JP6903613B2 - Speech recognition device, speech recognition method and program - Google Patents

Description

Embodiments of the present invention relate to voice recognition devices, voice recognition methods and programs.

A speech recognition technique that recognizes speech data using an acoustic model and a language model and outputs the utterance text included in the speech data has been conventionally known. The acoustic model is learned in advance using a large amount of data (for example, several hundred hours or more). However, it is difficult to learn an acoustic model that can obtain a high recognition rate (for example, 85% or more) under any conditions. For example, when an acoustic model learned using voice data recorded in a clean environment is used, the recognition rate in a conference room with a large reverberation deteriorates. One of the effective methods to prevent the deterioration of the recognition rate is the adaptation of the acoustic model.

Japanese Patent No. 5852550

However, in the conventional technique, when the acoustic model is adapted, an adverse effect also occurs. For example, when utterances of the same content are repeated, the adaptation of the acoustic model makes it easier to recognize this utterance, but makes it difficult to recognize other utterances. Also, for example, voice data includes both voice and non-voice, but when there are many non-voice parts, the non-voice recognition result can be easily obtained by adapting the acoustic model, and the voice recognition result can be obtained. It becomes difficult. An object to be solved by the present invention is to provide a voice recognition device, a voice recognition method, and a program capable of suppressing adverse effects due to adaptation of an acoustic model.

The voice recognition device of the embodiment includes a generation unit, a determination unit, a selection unit, and an adaptation unit. The generation unit recognizes the voice data using the language model and the first acoustic model, and generates a label that identifies the utterance included in the voice data. The determination unit specifies the number of voice data including the same utterance using the label, and determines the weight to be given to the voice data according to the number. The selection unit selects the voice data based on the weight. The adapting unit generates a second acoustic model by adapting the first acoustic model using the voice data selected by the selection unit.

The block diagram which shows the example of the functional structure of the voice recognition apparatus of 1st Embodiment. The figure which shows the example of the label information of 1st Embodiment. The flowchart which shows the example of the operation method of the voice recognition apparatus of 1st Embodiment. The block diagram which shows the example of the functional structure of the voice recognition apparatus of 2nd Embodiment. The figure which shows the example of the voice data of 2nd Embodiment. The block diagram which shows the example of the functional structure of the voice recognition apparatus of 3rd Embodiment. The block diagram which shows the example of the functional structure of the voice recognition apparatus of 4th Embodiment. The block diagram which shows the example of the functional structure of the voice recognition apparatus of 5th Embodiment. The figure which shows the example of the hardware composition of the voice recognition apparatus of 1st to 5th Embodiment.

The voice recognition device, the voice recognition method, and the embodiment of the program will be described in detail with reference to the accompanying drawings.

First, the adaptation of the acoustic model will be described. The adaptation of the acoustic model is performed by re-learning the acoustic model using the adaptation data based on the learned acoustic model. Hereinafter, the acoustic model learned first is referred to as a base acoustic model (first acoustic model), and the adapted acoustic model is referred to as an adaptive acoustic model (second acoustic model).

There are roughly two ways to adapt the acoustic model (supervised adaptation and unsupervised adaptation). In supervised adaptation, adaptive data including both audio data and the correct label of the audio data is used. In unsupervised adaptation, adaptive data containing only audio data is used (there is no correct label).

Supervised adaptation is good for adaptation because it has a correct label, but it is expensive because it is necessary to create a correct label by transcribing.

On the other hand, unsupervised adaptation is less costly because it does not require the creation of correct labels. In unsupervised adaptation, speech data is recognized and the speech recognition result is used as a label. Basically, the higher the speech recognition accuracy, the better, because an error in the speech recognition result may adversely affect adaptation. In the conventional unsupervised adaptation method, speech is recognized using a language model and a bass acoustic model, and labels, reliability, and acoustic likelihood are output. In the conventional unsupervised adaptation method, the acoustic model is adapted by selecting speech data having higher reliability and lower acoustic likelihood.

