JP6376486B2 - Acoustic model generation apparatus, acoustic model generation method, and program - Google Patents

Description

  The present invention relates to an acoustic model generation apparatus that generates an acoustic model used for speech recognition processing, for example.

  Conventionally, there has been an acoustic model generation apparatus that can perform gender identification without being affected by non-speech sections included in data to be identified (see, for example, Patent Document 1).

  Further, in the conventional speech recognition system, the accuracy of speech recognition has deteriorated due to a mismatch between the sound for generating the acoustic model and the speech uttered by the speech recognition system. In other words, the voices that can be prepared for generating the acoustic model are limited, such as voices recorded in the recording room and voices of reading the original. In order to solve this problem, acoustic model adaptation technology and the like existed.

  The acoustic model adaptation technology uses the speech accumulated in the speech recognition system (also referred to as “speech uttered in the speech environment”) to adapt the acoustic model generated from the speech recorded in the recording room. And generating an acoustic model that can be used for speech recognition processing in various environments (see, for example, Non-Patent Document 1). Note that the voices accumulated in the voice recognition system described above are various voices, and include, for example, voices mixed with background noise and voices with different utterance styles. Note that the adaptive process is a known technique, and thus detailed description thereof is omitted.

JP 2013-57789 A (first page, FIG. 1 etc.)

Sugaya et al., “National speech translation demonstration experiment and model adaptation of speech recognition using actual usage data”, IEICE Transactions D, Vol J96-D, No.1, pp.209-220

  However, the conventional acoustic model adaptation technology requires speech uttered in an utterance environment, and this approach cannot be used in languages without speech accumulation, and an acoustic model that improves speech recognition accuracy could not be generated. .

  The present invention has been made to solve such problems, and generates an acoustic model that is a language in which voice data does not exist in a suitable environment such as a speech environment and can improve voice recognition accuracy. It is an object.

  The acoustic model generation device according to the first aspect of the present invention is an acoustic model generation device that generates an acoustic model of a target language for speech recognition, and can store a target language new acoustic model that is a second acoustic model of the target language. Target language new acoustic model storage unit, one or more other language old acoustic models that are first acoustic models of one or more other languages different from the target language, and one or more second acoustic models of each other language 1 or more first correlation information that is information relating to the relationship with one or more other language new acoustic models, or one or more of each other language old acoustic model and the target language that is the first acoustic model of the target language Using one or more correlation information of one or more second correlation information, which is information related to the relationship with the acoustic model, the target language new from the target language old acoustic model or one or more other language new acoustic models. Generate an acoustic model And acoustic model generator, an acoustic model generating apparatus comprising an acoustic model storage unit for storing the target language new acoustic model acoustic model generator has generated the target language new acoustic model storage unit.

  With this configuration, even in a language that does not have voice data in a suitable environment such as a speech environment, an acoustic model corresponding to the language can be generated, and an acoustic model that improves voice recognition accuracy can be generated.

  Further, in the acoustic model generation device of the second invention, in contrast to the first invention, the acoustic model generation unit includes a target language old acoustic model storage unit that can store the target language old acoustic model, and one or more first acoustic models. The target language new acoustic model is obtained from the target language old acoustic model stored in the target language old acoustic model storage unit using the first correlation information storage unit capable of storing one correlation information and the one or more first correlation information. An acoustic model generation device comprising an acoustic model generation means for generating.

  With this configuration, by utilizing the correlation between the old acoustic model of another language and the new acoustic model of another language, even in languages where there is no speech data in a suitable environment such as an utterance environment, the acoustic model corresponds to the language. It is possible to generate an acoustic model that improves voice recognition accuracy.

  The acoustic model generation device according to the third aspect of the present invention is different from the first aspect in that the acoustic model generation unit includes an other language new acoustic model storage unit that can store one or more other language new acoustic models, and 1 From the second correlation information storage unit capable of storing the second correlation information and one or more other language new acoustic models stored in the other language new acoustic model storage unit using the one or more second correlation information. An acoustic model generation apparatus including an acoustic model generation unit that generates a target language new acoustic model.

  With such a configuration, by utilizing the correlation between the old acoustic model of another language and the new acoustic model of the target language, the acoustic model corresponding to the language can be used even in a language where there is no speech data in a suitable environment such as a speech environment. It is possible to generate an acoustic model that improves voice recognition accuracy.

  Further, in the acoustic model generation device according to the fourth aspect of the invention, in contrast to the first aspect, the acoustic model generation unit includes a target language old acoustic model storage unit that can store the target language old acoustic model, and one or more other acoustic model generation units. Another language new acoustic model storage unit that can store a language new acoustic model, a first correlation information storage unit that can store one or more first correlation information, and a second correlation that can store one or more second correlation information Using the information storage unit, the one or more first correlation information and the one or more second correlation information, the target language old acoustic model, or one or more other language new acoustic models, or the target language old acoustic model and one or more And an acoustic model generation means for generating a target language new acoustic model from another language new acoustic model.

  In such a configuration, by using the correlation between the old acoustic model of another language and the new acoustic model of another language, and the correlation between the old acoustic model of the other language and the old acoustic model of the target language, Even a language in which no voice data exists is an acoustic model corresponding to the language, and an acoustic model that improves voice recognition accuracy can be generated.

  The acoustic model generation device according to the fifth aspect of the present invention is different from the fourth aspect in that the acoustic model generation unit is a target language old acoustic model or one of two or more algorithms for generating a target language new acoustic model. According to the data possessed by the other language new acoustic model described above, further comprising a selection means for selecting any one algorithm, the acoustic model generation means, the first correlation information and the first correlation information according to the one algorithm selected by the selection means The acoustic model generation device generates a target language new acoustic model using one or more correlation information of the second correlation information.

  With this configuration, by using the correlation between the old acoustic model of the other language and the new acoustic model of the other language, and the correlation between the old acoustic model of the other language and the old acoustic model of the target language, using an algorithm suitable for the target data. Even in a language where there is no speech data in a suitable environment such as a speech environment, an acoustic model corresponding to the language can be generated, and an acoustic model that improves speech recognition accuracy can be generated.

  The acoustic model generation device according to the sixth aspect of the present invention relates to the first to fifth aspects of the invention, wherein the other language old acoustic model is an acoustic model before the adaptation processing of another language or another language new acoustic model. Is an acoustic model generated with different data, and the other language new acoustic model is an acoustic model generated with data different from the acoustic model after adaptation processing of other languages or the other language old acoustic model, and the target language old acoustic model Is an acoustic model generated with data similar to the acoustic model before adaptation processing of the target language or other language old acoustic model, and the new acoustic model of the target language is an acoustic model generation that is an acoustic model after adaptation processing of the target language Device.

  With this configuration, even in a language that does not have voice data in a suitable environment such as a speech environment, an acoustic model corresponding to the language can be generated, and an acoustic model that improves voice recognition accuracy can be generated.

  Further, the acoustic model generation device of the seventh aspect of the invention relates to the sixth aspect of the invention, wherein the first correlation information includes one or more vectors and one or more each corresponding to one or more other language old acoustic models. It is information acquired from one or more conversion functions that are differences from one or more vectors corresponding to other language new acoustic models, and the second correlation information is a vector corresponding to one or more other language old acoustic models. And an acoustic model generation apparatus which is information acquired from one or more conversion functions of a difference between a vector corresponding to an old acoustic model of a target language.

  Further, in the acoustic model generation device of the eighth invention, in contrast to the seventh invention, the acoustic model generation unit maps a vector corresponding to the target language old acoustic model using a conversion function of the first correlation information. To generate a new acoustic model of the target language, or by mapping from one or more other language new acoustic models stored in the other language new acoustic model storage unit using the conversion function of the second correlation information An acoustic model generation apparatus that generates an acoustic model.

  The acoustic model generation apparatus according to the present invention can generate an acoustic model that improves speech recognition accuracy, even in a language that does not have voice data in a suitable environment such as a speech environment, which is an acoustic model corresponding to the language.

