also represent speaker attributes.

  Next, identification learning in the acoustic model identification learning unit 29 will be described.

[Identification learning]
For identification learning in the acoustic model identification learning unit 29, identification learning using a phoneme error minimization criterion (MPE) is effective (for example, see Non-Patent Document 1). In MPE-based discriminative learning, the a posteriori probability of each branch of the phoneme lattice obtained from the recognition result is calculated, and the acoustic model parameters are estimated so as to reduce the expected value of the phoneme recognition error. Data necessary for this learning is a learning speech, a correct phoneme sequence (recognition result) corresponding to the speech, and a hypothesis lattice of phonemes as shown in FIG.

  In order to learn an utterance environment (speaker) dependent acoustic model, it is effective to adapt a large number of learned acoustic models of unspecified speakers (for example, see Non-Patent Document 3).

  The speech learning apparatus 1 according to the present embodiment gives an utterance environment identifier to a phoneme label of speech data used for learning an acoustic model, and also gives a similar utterance environment identifier to an acoustic model of each phoneme, and also an acoustic model of each phoneme. Are also given the same utterance environment identifier, and simultaneously learn acoustic models corresponding to a plurality of utterance environments from a speech segment in which a plurality of utterance environments are mixed.

  Further, the speech learning apparatus 1 according to the present embodiment creates a plurality of utterance environment-dependent acoustic models by merging (integrating) them using the utterance environment identifier when creating the acoustic model given the utterance environment identifier. Therefore, a plurality of acoustic models can be learned at once (collectively) using this acoustic model.

  In addition, the speech learning apparatus 1 according to the present embodiment gives a gender or speaker-specific identifier as an utterance environment identifier, corrects the recognition result of gender parallel speech recognition if necessary, and sets the phoneme label of the learned speech. Therefore, it is easy to determine the accuracy of the correct phoneme sequence corresponding to the learning speech.

  In order to introduce such discriminative learning into the learning of the acoustic model using the speech environment identifier, it is preferable that the speech learning apparatus 1 of the present embodiment uses a recognition hypothesis lattice including a recognition error. Hypothesis lattices can be generated by simultaneously searching for hypotheses obtained from a plurality of acoustic models such as gender-parallel speech recognition and acquiring lattices at the same time. Since each phoneme on this lattice is given a speech environment identifier to generate a hypothesis lattice, recognition errors of phonemes with different speech environments can be identified as model errors, so that recognition errors in the speech environment are reduced. It is possible to learn the statistics of the acoustic model.

  More specifically, a hypothesis lattice obtained from speech environment parallel speech recognition as shown in FIG. 6 is used for identification learning by the acoustic model identification learning unit 29. The hypothesis lattice obtained from the speech environment parallel speech recognition unit 26 is also given a speech environment identifier through which the hypothesis passes. For example, when performing parallel speech recognition by gender, the same speech environment identifier as the phoneme label with identifier shown in FIG. 3 is assigned to the hypothesis lattice. In the conventional gender parallel speech recognition, as can be seen from the example shown in FIG. 5, there is a mixture of gender speech in the temporary lattice, and this gender phoneme error may cause a word recognition error. Keep in mind.

  FIG. 7A shows a label of a learning phoneme provided with an utterance environment used for MPE-based discriminative learning, and FIG. 7B shows an example of a hypothesis lattice used for MPE-based discriminative learning. According to the present embodiment, at the time of MPE-based identification learning, identification learning can be performed in consideration of gender identification errors in addition to phoneme errors. The learning phoneme considering the speech environment is shown in the upper part of the branch of the lattice in FIG. 7B, and the correct accuracy of the phoneme of each branch is shown in the lower part of the branch of the lattice.

  In the discriminative learning, a branch having a correct phoneme accuracy of 1.0 is learned so as to have a high likelihood, and a phoneme having a correct answer accuracy of 0.0 or −1.0 is learned to have a low likelihood. It is an example. For example, when evaluating the phoneme replacement and deletion errors as the correct accuracy of the phonemes, the phoneme correct accuracy takes a value of 0.0 to 1.0, and when evaluating the phoneme insertion errors as the correct accuracy of the phonemes, The phoneme correct answer accuracy is learned to take a value of -1.0 to 1.0.

  For example, in the example of FIG. 7B, the paths “F_m, F_a, F_s, F_u, F_sp” and “M_t, M_a, M_d, M_a” are not phoneme errors, but the genders of men and women are incorrect. The accuracy of correct phonemes. 0.0 is given, that is, learning is performed so that the likelihood is low with respect to the feature vector representing the observed feature amount. In this way, it is possible to learn an acoustic model having high ability to discriminate phonemes of men and women.

  By using this learned acoustic model, for example, it is possible to perform speech learning collectively without individually learning the utterance environment dependent acoustic model, and the model accuracy of the post-learning utterance environment dependent acoustic model as a learning result Can be increased. This further increases the effect of facilitating the expansion of the utterance environment-dependent acoustic model, and reduces recognition errors caused by utterance environment identification errors even for voice recognition with multiple speakers. As a result, the speaker can be accurately identified and recognized.