(First Embodiment)
First, an example of the functional configuration of the voice recognition device 10 of the first embodiment will be described.

[Example of functional configuration]
FIG. 1 is a diagram showing an example of a functional configuration of the voice recognition device 10 of the first embodiment. The voice recognition device 10 of the first embodiment includes a generation unit 1, a determination unit 2, a selection unit 3, and an adaptation unit 4. Some or all the functions of the voice recognition device 10 may be realized by software (program) or by hardware.

Further, the voice recognition device 10 of the first embodiment stores the language model 101, the base acoustic model 102, and the adaptive acoustic model 103. The language model 101 is data that models the linguistic features of speech. The base acoustic model 102 and the adaptive acoustic model 103 are data modeling the acoustic characteristics of speech. The bass acoustic model 102 is the data initially learned. The adaptive acoustic model 103 is data obtained by re-learning the base acoustic model 102 using the adaptive data. A storage unit for storing the language model 101, the base acoustic model 102, and the adaptive acoustic model 103 may be provided in an external device.

The generation unit 1 recognizes the voice data using the language model 101 and the base acoustic model 102, and generates a label. The voice data is, for example, data separated for each utterance. The label is data converted from the voice recognition result of the voice data. The label is information that identifies the utterance contained in the voice data.

The determination unit 2 specifies the number of voice data including the same utterance using the label, and determines the weight to be given to the voice data according to the number.

The label and the number of the labels are stored in the voice recognition device 10 as label information shown in FIG. 2, for example.

FIG. 2 is a diagram showing an example of label information of the first embodiment. The label information of the first embodiment includes voice data, a voice recognition result, a label, a count number, and a weight.

The voice recognition result is a recognition result of voice data. In the example of FIG. 2, the label is data obtained by converting the voice recognition result into hiragana. The label is not limited to hiragana, but may be Roman characters or the like.

The count number indicates the number of labels. For example, the labels of utterance-1, utterance-3, and utterance-5 are the same. When the label of utterance-1 is generated, the count number of the label is 1. When the label of utterance-3 is generated, the count number of the label is 2. When the label of utterance-5 is generated, the count number of the label is 3.

The weight indicates the weight of the label. In the example of FIG. 2, the larger the number of labels counted, the smaller the weight of the label.

The generation unit 1 determines the weight of the label by, for example, the following equation (1).

μ = e ^1-x ... (1)

Here, μ is a weight and x is a count number. In the example of FIG. 2, the weight is determined by the equation (1). For example, the weights of the labels of utterance -1, utterance -3, and utterance -5 are 1.00, 0.37, and 0.14, respectively. The weight of the labels of utterance-2, utterance-4 and utterance-6 is 1.00.

The equation for determining the weight is not limited to the above equation (1), and other decreasing functions may be used.

Returning to FIG. 1, the selection unit 3 selects voice data (utterance) to be used as adaptive data based on the weight included in the label information generated by the generation unit 1. When there are multiple utterances with the same content in the adaptation data, the adaptation increases the posterior probability of the utterance with the same content, making it easier to recognize the utterance. On the other hand, in this case, since the posterior probability of other utterances is low, it becomes difficult to recognize other utterances.

Therefore, the selection unit 3 compares the weight of each utterance with the weight threshold value, and selects an utterance larger than the weight threshold value as adaptive data. As a result, the adverse effect of voice recognition can be suppressed by using the adaptive acoustic model 103 generated using the adaptive data.

The weight threshold is determined by, for example, the following equation (2).

θ = e ^1-αn ... (2)

Here, θ is a weight threshold, α is an utterance coefficient, and n is the total number of utterances. That is, for utterances having the same content, when the count number x is smaller than α times the total number of utterances n (x <αn), it is selected as adaptive data.