Block diagram of acoustic model generation apparatus 1 in Embodiment 1 of the present invention The flowchart explaining operation | movement of the acoustic model production | generation apparatus 1 The conceptual diagram explaining operation | movement of the acoustic model production | generation apparatus 1 The figure explaining the process of the acoustic model production | generation part 12 briefly The figure which shows the concept of the process of the acoustic model production | generation part 12 Diagram showing the amount of test data for evaluation of each language Diagram showing the total amount of data used for learning and adaptive processing in each language Figure showing the results of the experiment The figure which shows the experimental result of the same adaptation processing etc. Figure showing the results of the experiment Block diagram of acoustic model generation apparatus 2 in Embodiment 2 of the present invention The flowchart explaining operation | movement of the acoustic model production | generation apparatus 2 The conceptual diagram explaining operation | movement of the acoustic model production | generation apparatus 2 The figure explaining the specific operation | movement of the acoustic model production | generation apparatus 2 Block diagram of acoustic model generation apparatus 3 in Embodiment 3 of the present invention A flowchart for explaining the operation of the acoustic model generation device 3 Figure showing the same selection information management table Overview of a computer system that implements the acoustic model generation apparatus of the present invention Block diagram of the computer system

  Hereinafter, embodiments of an acoustic model generation device and the like will be described with reference to the drawings. In addition, since the component which attached | subjected the same code | symbol in embodiment performs the same operation | movement, description may be abbreviate | omitted again.

(Embodiment 1)
In the present embodiment, the first acoustic model of the target language using the other language old acoustic model that is the first acoustic model of the other language and the other language new acoustic model that is the second acoustic model of the other language. An acoustic model generation apparatus that generates a new target language acoustic model that is a second acoustic model of the target language from the target language old acoustic model will be described. Note that the target language is a language of an acoustic model generated by the acoustic model generation device, and is a speech language that is voice-recognized using the acoustic model. The target language has the same meaning as “Target Language” described later, and the other languages have the same meaning as “Source Language” described later. The other language is a language different from the target language of the generated acoustic model. Further, the other language is one or more languages.

  The other language old acoustic model is, for example, an acoustic model of another language before the adaptive processing is performed. Also, the other language old acoustic model may be an acoustic model generated with data different from the other language new acoustic model. In addition, the other language new acoustic model is, for example, an acoustic model of another language after the adaptive processing is performed. Further, the other language new acoustic model may be, for example, an acoustic model generated with data different from the other language old acoustic model. Further, the target language old acoustic model is, for example, an acoustic model of the target language before the adaptive process is performed. Further, the target language old acoustic model may be, for example, an acoustic model generated with data similar to the other language old acoustic model. The target language new acoustic model is, for example, an acoustic model of the target language after being subjected to adaptive processing. Here, the adaptive processing is processing for performing parameter conversion on an acoustic model that is normally used for speech recognition in one language, using stored speech in the one language. The stored voice is preferably a voice of one language stored in an environment where voice recognition is performed. Since the adaptive process is a conventional technique described in Non-Patent Document 1 or the like, detailed description thereof is omitted.

  Further, the other language old acoustic model may be, for example, an acoustic model of another language of a reading voice, and the other language new acoustic model may be, for example, an acoustic model of another language of spoken speech. Further, the target language old acoustic model may be, for example, an acoustic model of the target language of the reading speech, and the target language new acoustic model may be, for example, an acoustic model of the target language of spoken speech.

  More specifically, in the present embodiment, using the one or more first correlation information indicating the correlation between one or more other language old acoustic models and one or more other language new acoustic models, the target language old acoustics. An acoustic model generation apparatus that generates a new target language acoustic model from a model will be described.

  Note that the acoustic model is a model of acoustic features of speech for speech recognition. For example, a hidden Markov model (HMM) is used, and an output probability distribution of each state of the HMM is a Gaussian mixture distribution (GMM). It expresses with. The information (parameters) possessed by the acoustic model includes, for example, the state transition probability between states of the HMM for each symbol such as phonemes, and the average and variance of the Gaussian distribution in the GMM of each state. Usually, in speech recognition, it is common to use feature vectors of tens to hundreds of dimensions obtained by frequency analysis of speech as feature vectors for speech recognition. It is a one hundred dimensional vector.

  FIG. 1 is a block diagram of an acoustic model generation apparatus 1 in the present embodiment. The acoustic model generation device 1 includes a target language new acoustic model storage unit 11, an acoustic model generation unit 12, and an acoustic model storage unit 13.

  The acoustic model generation unit 12 includes a target language old acoustic model storage unit 121, another language old acoustic model storage unit 122, another language new acoustic model storage unit 123, a first correlation information storage unit 124, and a first correlation information generation unit. 125, and an acoustic model generation means 126.

  The target language new acoustic model storage unit 11 can store the target language new acoustic model.

  The acoustic model generation unit 12 uses one or more other language old acoustic models of one or more languages and one or more other language new acoustic models of one or more languages, and uses the target language. A target language new acoustic model is generated from the old acoustic model or one or more other language new acoustic models.

  More specifically, the acoustic model generation unit 12 is one or more first or more first information that is information regarding the relationship between one or more other language old acoustic models and one or more other language new acoustic models. Correlation information, or one or two or more correlation information of one or two or more second correlation information that is information related to the relationship between one or two or more other language old acoustic models and the target language old acoustic model. The target language new acoustic model is generated from the target language old acoustic model or one or more other language new acoustic models. The first correlation information is one or more that is a difference between one or more vectors corresponding to one or more other language old acoustic models and one or more vectors corresponding to one or more other language new acoustic models. It can be said that the information is obtained from the conversion function. Further, it can be said that the second correlation information is information acquired from one or more conversion functions of a difference between a vector corresponding to one or more other language old acoustic models and a vector corresponding to the target language old acoustic model. .

  The acoustic model generation unit 12 may generate a target language new acoustic model by mapping a vector corresponding to the target language old acoustic model using the conversion function of the first correlation information.

  Furthermore, in this Embodiment, the acoustic model production | generation part 12 demonstrates the case where a target language new acoustic model is produced | generated using 1 or 2 or more 1st correlation information from a target language old acoustic model.

  The other language new acoustic model is, for example, an acoustic model that is adaptively processed with respect to the other language old acoustic model.

  The target language old acoustic model storage unit 121 constituting the acoustic model generation unit 12 can store the target language old acoustic model.

  The other language old acoustic model storage unit 122 may store one or more other language old acoustic models of one or more other languages.

  The other language new acoustic model storage unit 123 can store one or two or more one or more other language new acoustic models.

  The first correlation information storage unit 124 can store one or more first correlation information. The first correlation information is information regarding the relationship between the other language old acoustic model stored in the other language old acoustic model storage unit 122 and the other language new acoustic model stored in the other language new acoustic model storage unit 123. is there. Here, the information on the relationship is usually information on the difference between the two acoustic models. That is, the first correlation information is usually a vector indicating a difference between a vector of the other language old acoustic model and a vector of the other language new acoustic model. A vector is a parameter set. In addition, the information regarding said relationship should just be the information which shows the relationship of two acoustic models.

  The first correlation information generation unit 125 generates one or more first correlation information using one or more other language old acoustic models and one or more other language new acoustic models. Specifically, the first correlation information generation unit 125 calculates, for example, the difference between the vector of the other language old acoustic model and the vector of the other language new acoustic model, and acquires the vector corresponding to the first correlation information To do.

  The acoustic model generation unit 126 generates a target language new acoustic model from the target language old acoustic model stored in the target language old acoustic model storage unit 121 using one or more first correlation information. For example, the acoustic model generation unit 126 adds the vector that is the first correlation information to the vector corresponding to the old acoustic model of the target language, and acquires the new target language acoustic model that is a new vector.

  The acoustic model storage unit 13 stores the target language new acoustic model generated by the acoustic model generation unit 12 in the target language new acoustic model storage unit 11.

  The target language new acoustic model storage unit 11, the target language old acoustic model storage unit 121, the other language old acoustic model storage unit 122, the other language new acoustic model storage unit 123, and the first correlation information storage unit 124 are nonvolatile recordings. A medium is preferred, but a volatile recording medium can also be realized.