  An example of adaptation processing in acoustic model identification learning will be described.

  As described above, according to the speech learning apparatus of the present embodiment, when the acoustic model is adapted by introducing the DLT estimated based on the MPE standard, the adaptation is not performed as in the present embodiment (MLLR). ), The word recognition rate (WER) is improved for a plurality of types of phoneme classes by gender.

  Moreover, as one aspect of the present invention, the speech learning apparatus 1 can be configured as a computer, and an acoustic model input unit 21, a speech input unit 22, a user interface unit 23, an identifier-equipped acoustic model generation unit 24, A program for realizing the functions of the speech environment parallel speech recognition unit 26, the recognition error correction unit 27, the transcription unit with identifier 28, the acoustic model identification learning unit 29, and the post-learning environment-dependent acoustic model generation unit 30 is as follows. These are stored in the storage unit 3 provided inside or outside each computer. Information and data used for each control can be stored in the storage unit. Such a storage unit can be realized by an external storage device such as an external hard disk or an internal storage device such as ROM or RAM. The control unit that executes the program can be realized by a central processing unit (CPU) or the like. That is, the CPU can appropriately read from the storage unit a program in which the processing content for realizing the function of each component is described, and implement each device on the computer. Here, the function of any means may be realized by all or part of the hardware.

  In the embodiment described above, the program describing the processing content for realizing the function of the speech learning apparatus 1 can be recorded on a computer-readable recording medium. As the computer-readable recording medium, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording device, and a semiconductor memory may be used.

  The speech learning apparatus 1 may be input via a portable recording medium such as a DVD or CD-ROM, or may be input via an interface dedicated to video input.

  The speech learning apparatus 1 of the above-described embodiment has been described as a representative example, but it will be apparent to those skilled in the art that many changes and substitutions can be made within the spirit and scope of the present invention. Accordingly, the invention should not be construed as limited by the embodiments described above, but only by the claims.

  According to the present invention, it is possible to perform speech learning collectively without individually learning the utterance environment-dependent acoustic model, and to improve the model accuracy of the post-learning utterance environment-dependent acoustic model as a learning result. Useful for any speech recognition application.

DESCRIPTION OF SYMBOLS 1 Speech learning apparatus 2 Control part 3 Memory | storage part 21 Acoustic model input part 22 Speech input part 23 User interface part 24 Acoustic model generation part with identifier 26 Speech environment parallel speech recognition part 27 Recognition error correction part 28 Transcript part with identifier 29 Acoustic Model identification learning unit 30 Post-learning environment-dependent acoustic model generation unit 101 Speech recognition device 102 Control unit 103 Storage unit 121 Acoustic model input unit 122 Speech input unit 126 Speech environment parallel speech recognition unit (gender parallel speech recognition unit)
126a Utterance detection / gender change control unit

Claims (5)

A speech learning device for learning an acoustic model used for speech recognition,
Each of the environment-dependent acoustic models for each utterance environment is merged with each phoneme label of each environment-dependent acoustic model with an utterance environment identifier for identifying each utterance environment, and a series of acoustic models with identifiers is created. An acoustic model generation unit with an identifier to be generated;
Using each of the environment-dependent acoustic models for each of a plurality of utterance environments to which the utterance environment identifier is attached, the speech recognition is performed in parallel on the learning speech in which the plurality of utterance environments are mixed, and an utterance environment parallel that generates a recognition result A voice recognition unit;
A transcription unit with an identifier for creating a transcript with the utterance environment identifier attached to the generated recognition result;
An acoustic model identification learning unit for identifying and learning an acoustic model with an identifier using the learning speech and the transcript with the identifier,
The acoustic model identification learning unit learns a plurality of acoustic models with identifiers adapted for each utterance environment from learning speech in which utterances of the plurality of utterance environments are mixed.   The speech environment parallel speech recognition unit performs speech recognition in parallel on the learning speech in which the plurality of speech environments are mixed, using each of the environment dependent acoustic models for each of the plurality of speech environments with the speech environment identifier. The speech learning apparatus according to claim 1, wherein a speech recognition identifier is automatically added to the recognition result to generate a recognition result with the identifier.   A post-learning environment-dependent sound for generating a plurality of utterance environment-dependent acoustic models after learning by removing the utterance environment identifier from the acoustic model with an identifier for each utterance environment generated by the acoustic model identification learning unit The speech learning apparatus according to claim 1, further comprising a model generation unit.   The said acoustic model identification learning part learns a several acoustic model using the speech environment identifier for every man and woman or speaker as speech environment, It is characterized by the above-mentioned. Voice learning device. The speech environment parallel speech recognition unit generates a hypothesis lattice with an identifier by assigning the speech environment identifier to the hypothesis lattice in speech recognition,
The acoustic model identification learning unit acquires a hypothesis lattice with an identifier from the speech environment parallel speech recognition unit, and uses the hypothesis lattice with the identifier, the learning speech, and the transcript with the identifier, The speech learning apparatus according to any one of claims 1 to 4, wherein a plurality of acoustic models with identifiers adapted for each utterance environment are learned from learning speech in which utterances in the utterance environment are mixed. .

Images