The utterance coefficient α is, for example, 0.2. In the example of FIG. 2, since the total number of utterances n is 6, the weight threshold θ is 0.82. Since the weights of utterance-1, utterance-2, utterance-3, and utterance-5 are larger than the weight threshold value θ, they are selected as adaptive data by the selection unit 3. On the other hand, since the weights of utterance -4 and utterance -6 are smaller than the weight threshold value θ, they are not selected as adaptive data by the selection unit 3.

In the description of the first embodiment, the case where the utterance coefficient is 0.2 has been described, but if necessary, another numerical value of 1 or less may be set as the utterance coefficient. Further, the weight threshold θ may be determined based on the absolute number of utterances (total number of utterances n) instead of the ratio αn of the total number of utterances n. In this case, αn in the above equation (2) may be changed to n.

The adaptive unit 4 generates the adaptive acoustic model 103 by adapting the base acoustic model 102 using the adaptive data selected by the selection unit 3. Specifically, the adaptation of the base acoustic model 102 is performed by optimizing the parameters of the base acoustic model 102 using the adaptation data. As a method of adapting the bass acoustic model 102, for example, there is a method using DNN (Deep Neural Network), CNN (Convolutional Neural Network), RNN (Recurrent Neural Network) and the like. The adaptive acoustic model 103 may be stored in an external storage unit of the voice recognition device 10.

[Example of operation method]
FIG. 3 is a flowchart showing an example of an operation method of the voice recognition device 10 of the first embodiment. First, the generation unit 1 recognizes voice data using the language model 101 and the base acoustic model 102 (step S1). Next, the generation unit 1 generates a label that identifies the utterance included in the voice data recognized by the process of step S1 (step S2).

Next, the determination unit 2 specifies the number of voice data including the same utterance using the label, and determines the weight to be given to the voice data according to the number (step S3). Next, the selection unit 3 selects voice data to be used as adaptive data based on the weight (step S4). Next, the adaptation unit 4 generates the adaptive acoustic model 103 by adapting the base acoustic model 102 using the voice data (adaptive data) selected by the selection unit 3 (step S5).

As described above, in the voice recognition device 10 of the first embodiment, the generation unit 1 recognizes the voice data by using the language model 101 and the base sound model 102 (first sound model), and uses the voice data as the voice data. Generates a label that identifies the included speech. The determination unit 2 specifies the number of voice data including the same utterance using the label, and determines the weight to be given to the voice data according to the number. The selection unit 3 selects voice data based on the weight. Then, the adaptive unit 4 adapts the base acoustic model 102 (first acoustic model) using the voice data (adaptive data) selected by the selection unit 3, thereby causing the adaptive acoustic model 103 (second acoustic model). To generate.

As a result, according to the voice recognition device 10 of the first embodiment, it is possible to suppress the adverse effect on voice recognition that occurs when the acoustic model is adapted.

(Second Embodiment)
Next, the second embodiment will be described. In the description of the second embodiment, the same description as that of the first embodiment will be omitted.

The more non-speech parts included in the adaptive data, the higher the probability of non-speech (lower the probability of speech) due to the adaptation using the adaptive data, so that the speech recognition result may become non-speech. More. On the other hand, the smaller the non-speech part included in the adaptive data, the lower the probability of non-speech (higher the probability of speech) due to the adaptation using the adaptive data, so that the non-speech recognition result becomes speech. There are many things.

In the case of supervised learning, the amount of data in the non-voice part can be controlled because the voice data is manually cut out for each utterance. On the other hand, in the case of unsupervised learning, it is difficult to control the amount of data in the non-voice portion because the utterance is automatically cut out by the voice section detection process such as VAD (voice activity detection).

In the second embodiment, a configuration capable of suppressing the adverse effect of adaptation even when there are many voice (or non-voice) parts included in the voice data will be described.

[Example of functional configuration]
FIG. 4 is a block diagram showing an example of the functional configuration of the voice recognition device 10-2 of the second embodiment. The voice recognition device 10-2 of the second embodiment includes a generation unit 1, a selection unit 3-2, an adaptation unit 4, and a calculation unit 5. The description of the generation unit 1 and the adaptation unit 4 will be omitted because they are the same as those in the first embodiment.