  The process of storing the target language new acoustic model or the like in the target language new acoustic model storage unit 11 or the like is not limited. For example, the target language new acoustic model or the like may be stored in the target language new acoustic model storage unit 11 or the like via a recording medium, and the target language new acoustic model or the like transmitted via a communication line or the like may be stored. The target language new acoustic model storage unit 11 or the like may be stored, or the target language new acoustic model or the like input via the input device is stored in the target language new acoustic model storage unit 11 or the like. It may be like that.

  The acoustic model generation unit 12, the first correlation information generation unit 125, the acoustic model generation unit 126, and the acoustic model storage unit 13 can be usually realized by an MPU, a memory, or the like. The processing procedure of the acoustic model generation unit 12 is usually realized by software, and the software is recorded in a recording medium such as a ROM. However, it may be realized by hardware (dedicated circuit).

  Next, operation | movement of the acoustic model production | generation apparatus 1 is demonstrated using the flowchart of FIG.

  (Step S <b> 201) The first correlation information generation unit 125 acquires another language old acoustic model from the other language old acoustic model storage unit 122.

  (Step S202) The first correlation information generation means 125 acquires another language new acoustic model from the other language new acoustic model storage unit 123.

(Step S203) The first correlation information generating unit 125, a vector corresponding to other languages new acoustic model acquired in step S202 and the vector (mu _s ^I) corresponding to another language old acoustic model acquired at step S201 (mu _s First correlation information (for example, V _s = μ _s ^R −μ _s ^I ), which is information indicating a difference from ^{R 2} ), is calculated.

(Step S204) The first correlation information generation means 125 accumulates the first correlation information (V _s ) calculated in Step S203 in the first correlation information storage unit 124.

  (Step S205) The acoustic model generation means 126 determines whether to generate an acoustic model. If an acoustic model is to be generated, the process goes to step S206. If an acoustic model is not to be generated, the process returns to step S205. Note that, for example, an acoustic model may be generated by receiving a user instruction, or an acoustic model may be generated with the accumulation of the first correlation information as a trigger.

  (Step S206) The acoustic model generation means 126 acquires the target language old acoustic model from the target language old acoustic model storage unit 121.

(Step S207) The acoustic model generation means 126 acquires the first correlation information (V _s ) from the first correlation information storage unit 124.

(Step S208) The acoustic model generation unit 126 applies the first correlation information acquired in Step S207 to the target language old acoustic model acquired in Step S206, and generates a target language new acoustic model. The acoustic model generation unit 126 adds, for example, the first correlation information (V _s ) to the vector (μ _t ^I ) corresponding to the target language old acoustic model, and the target language new acoustic model (μ _t ^R = μ _t ^I + V). _s ).

(Step S209) The acoustic model storage unit 13 stores the new acoustic model (μ _t ^R ) generated in step S208 in the target language new acoustic model storage unit 11, and ends the process.

  In the flowchart of FIG. 2, the case where there is one other language has been described. However, two or more other languages may be used. In this case, the first correlation information generation unit 125 generates two or more first correlation information in two or more other languages. The acoustic model generation unit 126 generates a new target language acoustic model from the target language old acoustic model using two or more pieces of first correlation information.

  Further, step S205 in the flowchart of FIG. 2 may be omitted. That is, it goes without saying that the acoustic model generation processing may be performed immediately after the first correlation information is accumulated.

  Hereinafter, a specific operation of the acoustic model generation device 1 according to the present embodiment will be described.

  First, the concept of the operation of the acoustic model generation device 1 will be described. FIG. 3 is a conceptual diagram for explaining the operation of the acoustic model generation device 1. The acoustic model generation device 1 generates an acoustic model using the correlation between models before adaptation and after adaptation.

  The second acoustic model generation unit (not shown) of the acoustic model generation device 1 generates the other language old acoustic model 303 from the acoustic model generation speech 301 of another language (Source Language) by the speech model generation processing 302. . The other language old acoustic model 303 is “Lab AM (source)” in FIG. 3. The voice model generation process 302 is a known technique, and thus detailed description thereof is omitted. Next, the adaptation processing means (not shown) of the acoustic model generation apparatus 1 performs an adaptation process (acoustic model generation (adaptation)) 305 using the accumulated speech 304 to generate another language new acoustic model from the other language old acoustic model 303. 306 is generated. The other language new acoustic model 306 is “Fld AM (source)” in FIG. 3. Then, the first correlation information generation unit 125 calculates first correlation information f (307) that is a difference between the other language old acoustic model 303 and the other language new acoustic model 306.

  Next, a second acoustic model generation unit (not shown) of the acoustic model generation device 1 performs a target language old acoustic model 310 from a target language (Target Language) acoustic model generation speech 308 by a speech model generation process 309. Is generated. The target language old acoustic model 310 is “Lab AM (target)” in FIG. 3. Here, when the target language accumulated speech 311 does not exist, the acoustic model generation (adaptation) 312 cannot be performed on the target language old acoustic model 310. That is, the broken line in FIG. 3 indicates data that does not exist or processing that cannot be performed. Then, the acoustic model generation unit 126 applies the first correlation information f (313) to the target language old acoustic model 310 to generate the target language new acoustic model 314. This generated target language new acoustic model is a model to be created. Also, the target language new acoustic model 314 of FIG. 3 is “Fld AM (target)”.

  Hereinafter, two specific examples and experimental results will be described for the processing of the acoustic model generation unit 12.

(Specific example 1)
First, the process of the acoustic model generation unit 12 will be briefly described with reference to FIG. In the first specific example, a case where one state of speech is modeled by a two-dimensional normal distribution and only the average of the normal distribution is applied will be described.

Now, the baseline model _{S I} in other languages (Source Language), and the average ^{_{"μ s I = (1,1 / 2}} ) " 2-dimensional normal distribution of variance sigma _s ^I. In addition, the baseline model S _I is another language old acoustic model. Further, the adaptation model S _R of another language adapted by the stored speech is assumed to be a two-dimensional normal distribution having an average “μ _s ^R = (0, 1)” and a variance σ _s ^R (= σ _s ^I ). It should be noted that adaptive model S _R is another language new acoustic model.

At this time, the first correlation information generating unit 125 calculates a difference V _s (41 in FIG. 4) between the average vectors of the adaptive model S _R and the baseline model S _{I according} to the following equation: “V _s = μ _s ^R −μ _s ^I = (1,1 / 2) ”.

In addition, the baseline model T _I of the target language (Target Language) is a two-dimensional normal distribution having an average “μ _t ^I = (0, 0)” and a variance σ _t ^I. In addition, the baseline model T _I is a target language old acoustic model.

Then, when adapting by using the difference vector V _s of the average vectors of other languages as it is, the acoustic model generation unit 126 calculates the average “μ _t ^R = μ _t ^I + V _s = (1,1 / 2)”, the variance σ _t It gets the two-dimensional normal distribution with ^R (= ^σ _t ^I), which is referred to as adaptive model _{T R} of the target language. It should be noted that adaptive model T _R is a target language new acoustic model.

  In the specific example 1, it is assumed that one state of speech is modeled by a two-dimensional normal distribution. However, the state is not limited to a two-dimensional normal distribution, and is modeled by a mixed normal distribution of several tens of dimensions. It is further preferable.

  In addition, the model is not necessarily modeled with a mixed normal distribution. For example, in the case of an acoustic model using a neural network, adaptation can be performed using the first correlation information that is a difference between two acoustic models.

(Specific example 2)
In the second specific example, the adaptation processing unit (not shown) of the acoustic model generation apparatus 1 is based on the MAP adaptation method, and the other language old acoustic model (also referred to as “an initial acoustic model of another language”) in another language (Source Language). Then, another language new acoustic model (also referred to as “an acoustic model adapted to another language”) is generated using the accumulated speech of the other language. Here, the average vector of the sth Gaussian distribution of the other language old acoustic model is μ _s ^I , and the average vector of the sth Gaussian distribution of the other language new acoustic model is μ _s ^R.

In the MAP adaptation method, when the average vector of the other language new acoustic model is adapted, the average vector (μ _s ^R ) is defined as each average vector (μ _s ^I ) of the other language old acoustic model as a prior distribution average vector, It is calculated by the following formula 1.

In Equation 1, m _s is the maximum likelihood estimate of the sth Gaussian distribution obtained from the stored speech. n is the total number of learning samples obtained from the accumulated speech for the corresponding Gaussian distribution. Also, τ is a parameter for adjusting the relative balance between the prior distribution and the sample obtained from the accumulated speech.

  And in the specific example 2, a difference vector is calculated | required as follows. The concept of the process in the specific example 2 is shown in FIG.

The first correlation information generation means 125 shows the first correlation information (V _s ), which is the difference between the other language old acoustic model (μ _s ^I ) and the other language new acoustic model (μ _s ^R ), as shown in Equation 2. Calculate as follows. This first correlation information (V _s ) is a transition vector of an average vector in another language.

Here, sεK ₁ (K ₁ is a Gaussian distribution set of training data in other languages)

The first correlation information generation unit 125 calculates the first correlation information that is a transition vector by substituting the average vector (μ _s ^R ) of Equation 1 into Equation 2 (see Equation 3).

In Equation 3, the transition vector (V _s ) obtained by the MAP application method is expressed as “V _s ^ML = (m _s −μ _s ^I )”, and is calculated by maximum likelihood (ML) estimation.

Further, in the following Expression 4, it is indicated that the transition vector by the MAP application method is obtained by the transition vector (V _s ^ML ) corrected by using the ML estimation by the weighting factor. The weighting factor depends on the total number n of learning samples.

  Next, as in other languages, an average vector of the Gaussian distribution of the target language old acoustic model (also referred to as “an initial model of the target language”) is generated by acoustic model learning.

Here, the t-th average vector of the Gaussian distribution of the target language old acoustic model is denoted by μ _t ^I. Here, since there is no accumulated voice data for the target language adaptation processing, the transition vector of each Gaussian distribution of the target language is estimated by the transition vector of the other language.

_{T in} the transition vector (μ _t ^I ) of the target language is “tεK ₂ ”. Here, K ₂ is a Gaussian distribution set of the target language. The transition vector (V _t ) in μ _t ^I is interpolated by the following equation 5 of the learned transition vector V _s .

In Equation 5, N (t) is a set of Gaussian distributions near K of the vector (μ _t ^I ). λ _{t, sk} is a weighting factor and depends on the distance between μ _t ^I and μ _sk ^I. Summed transition vector _{V t} is the vector (mu _t ^I), vector (mu _t ^R) is obtained (see Equation 6). In Equation 5, s _k is the k-th s [sεK ₁ (K ₁ is a Gaussian distribution set of training data in other languages)].

  In addition, for example, a set of Gaussian distributions in which K is close is the conventional technique of Kullback-Leibler divergence (KL-divergence) (“S. Kullback, and RA Leibler,“ On information and sufficiency, ”Annals of Mathematical Statistics, vol. 22, no. 1, pp. 79-86, 1951 ”).

Further, the weight coefficient (λ _{a, b} ) is calculated by, for example, the following formula 7.

In Equation 7, d _{a and b} are distances between the vector (μ _a ^I ) and the vector (μ _b ^R ) calculated based on KL-divergence, and f is a parameter for weight adjustment.
(Experiment)

  Hereinafter, experimental results will be described. In this experiment, the acoustic model generation apparatus 1 that operates according to the method of the specific example 2 is used. In this experiment, the other language is Japanese and the target language is Indonesian. In other words, an acoustic model before the Japanese adaptation process, an acoustic model after the Japanese adaptation process, and an acoustic model before the Indonesian adaptation process exist in advance.

  The amount of test data for evaluation in each language is shown in FIG. In this experiment, two utterance data were used. One is a travel conversation basic expression corpus (BTEC), and the other is voice data (VTlog) recorded in an actual environment. BTEC is voice data of travel conversation basic expressions acquired in a clean environment. The sound data acquired in a clean environment is, for example, sound data recorded in a recording room, sound data acquired when a document is read out, and the like. VTlog is voice data recorded by VoiceTra (refer to URL “http://mastar.jp/translation/index.html”). It contains voice data including noise and voice data of various utterance styles. Including. In FIG. 6, “time” indicates a recording time (unit: time), and “utterance” indicates the number of utterances.

  FIG. 7 is a table showing the total amount of data used for learning and adaptive processing in each language. The learning data is voice data uttered in the laboratory (data in the column “Learning” in FIG. 7) and voice data uttered in the real environment, and is recorded by Voice Tra (“adaptive processing ( VTlog) "column data). The acoustic model is a three-state Left-to-Right, gender-independent HMM learned from learning data in each language. The number of states is 5000 in Indonesian and 500 in Japanese, and a decision tree-based clustering method is used as a state sharing method. For Indonesian, 4 Gaussian distributions were used for each state, and for Japanese, 16 Gaussian distributions were used for each state.

  Moreover, the language model (LMs) of each language was learned using the BTEC corpus.

  FIG. 8 shows the experimental results. BTEC had a word error rate (WER) of 17.74% in Japanese and 15.97% in Indonesian.

  On the other hand, because the learning model and test speech differ greatly, the VTlog WER was 37.75% in Japanese and 55.31% in Indonesian. From this experimental result, it can be seen that a mismatch in the learning model and the test speech causes a decrease in accuracy.

  Next, in order to reduce the mismatch, MAP adaptation was performed on the Japanese acoustic model using the accumulated speech in the Japanese actual speech environment, which is speech data recorded by VoiceTra. FIG. 9 shows the results of the adaptation experiment. The VTlog WER was 24.66%, a significant improvement over the baseline (37.75%). This shows the effect of adapting the acoustic model using speech data that matches the actual utterance.

  Next, the method of the specific example 2 was evaluated. In the evaluation, the parameters were experimentally determined as “τ = 10”, “f = 3”, and “k <= 10”. FIG. 10 shows the experimental results. The VTlog WER improved from 55.31% to 50.40% and achieved an error reduction rate (ERR) of 8.9% (see “Proposed” row in FIG. 10). This result verified our hypothesis below. Our hypothesis is to improve the recognition accuracy by applying the transition vector estimated by the transition vector of another language (here, Japanese) to the acoustic model of the target language (here, Indonesian). Thereby, the method of the acoustic model generation device 1 adapts the acoustic model of the target language to the actual speech environment without using the accumulated speech of the target language (here Indonesian language) related to the actual speech.

In specific example 2, one state of speech is a mixed normal distribution, and the average is adapted based on the MAP adaptation method to generate a new acoustic model in another language. For example, other parameters such as dispersion of normal distribution and state transition probability of HMM can be similarly applied. Also, the acoustic model is not necessarily modeled with a mixed normal distribution. For example, even in the case of an acoustic model using a neural network, it is possible to adapt using the first correlation information that is the difference between the two acoustic models. it can.

  As described above, according to the present embodiment, even in a language that does not have voice data in a suitable environment such as a speech environment, an acoustic model corresponding to the language can be generated, and an acoustic model that improves voice recognition accuracy can be generated.

  In addition, according to the present embodiment, by using the correlation between the acoustic model before the adaptation process of the other language and the acoustic model after the adaptation process of the other language, the voice data in a suitable environment such as the speech environment is obtained. Even a language that does not exist is an acoustic model corresponding to the language, and an acoustic model that improves speech recognition accuracy can be generated.

  Note that the processing in the present embodiment may be realized by software. Then, this software may be distributed by software download or the like. Further, this software may be recorded and distributed on a recording medium such as a CD-ROM. This also applies to other embodiments in this specification. Note that the software that implements the acoustic model generation apparatus according to the present embodiment is the following program. In other words, this program causes the computer to execute one or more other language old acoustic models which are first acoustic models of one or more other languages different from the target language and second acoustics of the one or more other languages. First correlation information that is information related to a relationship with one or more other language new acoustic models that are models, or a target language that is a first acoustic model of each of the one or more other language old acoustic models and the target language The target language new acoustic model is obtained from the target language old acoustic model or one or more other language new acoustic models using any one or more correlation information of the second correlation information that is information related to the acoustic model. A program for causing an acoustic model generation unit to generate and an acoustic model storage unit that stores a target language new acoustic model generated by the acoustic model generation unit in a recording medium.