The calculation unit 5 calculates the ratio of the audio frame included in the audio data to the non-audio frame included in the audio data by using the label generated by the generation unit 1.

FIG. 5 is a diagram showing an example of voice data of the second embodiment. In the example of FIG. 5, the case where the number of frames included in the voice data is 20 is shown. The 1, 2, 18, 19 and 20th frames are examples of non-audio frames. Sil is an abbreviation for silence. The 3rd to 17th frames are audio frames. The utterance included in the voice data of FIG. 5 is "Good morning", and the label of the utterance is also "Good morning".

The calculation unit 5 performs alignment using the generated label in order to represent the phoneme for each frame. Depending on the length of the pronunciation, one phoneme may correspond to two or more frames. In the example of FIG. 5, for example, the phonemes corresponding to frames 4 and 5 are the same.

The calculation unit 5 calculates the ratio of the audio frame to the non-audio frame. In the example of FIG. 5, the ratio of audio frames is 15/20 = 0.75. The ratio of non-audio frames is 5/20 = 0.25.

Returning to FIG. 4, the selection unit 3-2 selects audio data in which the ratio of audio frames is within a predetermined selection range as adaptive data. The predetermined selection range is, for example, 0.3 to 0.9. In the example of FIG. 5, since the ratio of audio frames is 0.75, the audio frames are selected as adaptive data by the selection unit 3-2.

The predetermined selection range may be set according to the purpose of adaptation. When it is desired to output the voice recognition result as much as possible from the voice recognition device 10-2, a range of a section having a higher value is used as a predetermined selection range (for example, 0.4 to 1.0). On the other hand, since the voice data contains background noise, if the voice recognition device 10-2 does not want to output the background noise recognition result as much as possible, a range of a section having a lower value is used as a predetermined selection range. (For example, 0.0 to 0.5).

As described above, according to the voice recognition device 10-2 of the second embodiment, even when voice data having a high ratio of non-voice frames (for example, 0.7 or more) is included, the selection unit 3- Due to 2, the voice data is not selected. As a result, it is possible to suppress an adverse effect on the speech recognition result using the adaptive speech model 103.

(Third Embodiment)
Next, the third embodiment will be described. In the description of the third embodiment, the same description as in the first and second embodiments will be omitted. In the third embodiment, the operation when the first and second embodiments are combined will be described.

[Example of functional configuration]
FIG. 6 is a block diagram showing an example of the functional configuration of the voice recognition device 10-3 according to the third embodiment. The voice recognition device 10-3 of the third embodiment includes a generation unit 1, a determination unit 2, a selection unit 3-3, an adaptation unit 4, and a calculation unit 5. The description of the generation unit 1, the determination unit 2, and the adaptation unit 4 will be omitted because they are the same as those in the first embodiment. Since the description of the calculation unit 5 is the same as that of the second embodiment, the description thereof will be omitted.

The method of selecting adaptive data by the selection unit 3 of the first embodiment is referred to as the selection method A, and the method of selecting the adaptive data by the selection unit 3-2 of the second embodiment is referred to as the selection method B. Selection methods A and B are independent. Therefore, it is possible to select the voice data to be used as the adaptive data by the combination of the selection methods A and B.

The selection unit 3-3 selects audio data to be used as adaptive data based on the weight determined by the determination unit 2 and the ratio of the audio frames calculated by the calculation unit 5. Specifically, the selection unit 3-3 selects, for example, the adaptation data candidate by the selection method A, and then selects the adaptation data from the adaptation data candidates by the selection method B. Further, for example, the selection unit 3-3 selects the adaptation data candidate by the selection method B, and then selects the adaptation data from the adaptation data candidates by the selection method A.

As a result, according to the voice recognition device 10-3 of the third embodiment, the effects of the first and second embodiments can be obtained.