  In the program, the acoustic model generation unit includes a target language old acoustic model storage unit that can store a target language old acoustic model, a first correlation information storage unit that can store first correlation information, and the first Using the correlation information, the acoustic model generation means for generating the target language new acoustic model from the target language old acoustic model stored in the target language old acoustic model storage unit, and causing the computer to function Is preferred.

(Embodiment 2)
In this embodiment, an acoustic model generation apparatus that generates a target language new acoustic model from another language new acoustic model using the other language old acoustic model and the target language old acoustic model will be described.

  More specifically, in the present embodiment, the target language new acoustic model is obtained from the other language new acoustic model using the second correlation information indicating the correlation between the other language old acoustic model and the target language old acoustic model. An acoustic model generation device to be generated will be described.

  FIG. 11 is a block diagram of the acoustic model generation apparatus 2 in the present embodiment. The acoustic model generation device 2 includes a target language new acoustic model storage unit 11, an acoustic model generation unit 22, and an acoustic model storage unit 13.

  The acoustic model generation unit 22 includes a target language old acoustic model storage unit 121, another language old acoustic model storage unit 122, another language new acoustic model storage unit 123, a second correlation information storage unit 224, a second correlation information generation unit 225, And an acoustic model generation means 226.

  The acoustic model generation unit 22 is one or two or more pieces of first correlation information, or one or two or more pieces of information related to the relationship between one or more other language old acoustic models and one or more other language new acoustic models. The target language old acoustic model or the other using one or more correlation information of one or two or more second correlation information, which is information related to the relationship between the other language old acoustic model and the target language old acoustic model A target language new acoustic model is generated from the language new acoustic model. The acoustic model generation unit 22 generates another language new acoustic model by mapping from one or more other language new acoustic models stored in the other language new acoustic model storage unit 122 using the conversion function of the second correlation information. You may do it.

  Furthermore, in the present embodiment, the acoustic model generation unit 22 generates a target language new acoustic model from one or more other language new acoustic models using one or more second correlation information. explain.

  The second correlation information storage unit 224 can store one or more second correlation information. The second correlation information is information related to the relationship between the other language old acoustic model and the target language old acoustic model. Here, the information on the relationship is usually information on the difference between the two acoustic models. That is, the second correlation information is usually a vector indicating a difference between a vector corresponding to the other language old acoustic model and a vector corresponding to the target language old acoustic model. A vector is a parameter set. The structure of the second correlation information may be the same as the structure of the first correlation information.

  The second correlation information generation unit 225 generates one or more second correlation information using one or more other language old acoustic models and the target language old acoustic model. Specifically, the second correlation information generation unit 225 calculates, for example, a difference between a vector corresponding to one or two or more other language old acoustic models and a vector corresponding to the target language old acoustic model. Vectors corresponding to two or more pieces of second correlation information are acquired.

  The acoustic model generation unit 226 generates a target language new acoustic model from the other language new acoustic model stored in the other language new acoustic model storage unit 123 using one or more second correlation information. For example, the acoustic model generation unit 126 adds the vector that is the second correlation information to the vector corresponding to the other language new acoustic model, and acquires the target language new acoustic model that is a new vector.

  The acoustic model generation unit 22, the second correlation information generation unit 225, and the acoustic model generation unit 226 can be usually realized by an MPU, a memory, or the like. The processing procedure of the acoustic model generation unit 22 and the like is usually realized by software, and the software is recorded on a recording medium such as a ROM. However, it may be realized by hardware (dedicated circuit).

  The second correlation information storage unit 224 is preferably a non-volatile recording medium, but can also be realized by a volatile recording medium. The process in which the second correlation information is stored in the second correlation information storage unit 224 does not matter. For example, the second correlation information may be stored in the second correlation information storage unit 224 via the recording medium, and the second correlation information transmitted via the communication line or the like is stored in the second correlation information storage unit. 224 may be stored, or the second correlation information input via the input device may be stored in the second correlation information storage unit 224.

  Next, operation | movement of the acoustic model production | generation apparatus 2 is demonstrated using the flowchart of FIG.

  (Step S <b> 1201) The second correlation information generation unit 225 acquires another language old acoustic model from the other language old acoustic model storage unit 122.

  (Step S1202) The second correlation information generation unit 225 acquires the target language old acoustic model from the target language old acoustic model storage unit 121.

(Step S1203) The second correlation information generating unit 225, vector corresponding to the target language old acoustic model acquired acquired vectors corresponding to other languages old acoustic model (mu _s ^I) in step S1202 in step S1201 (mu _t the second correlation information indicating a difference between ^I) (e.g., to calculate the ^{_{V I = μ t I -μ s}} I).

(Step S1204) The second correlation information generation unit 225 accumulates the second correlation information (V _I ) calculated in Step S1203 in the second correlation information storage unit 224.

  (Step S1205) The acoustic model generation means 226 determines whether to generate an acoustic model. If an acoustic model is to be generated, the process goes to step S1206. If an acoustic model is not to be generated, the process returns to step S1205. Note that, for example, an acoustic model may be generated by accepting a user instruction, or an acoustic model may be generated with the accumulation of second correlation information or the like as a trigger.

(Step S1206) The acoustic model generation means 226 acquires another language new acoustic model (μ _s ^I ) from the other language new acoustic model storage unit 123.

(Step S1207) The acoustic model generation unit 226 acquires the second correlation information (V _I ) from the second correlation information storage unit 224.

(Step S1208) The acoustic model generation unit 226 applies the first correlation information acquired in Step S1207 to the other language new acoustic model acquired in Step S1206, and generates a target language new acoustic model. For example, the acoustic model generation unit 126 adds the second correlation information (V _I ) to the vector (μ _s ^I ) corresponding to the target language old acoustic model, and the target language new acoustic model (μ _t ^R = μ _s ^I). + V _I ).

(Step S1209) The acoustic model storage unit 13 stores the new acoustic model (μ _t ^R ) generated in step S1208 in the target language new acoustic model storage unit 11, and ends the process.

  In the flowchart of FIG. 12, the case where there is one other language has been described, but the number of other languages may be two or more. In this case, the second correlation information generation unit 225 generates two or more second correlation information in two or more other languages. The acoustic model generation unit 226 generates a target language new acoustic model from the target language old acoustic model using two or more pieces of second correlation information. For example, the acoustic model generation unit 226 acquires an average vector of two or more pieces of second correlation information, adds the average vector to a vector corresponding to the target language old acoustic model, and calculates a target language new acoustic model.

  Further, step S1205 in the flowchart of FIG. 12 may be omitted. That is, it goes without saying that an acoustic model generation process may be performed immediately after the second correlation information is accumulated.

  Hereinafter, a specific operation of the acoustic model generation device 2 in the present embodiment will be described. First, the concept of the operation of the acoustic model generation device 1 will be described. FIG. 13 is a conceptual diagram illustrating the operation of the acoustic model generation device 2.

  Here, the acoustic model generation apparatus 2 generates an acoustic model by using the correlation between the languages of the pre-adaptation model.

  The second acoustic model generation means (not shown) of the acoustic model generation device 2 generates the other language old acoustic model 1303 from the acoustic language generation source 1301 of another language (Source Language) by the speech model generation processing 1302. . The other language old acoustic model 1303 is “Lab AM (source)” in FIG. 13. Next, the adaptive processing means (not shown) of the acoustic model generation apparatus 2 performs the adaptive processing (acoustic model generation (adaptive)) 1305 using the accumulated speech 1304 to convert the other language old acoustic model 1303 to the other language new acoustic model. 1306 is generated. The other language new acoustic model 1306 is “Fld AM (source)” in FIG. 13.

  Next, the second acoustic model generation unit (not shown) of the acoustic model generation device 2 performs the target language old acoustic model 1309 from the target language (Target Language) acoustic model generation speech 1307 by the speech model generation processing 1308. Is generated. The target language old acoustic model 1309 is “Lab AM (target)” in FIG. Here, since the target language accumulated speech 1310 does not exist, the acoustic model generation (adaptation) 1311 cannot be performed on the target language old acoustic model 1309. That is, the broken line in FIG. 13 indicates data that does not exist or processing that cannot be performed.