(Fourth Embodiment)
Next, the fourth embodiment will be described. In the description of the fourth embodiment, the same description as that of the first embodiment will be omitted. In the fourth embodiment, a configuration for performing voice recognition will be described using the adaptive acoustic model 103.

[Example of functional configuration]
FIG. 7 is a diagram showing an example of the functional configuration of the voice recognition device 10-4 according to the fourth embodiment. The voice recognition device 10-4 of the fourth embodiment includes a generation unit 1, a determination unit 2, a selection unit 3, an adaptation unit 4, and a recognition unit 6. The description of the generation unit 1, the determination unit 2, the selection unit 3, and the adaptation unit 4 will be omitted because they are the same as those in the first embodiment.

The recognition unit 6 uses the language model 101 and the adaptive acoustic model 103 to perform voice recognition of voice data. For example, when performing voice recognition of voice data acquired in an environment similar to the environment in which the adaptive data is acquired, the parameters of the adaptive acoustic model 103 are preferable to the parameters of the base acoustic model 102. Further, for example, in the case of voice recognition of voice data of a speaker (or the same speaker) similar to the speaker of the utterance included in the adaptive data, the parameters of the adaptive acoustic model 103 are preferable to the parameters of the base acoustic model 102. Therefore, when voice recognition is performed using the adaptive acoustic model 103, higher voice recognition accuracy can be obtained.

(Fifth Embodiment)
Next, the fifth embodiment will be described. In the description of the fifth embodiment, the same description as that of the first embodiment will be omitted. In the first embodiment, adaptation is performed using two types of models, a language model 101 and a bass acoustic model 102. In the fifth embodiment, the configuration in the case of performing adaptation by the end-to-end voice recognition method without distinguishing between the language model 101 and the bass acoustic model 102 will be described.

[Example of functional configuration]
FIG. 8 is a diagram showing an example of the functional configuration of the voice recognition device 10-5 according to the fifth embodiment. The voice recognition device 10-5 of the fifth embodiment includes a generation unit 1-2, a determination unit 2, a selection unit 3, and an adaptation unit 4-2. The description of the determination unit 2 and the selection unit 3 will be omitted because they are the same as those in the first embodiment.

The voice recognition device 10-5 of the fifth embodiment stores the voice recognition base model 104 and the voice recognition adaptation model 105. The speech recognition-based model 104 is data modeled without distinguishing between the linguistic features of speech and the acoustic features of speech.

The generation unit 1-2 recognizes the voice data using the voice recognition base model 104 and generates a label. In the fifth embodiment, the speech recognition base model 104 plays the role of the language model 101 and the base acoustic model 102. The description of the label generation method is the same as that of the first embodiment, and thus the description thereof will be omitted.

The adaptation unit 4-2 generates the voice recognition adaptation model 105 by adapting the voice recognition base model 104 using the adaptation data selected by the selection unit 3. Specifically, the adaptation of the speech recognition base model 104 is performed by optimizing the parameters of the speech recognition base model 104 using the adaptation data. As a method of adapting the speech recognition base model 104, for example, there is a method of using DNN, CNN, RNN (Recurrent Neural Network) and the like. The voice recognition adaptation model 105 may be stored in an external storage unit of the voice recognition device 10.

Finally, an example of the hardware configuration of the voice recognition device 10 (10-2, 10-3, 10-4, 10-5) of the first to fifth embodiments will be described.

[Example of hardware configuration]
FIG. 9 is a diagram showing an example of the hardware configuration of the voice recognition device 10 (10-2, 10-3, 10-4, 10-5) of the first to fifth embodiments. Hereinafter, the case of the voice recognition device 10 of the first embodiment will be described as an example. The hardware configuration of the voice recognition device 10-2 (10-3, 10-4, 10-5) of the second to fifth embodiments is also the same as the hardware configuration of the voice recognition device 10 of the first embodiment. The same is true.

The voice recognition device 10 of the first embodiment includes a control device 301, a main storage device 302, an auxiliary storage device 303, a display device 304, an input device 305, and a communication device 306. The control device 301, the main storage device 302, the auxiliary storage device 303, the display device 304, the input device 305, and the communication device 306 are connected via the bus 310.