  Then, the second correlation information generation means 225 generates the second correlation information g (1312) using the other language old acoustic model 1303 and the target language old acoustic model 1309. Specifically, the second correlation information generation unit 225 calculates, for example, a difference between a vector corresponding to the other language old acoustic model 1303 and a vector corresponding to the target language old acoustic model 1309. A vector which is the difference between the vectors is the second correlation information g.

  Next, the acoustic model generation unit 226 applies the second correlation information g to the target language old acoustic model 1309 to generate a target language new acoustic model 1313. Specifically, the acoustic model generation unit 226 adds a vector corresponding to the second correlation information g to a vector corresponding to the target language old acoustic model 1309 to generate a target language new acoustic model 1313. The generated target language new acoustic model is a model that the user wants to create. Further, the target language new acoustic model 1313 of FIG. 13 is “Fld AM (target)”.

  Hereinafter, further specific examples of the processing of the acoustic model generation unit 22 will be described.

(Concrete example)
Here, a specific operation of the acoustic model generation device 2 will be described with reference to FIG. 14, the baseline model _{S I} in other languages (Source Language), mean ^{_{"μ s I = (0,1 / 2}} ) ", a two-dimensional normal distribution of variance sigma _s ^I. In addition, the baseline model S _I is another language old acoustic model. In addition, the adaptation model S _R of another language adapted with the stored speech is a two-dimensional normal distribution having an average “μ _s ^R = (1, 1)” and a variance σ _s ^R (= σ _s ^I ). It should be noted that adaptive model S _R is another language new acoustic model. Further, the target language old acoustic model T _I is a two-dimensional normal distribution having an average “μ _t ^R = (0, 0)” and a variance σ _t ^R (= σ _t ^I ). In addition, the target language new acoustic model and T _R. The other language old acoustic model (S _I ), the other language new acoustic model (S _R ), and the target language old acoustic model (T _I ) were acquired by the processing described in the first specific example of the first embodiment. , And.

In such a situation, the second correlation information generation means 225 sets the difference (V _I ) between the average vectors of S _I and T _I as “V _I = μ _t ^I −μ _s ^I = (0, −1/2)”. get. This second correlation information is an arrow 141 in FIG.

Next, the acoustic model generation means 226 applies the difference (V _I ) of this average vector to the other language new acoustic model (142 in FIG. 14), and the average “μ _t ^R = μ _s ^I + V _I = (1) , ½) ”, a two-dimensional normal distribution with variance σ _t ^R (= σ _t ^I ). The 2-dimensional normal distribution, an adaptive model T _R of the target language. T _R is the target language new acoustic model.

In this specific example, as in specific example 1, it is assumed that one state of speech is modeled by a two-dimensional normal distribution. However, the present invention is not limited to a two-dimensional normal distribution, and is mixed by several tens of dimensions. It is more preferable to model with a normal distribution or the like. In addition, the model is not necessarily modeled with a mixed normal distribution. For example, in the case of an acoustic model using a neural network, adaptation can be performed using the second correlation information that is a difference between two acoustic models.

  As described above, according to the present embodiment, by using the correlation between the acoustic model before the adaptation process of the other language and the acoustic model before the adaptation process of the target language, the voice data in a suitable environment such as the speech environment is obtained. Even a language that does not exist is an acoustic model corresponding to the language, and an acoustic model that improves speech recognition accuracy can be generated.

  In addition, the software which implement | achieves the acoustic model production | generation apparatus 2 in this Embodiment is the following programs. In other words, this program causes the computer to execute one or more other language old acoustic models which are first acoustic models of one or more other languages different from the target language and second acoustics of the one or more other languages. First correlation information that is information related to a relationship with one or more other language new acoustic models that are models, or a target language that is a first acoustic model of each of the one or more other language old acoustic models and the target language The target language new acoustic model is obtained from the target language old acoustic model or one or more other language new acoustic models using any one or more correlation information of the second correlation information that is information related to the acoustic model. A program for causing an acoustic model generation unit to generate and an acoustic model storage unit that stores a target language new acoustic model generated by the acoustic model generation unit in a recording medium.

  Further, in the above program, the acoustic model generation unit includes an other language new acoustic model storage unit that can store the other language new acoustic model, a second correlation information storage unit that can store second correlation information, and the first Using the two-correlation information, the computer functions as an acoustic model generation unit that generates a target language new acoustic model from the other language new acoustic model stored in the other language new acoustic model storage unit. Is preferred.

(Embodiment 3)
In the present embodiment, in the present embodiment, the target language old acoustic model or the other language new acoustic is obtained using one or more correlation information of one or more first correlation information and one or more second correlation information. An acoustic model generation apparatus that generates a target language new acoustic model from the model or the target language old acoustic model and another language new acoustic model will be described.

  More specifically, an acoustic model generation apparatus in which the usage (algorithm) of the first correlation information and the second correlation information dynamically changes in the present embodiment will be described.

  FIG. 15 is a block diagram of the acoustic model generation device 3 in the present embodiment.

  The acoustic model generation device 3 includes a target language new acoustic model storage unit 11, an acoustic model generation unit 32, and an acoustic model storage unit 13.

  The acoustic model generation unit 32 includes a target language old acoustic model storage unit 121, another language old acoustic model storage unit 122, another language new acoustic model storage unit 123, a first correlation information storage unit 124, a first correlation information generation unit 125, A second correlation information storage unit 224, a second correlation information generation unit 225, a selection unit 321 and an acoustic model generation unit 326 are provided. The selection unit 321 includes a selection information management unit 3211.

  The acoustic model generation unit 32 is one or more first correlation information that is information related to the relationship between the other language old acoustic model and the other language new acoustic model, or the relationship between the other language old acoustic model and the target language old acoustic model. A target language new acoustic model is generated from the target language old acoustic model or the other language new acoustic model by using any one or more correlation information of one or two or more second correlation information that is information regarding.

  Furthermore, in this Embodiment, the acoustic model production | generation part 32 uses a 1 or 2 or more 1st correlation information and a 1 or 2 or more 2nd correlation information, and becomes a target language old acoustic model or another language new acoustic. A target language new acoustic model is generated from the model or the target language old acoustic model and another language new acoustic model.

  The selection means 321 selects one of the two or more algorithms for generating the target language new acoustic model according to the data of the target language old acoustic model or the other language new acoustic model. For example, a 1st algorithm is an algorithm which produces | generates a target language new acoustic model from a target language old acoustic model using 1st correlation information. For example, the second algorithm is an algorithm for generating a target language new acoustic model from another language new acoustic model using the second correlation information. Further, for example, the third algorithm is an algorithm for generating a target language new acoustic model from the target language old acoustic model and the other language new acoustic model using the first correlation information and the second correlation information.

  The selection information management unit 3211 can store one or more selection information that is information for the selection means 321 to determine an algorithm. The selection information is, for example, information on a pair of a phoneme identifier that identifies a phoneme and an algorithm identifier that identifies an algorithm. Note that the selection information may include selection information for switching the algorithm in units smaller than phonemes. Further, the selection information may include selection information for switching the algorithm in units rougher than phonemes.

  The acoustic model generation means 326 uses the one or more pieces of first correlation information and one or more pieces of correlation information among the one or more pieces of second correlation information, and uses the target language old acoustic model or other language new acoustic model. Alternatively, the target language new acoustic model is generated from the target language old acoustic model and the other language new acoustic model.

  More specifically, the acoustic model generation unit 326 performs one or more correlations among one or more first correlation information and one or two or more second correlation information according to one algorithm selected by the selection unit 321. The target language new acoustic model is generated using the information.

  The acoustic model generation unit 32, the selection unit 321 and the acoustic model generation unit 326 can usually be realized by an MPU, a memory, or the like. The processing procedure of the acoustic model generation unit 32 and the like is usually realized by software, and the software is recorded on a recording medium such as a ROM. However, it may be realized by hardware (dedicated circuit).

  The selection information management unit 3211 is preferably a nonvolatile recording medium, but can also be realized by a volatile recording medium. The process in which the selection information is stored in the selection information management unit 3211 is not limited. For example, the selection information may be stored in the selection information management unit 3211 via a recording medium, and the selection information transmitted via a communication line or the like is stored in the selection information management unit 3211. Alternatively, the selection information input via the input device may be stored in the selection information management unit 3211.