The control device 301 executes the program read from the auxiliary storage device 303 to the main storage device 302. The main storage device 302 is a memory such as a ROM (Read Only Memory) and a RAM (Random Access Memory). The auxiliary storage device 303 is an HDD (Hard Disk Drive), a memory card, or the like.

The display device 304 displays the display information. The display device 304 is, for example, a liquid crystal display or the like. The input device 305 is an interface for operating the voice recognition device 10. The input device 305 is, for example, a keyboard, a mouse, or the like. When the voice recognition device 10 is a smart device such as a smartphone or a tablet terminal, the display device 304 and the input device 305 are, for example, a touch panel. The communication device 306 is an interface for communicating with another device.

The program executed by the voice recognition device 10 of the first embodiment is a file in an installable format or an executable format, which is read by a computer such as a CD-ROM, a memory card, a CD-R, and a DVD (Digital Versaille Disc). It is recorded on a possible storage medium and provided as a computer program product.

Further, the program executed by the voice recognition device 10 of the first embodiment may be stored on a computer connected to a network such as the Internet and provided by downloading via the network. Further, the program executed by the voice recognition device 10 of the first embodiment may be configured to be provided via a network such as the Internet without being downloaded.

Further, the program of the voice recognition device 10 of the first embodiment may be configured to be provided by incorporating it into a ROM or the like in advance.

The program executed by the voice recognition device 10 of the first embodiment has a module configuration including functional blocks that can be realized by the program among the above-mentioned functional blocks. As the actual hardware, each functional block is loaded on the main storage device 302 by the control device 301 reading a program from the storage medium and executing the program. That is, each of the above functional blocks is generated on the main storage device 302.

It should be noted that a part or all of the above-mentioned functional blocks may not be realized by software, but may be realized by hardware such as an IC (Integrated Circuit).

Further, when each function is realized by using a plurality of processors, each processor may realize one of each function, or may realize two or more of each function.

Further, the operation mode of the voice recognition device 10 of the first embodiment may be arbitrary. The voice recognition device 10 of the first embodiment may be operated as, for example, a cloud system on a network.

Although some embodiments of the present invention have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These novel embodiments can be implemented in various other embodiments, and various omissions, replacements, and changes can be made without departing from the gist of the invention. These embodiments and modifications thereof are included in the scope and gist of the invention, and are also included in the scope of the invention described in the claims and the equivalent scope thereof.

1 Generation unit 2 Decision unit 3 Selection unit 4 Adaptation unit 5 Calculation unit 6 Recognition unit 101 Language model 102 Base acoustic model 103 Adaptation acoustic model 104 Speech recognition base model 105 Speech recognition adaptive model 301 Control device 302 Main storage device 303 Auxiliary storage device 304 Display device 305 Input device 306 Communication device 310 Bus

Claims (13)