  Next, operation | movement of the acoustic model production | generation apparatus 3 is demonstrated using the flowchart of FIG. In the flowchart of FIG. 16, the process of accumulating the first correlation information in the first correlation information storage unit 124 and the process of accumulating the second correlation information in the second correlation information storage unit 224 are described in the first and second embodiments. Therefore, the description here is omitted. In the flowchart of FIG. 16, processing for generating a target language new acoustic model from the target language old acoustic model and / or another language new acoustic model will be described. In the flowchart of FIG. 16, the description of the same steps as those in the flowchart of FIG. 2 is omitted.

  (Step S1601) The selection unit 321 substitutes 1 for the counter i.

  (Step S1602) The selection unit 321 determines whether or not the i-th processing unit (for example, the i-th phoneme) exists in the target language old acoustic model or the other language new acoustic model. If the i-th processing unit exists, the process goes to step S1603, and if it does not exist, the process ends.

  (Step S1603) The selection unit 321 acquires the processing unit identifier (for example, “a” of the phoneme identifier) of the i-th processing unit in the target language old acoustic model or the other language new acoustic model.

  (Step S1604) The selection unit 321 acquires an algorithm identifier corresponding to the processing unit identifier acquired in step S1603 from the selection information management unit 3211.

  (Step S1605) The acoustic model generation unit 326 determines whether the algorithm identifier acquired in Step S1604 is information indicating that the algorithm is the first algorithm. If it is the first algorithm, go to step S206, and if it is not the first algorithm, go to step S1606. Here, the first algorithm is an algorithm for generating the target language new acoustic model from the target language old acoustic model using the first correlation information, as described in the first embodiment. is there.

  (Step S1606) The acoustic model generation means 326 determines whether or not the algorithm identifier acquired in Step S1604 is information indicating that the algorithm is the second algorithm. If it is the second algorithm, go to step S1206, and if it is not the second algorithm, go to step S206. If it is not the second algorithm, it is the third algorithm. The second algorithm is an algorithm in which the acoustic model generation unit 22 described in Embodiment 2 generates a target language new acoustic model from another language new acoustic model using the second correlation information. The third algorithm is an algorithm for generating a target language new acoustic model from the target language old acoustic model and the other language new acoustic model using the first correlation information and the second correlation information.

  (Step S1607) The selection unit 321 increments the counter i by 1, and returns to step S1602.

  (Step S1608) The acoustic model generation means 326 generates a target language new acoustic model from the target language old acoustic model and the other language new acoustic model using the first correlation information and the second correlation information. Go to step S209.

  Hereinafter, a specific operation of the acoustic model generation device 3 in the present embodiment will be described. Here, the selection information management unit 3211 stores the selection information management table shown in FIG. The selection information management table stores two or more records having “phoneme identifier” and “algorithm identifier”. In addition, the first algorithm is executed for the phoneme corresponding to the algorithm identifier “1”, and the second algorithm is executed for the phoneme corresponding to the algorithm identifier “2”. This means that the third algorithm is executed for the phoneme corresponding to the algorithm identifier “3”.

  Hereinafter, the operation of the acoustic model generation unit 32 will be described. First, it is assumed that the selection unit 321 has acquired the phoneme identifier “a” of the first phoneme in the target language old acoustic model. Next, the selection unit 321 acquires the algorithm identifier “3” paired with the phoneme identifier “a” from the selection information management table.

  Then, the acoustic model generation unit 326 executes the third algorithm as follows according to the algorithm identifier “3”. Here, it is assumed that the acoustic model is a two-dimensional normal distribution, as in the first specific example of the first embodiment.

That is, the acoustic model generation unit 326 acquires the first correlation information “V _s = μ _s ^R −μ _s ^I = (1, 1/2)” in the first correlation information storage unit 124.

In addition, the acoustic model generation unit 326 converts the two-dimensional normal distribution that is the baseline model T _I (average μ _t ^I = (0, 0), variance σ _t ^I ) of the target language into the target language old sound. Obtained from the model storage unit 121.

Next, the acoustic model generation unit 326 acquires a two-dimensional normal distribution having an average “μ _t ^R1 = μ _t ^I + V _s = (1, 1/2)” and a variance σ _t ^R (= σ _t ^I ). .

Next, the acoustic model generation unit 326 acquires the second correlation information “V _I = μ _t ^I −μ _s ^I = (0, −1/2)” in the second correlation information storage unit 224.

Further, the acoustic model generation unit 326 applies the difference (V _I ) of the average vectors to the other language new acoustic model, and calculates the average “μ _t ^R2 = μ _s ^I + V _I = (1, 1/2)”, A two-dimensional normal distribution having a variance σ _t ^R (= σ _t ^I ) is acquired.

Next, the acoustic model generation unit 326 executes “½ (μ _t ^R1 + μ _t ^R2 )” to obtain a final other language new acoustic model μ _t ^R. Here, the acoustic model generation unit 326 generates 50% and 50% of the application of μ _t ^R1 and μ _t ^R2 , but may generate an acoustic model with different weights.

In this specific example, as in specific example 1, it is assumed that one state of speech is modeled by a two-dimensional normal distribution. However, the present invention is not limited to a two-dimensional normal distribution, and is mixed by several tens of dimensions. It is more preferable to model with a normal distribution or the like. In addition, the model is not necessarily modeled with a mixed normal distribution. For example, in the case of an acoustic model using a neural network, adaptation can be performed using the second correlation information that is a difference between two acoustic models.

  As described above, according to the present embodiment, the correlation between the acoustic model before the adaptation process of the other language and the acoustic model after the adaptation process of the other language, and the adaptation process of the acoustic model and the target language before the adaptation process of the other language. By using the correlation with the previous acoustic model, even in languages where there is no speech data in a suitable environment such as a speech environment, an acoustic model corresponding to the language can be generated and an acoustic model that improves speech recognition accuracy can be generated .

  Note that the software that implements the acoustic model generation apparatus according to the present embodiment is the following program. In other words, this program causes the computer to execute one or more other language old acoustic models which are first acoustic models of one or more other languages different from the target language and second acoustics of the one or more other languages. First correlation information that is information related to a relationship with one or more other language new acoustic models that are models, or a target language that is a first acoustic model of each of the one or more other language old acoustic models and the target language The target language new acoustic model is obtained from the target language old acoustic model or one or more other language new acoustic models using any one or more correlation information of the second correlation information that is information related to the acoustic model. A program for causing an acoustic model generation unit to generate and an acoustic model storage unit that stores a target language new acoustic model generated by the acoustic model generation unit in a recording medium.

  In the above program, the acoustic model generation unit includes a target language old acoustic model storage unit that can store the target language old acoustic model, another language new acoustic model storage unit that can store another language new acoustic model, Using the first correlation information storage unit capable of storing one correlation information, the second correlation information storage unit capable of storing second correlation information, the first correlation information and the second correlation information; An acoustic model generating means for generating a target language new acoustic model from the old acoustic model, the other language new acoustic model, or the target language old acoustic model and the other language new acoustic model, and a computer, It is preferable to make it function.

  In the above program, the acoustic model generation unit may be any one of two or more algorithms for generating a target language new acoustic model according to data of the target language old acoustic model or the other language new acoustic model. The acoustic model generation means further comprises one or more correlations of the first correlation information and the second correlation information according to the one algorithm selected by the selection means. It is preferable to cause a computer to function as information that generates the new target language acoustic model.

  FIG. 18 shows the external appearance of a computer that executes the program described in this specification to realize the acoustic model generation apparatus according to the various embodiments described above. The above-described embodiments can be realized by computer hardware and a computer program executed thereon. FIG. 18 is an overview diagram of the computer system 300, and FIG. 19 is a block diagram of the system 400.

  In FIG. 18, a computer system 400 includes a computer 401 including a CD-ROM drive, a keyboard 402, a mouse 403, and a monitor 404.