A generation unit that recognizes voice data using a language model and a first acoustic model and generates a label that identifies an utterance included in the voice data.
Using the label, a determination unit that specifies the number of voice data including the same utterance and determines the weight to be given to the voice data according to the number.
A selection unit that selects the voice data based on the weight, and
An adaptation unit that generates a second acoustic model by adapting the first acoustic model using the voice data selected by the selection unit.
A voice recognition device equipped with. A generation unit that recognizes voice data using a language model and a first acoustic model and generates a label that identifies an utterance included in the voice data.
Using the label, a calculation unit that calculates the ratio of the audio frame included in the audio data to the non-audio frame included in the audio data.
A selection unit that selects audio data for which the ratio of audio frames is within a predetermined selection range, and
An adaptation unit that generates a second acoustic model by adapting the first acoustic model using the voice data selected by the selection unit.
A voice recognition device equipped with. A generation unit that recognizes voice data using a language model and a first acoustic model and generates a label that identifies an utterance included in the voice data.
Using the label, a determination unit that specifies the number of voice data including the same utterance and determines the weight to be given to the voice data according to the number.
Using the label, a calculation unit that calculates the ratio of the audio frame included in the audio data to the non-audio frame included in the audio data.
A selection unit that selects the audio data based on the weight and the ratio of the audio frames.
An adaptation unit that generates a second acoustic model by adapting the first acoustic model using the voice data selected by the selection unit.
A voice recognition device equipped with. The determination unit determines the weight to be smaller as the number of the determination units increases.
The voice recognition device according to claim 1. The selection unit determines whether or not the weight is larger than the threshold value, and selects voice data to which a weight larger than the threshold value is given.
The voice recognition device according to claim 1. A recognition unit that performs voice recognition of the voice data using the language model and the second acoustic model.
The voice recognition device according to claim 1. The language model and the first acoustic model are represented by a speech recognition-based model modeled without distinguishing between the linguistic features of speech and the acoustic features of speech.
The adapting unit adapts the voice recognition base model using the voice data selected by the selection unit.
The voice recognition device according to claim 1. A step of recognizing voice data using a language model and a first acoustic model and generating a label for identifying an utterance included in the voice data.
A step of specifying the number of voice data containing the same utterance using the label and determining the weight to be given to the voice data according to the number.
The step of selecting the voice data based on the weight, and
A step of generating a second acoustic model by adapting the first acoustic model using the voice data selected by the selected step, and a step of generating the second acoustic model.
Speech recognition methods including. A step of recognizing voice data using a language model and a first acoustic model and generating a label for identifying an utterance included in the voice data.
Using the label, a step of calculating the ratio of the audio frame included in the audio data to the non-audio frame included in the audio data, and
A selection unit that selects audio data for which the ratio of audio frames is within a predetermined selection range, and
A step of generating a second acoustic model by adapting the first acoustic model using the voice data selected by the selected step, and a step of generating the second acoustic model.
Speech recognition methods including. A step of recognizing voice data using a language model and a first acoustic model and generating a label for identifying an utterance included in the voice data.
A step of specifying the number of voice data containing the same utterance using the label and determining the weight to be given to the voice data according to the number.
Using the label, a step of calculating the ratio of the audio frame included in the audio data to the non-audio frame included in the audio data, and
A step of selecting the audio data based on the weight and the ratio of the audio frames,
A step of generating a second acoustic model by adapting the first acoustic model using the voice data selected by the selected step, and a step of generating the second acoustic model.
Speech recognition methods including. Computer,
A generation unit that recognizes voice data using a language model and a first acoustic model and generates a label that identifies an utterance included in the voice data.
Using the label, a determination unit that specifies the number of voice data including the same utterance and determines the weight to be given to the voice data according to the number.
A selection unit that selects the voice data based on the weight, and
An adaptation unit that generates a second acoustic model by adapting the first acoustic model using the voice data selected by the selection unit.
A program to function as. Computer,
A generation unit that recognizes voice data using a language model and a first acoustic model and generates a label that identifies an utterance included in the voice data.
Using the label, a calculation unit that calculates the ratio of the audio frame included in the audio data to the non-audio frame included in the audio data.
A selection unit that selects audio data for which the ratio of audio frames is within a predetermined selection range, and
An adaptation unit that generates a second acoustic model by adapting the first acoustic model using the voice data selected by the selection unit.
A program to function as. Computer,
A generation unit that recognizes voice data using a language model and a first acoustic model and generates a label that identifies an utterance included in the voice data.
Using the label, a determination unit that specifies the number of voice data including the same utterance and determines the weight to be given to the voice data according to the number.
Using the label, a calculation unit that calculates the ratio of the audio frame included in the audio data to the non-audio frame included in the audio data.
A selection unit that selects the audio data based on the weight and the ratio of the audio frames.
An adaptation unit that generates a second acoustic model by adapting the first acoustic model using the voice data selected by the selection unit.
A program to function as.

Images