  In FIG. 19, a computer 401 includes an MPU 4013, a bus 4014, a ROM 4015, a RAM 4016, and a hard disk 4017 in addition to a CD-ROM drive 4012. The bus 4014 is connected to the MPU 4013 and the CD-ROM drive 4012. The ROM 4015 stores a program such as a bootup program. A RAM 4016 is connected to the MPU 4013, and temporarily stores application program instructions and provides a temporary storage space. The hard disk 4017 is for storing application programs, system programs, and data. Although not shown here, the computer 401 may further include a network card that provides connection to a LAN.

  A program that causes the computer system 400 to execute the functions of the acoustic model generation apparatus according to the above-described embodiment may be stored in the CD-ROM 4101, inserted into the CD-ROM drive 4012, and further transferred to the hard disk 4017. Alternatively, the program may be transmitted to the computer 401 via a network (not shown) and stored in the hard disk 4017. The program is loaded into the RAM 4016 at the time of execution. The program may be loaded directly from the CD-ROM 4101 or the network.

  The program does not necessarily include an operating system or a third-party program that causes the computer 401 to execute the functions of the acoustic model generation apparatus according to the above-described embodiment. The program only needs to include an instruction portion that calls an appropriate function (module) in a controlled manner and obtains a desired result. How the computer system 400 operates is well known and will not be described in detail.

  Further, the computer that executes the program may be singular or plural. That is, centralized processing may be performed, or distributed processing may be performed.

  In each of the above embodiments, each process (each function) may be realized by centralized processing by a single device (system), or by distributed processing by a plurality of devices. May be.

  The present invention is not limited to the above-described embodiments, and various modifications are possible, and it goes without saying that these are also included in the scope of the present invention.

  As described above, the acoustic model generation device according to the present invention generates an acoustic model that improves speech recognition accuracy, even in a language that does not have speech data in a suitable environment such as an utterance environment, which is an acoustic model corresponding to the language. It is useful as an acoustic model generation device or the like.

1, 2, 3 Acoustic model generation device 11 Target language new acoustic model storage unit 12, 22, 32 Acoustic model generation unit 13 Acoustic model storage unit 121 Target language old acoustic model storage unit 122 Other language old acoustic model storage unit 123 Other language New acoustic model storage unit 124 First correlation information storage unit 125 First correlation information generation unit 126, 226, 326 Acoustic model generation unit 224 Second correlation information storage unit 225 Second correlation information generation unit 321 Selection unit

Claims (6)

An acoustic model generation device that generates an acoustic model of a target language for speech recognition,
A target language new acoustic model storage unit capable of storing a target language new acoustic model which is a second acoustic model of the target language;
One or more other language old acoustic models that are first acoustic models of one or more other languages different from the target language and one or more other languages that are second acoustic models of the one or more other languages. First correlation information that is information related to a new acoustic model, or information related to a relationship between the one or more other language old acoustic models and a target language old acoustic model that is a first acoustic model of the target language. An acoustic model generation unit that generates a target language new acoustic model from the target language old acoustic model or one or more other language new acoustic models using any one or more correlation information of the second correlation information;
An acoustic model storage unit that stores the target language new acoustic model generated by the acoustic model generation unit in the target language new acoustic model storage unit ;
The acoustic model generation unit
A target language old acoustic model storage unit capable of storing the target language old acoustic model; and
Another language new acoustic model storage unit capable of storing one or more other language new acoustic models;
A first correlation information storage unit capable of storing one or more pieces of first correlation information;
A second correlation information storage unit capable of storing one or more second correlation information;
A selection means for selecting any one of the two or more algorithms for generating the target language new acoustic model according to data of the target language old acoustic model or the one or more other language new acoustic models;
Using the one or more first correlation information and the one or more second correlation information, the target language old acoustic model, the one or more other language new acoustic models, or the target language old acoustic model and the one An acoustic model generation means for generating a target language new acoustic model from the above-mentioned other language new acoustic model,
The acoustic model generation means includes
According to the one algorithm selected by the selection unit, the target language new acoustic model is generated using one or more correlation information of the one or more first correlation information and the one or more second correlation information. Acoustic model generation device. The other language old acoustic model is an acoustic model generated with data different from the acoustic model before adaptation processing of another language or the other language new acoustic model,
The other language new acoustic model is an acoustic model generated with data different from the acoustic model after adaptive processing of another language or the other language old acoustic model,
The target language old acoustic model is an acoustic model generated with data similar to an acoustic model before adaptation processing of the target language or another language old acoustic model,
The target language new acoustic model, the acoustic model generating device according to claim 1, wherein an acoustic model after adaptation processing of the target language. The first correlation information is
Obtained from one or more transformation functions that are differences between one or more vectors corresponding to the one or more other language old acoustic models and one or more vectors corresponding to the one or more other language new acoustic models. Information
The second correlation information is
The acoustic model generation apparatus according to claim 2 , wherein the acoustic model generation apparatus is information acquired from one or more conversion functions of a difference between a vector corresponding to each of the one or more other language old acoustic models and a vector corresponding to the target language old acoustic model. . The acoustic model generation unit
The acoustic model generation apparatus according to claim 3 , wherein a target language new acoustic model is generated by mapping a vector corresponding to the target language old acoustic model using a conversion function of the first correlation information. The target language old acoustic model storage unit that can store the target language old acoustic model, the other language new acoustic model storage unit that can store one or more other language new acoustic models, and one or more first correlation information can be stored. An acoustic model generation method that can be realized by an acoustic model generation unit including a first correlation information storage unit and a second correlation information storage unit that can store one or more second correlation information , and an acoustic model storage unit. ,
The acoustic model generation unit includes at least one other language old acoustic model that is a first acoustic model of one or more other languages different from the target language and a second acoustic model of the one or more other languages. One or more first correlation information that is information related to a relationship with one or more other language new acoustic models, or a target language that is a first acoustic model of the one or more other language old acoustic models and the target language Using one or more correlation information of one or more second correlation information, which is information relating to the relationship with the old acoustic model, from the target language old acoustic model or one or more other language new acoustic models, the target language An acoustic model generation step for generating a new acoustic model;
The acoustic model storage unit includes an acoustic model storage step of storing the target language new acoustic model generated in the acoustic model generation step in a recording medium ,
In the acoustic model generation step,
A selection sub-step of selecting any one of the two or more algorithms for generating a target language new acoustic model according to data of the target language old acoustic model or the one or more other language new acoustic models; ,
Using the one or more first correlation information and the one or more second correlation information, the target language old acoustic model, the one or more other language new acoustic models, or the target language old acoustic model and the one An acoustic model generation substep for generating a target language new acoustic model from the above-mentioned other language new acoustic model,
In the acoustic model generation sub-step,
According to the one algorithm selected in the selection sub-step, the target language new acoustic model is determined using one or more correlation information of the one or more first correlation information and the one or more second correlation information acoustic model generating method to be generated. The target language old acoustic model storage unit that can store the target language old acoustic model, the other language new acoustic model storage unit that can store one or more other language new acoustic models, and one or more first correlation information can be stored. A computer accessible to the first correlation information storage unit and the second correlation information storage unit capable of storing one or more second correlation information ;
One or more other language old acoustic models that are first acoustic models of one or more other languages different from the target language and one or more other languages that are second acoustic models of the one or more other languages. One or more first correlation information that is information related to a new acoustic model, or a relationship between the one or more other language old acoustic models and a target language old acoustic model that is a first acoustic model of the target language Sound that generates a target language new acoustic model from the target language old acoustic model or one or more other language new acoustic models using any one or more correlation information of one or more second correlation information that is information A model generator,
Function as an acoustic model storage unit that stores the target language new acoustic model generated by the acoustic model generation unit in a recording medium ;
The acoustic model generation unit
A selection means for selecting any one of the two or more algorithms for generating the target language new acoustic model according to data of the target language old acoustic model or the one or more other language new acoustic models;
Using the one or more first correlation information and the one or more second correlation information, the target language old acoustic model, the one or more other language new acoustic models, or the target language old acoustic model and the one From the above other language new acoustic models, function as acoustic model generation means to generate the target language new acoustic model,
The acoustic model generating means;
According to the one algorithm selected by the selection unit, the target language new acoustic model is generated using one or more correlation information of the one or more first correlation information and the one or more second correlation information. as things, because of the program cause the computer to function.

Images