JP2021026045A - Storage device, storage method and program - Google Patents

Abstract

To provide a technology for storing utterance data used for learning an acoustic model by concealing the content.SOLUTION: A storage device 1 for storing utterance data which is utterance data formed of data on voice and a text corresponding to the voice, and which is used for learning an acoustic model comprises: an input unit 10 for accepting input if a plurality of pieces of original utterance data formed of the voice and the text corresponding to the voice; a fragment data generation unit 11 for dividing the original utterance data and generating a plurality of pieces of fragment data formed of the voice including one or more clauses and the text corresponding to the voice; a fragment data connection unit 12 which randomly connects the plurality of pieces of fragment data generated from the plurality of pieces of original utterance data and generates the plurality of utterance data in prescribed length; and a storage unit 13 for storing the plurality of utterance data generated by the fragment data connection unit 12.SELECTED DRAWING: Figure 1

Description

The present invention relates to a technique for storing utterance data used for learning an acoustic model for speech recognition.

When creating an acoustic model for speech recognition, it is effective to perform learning using the speech data in the domain that is actually used, and this can significantly improve the performance of the acoustic model. However, since the data in the domain contains personal information, it may infringe on the privacy of the speaker. In addition, the data in the domain may identify the speaker as belonging to a certain group. Therefore, in general, the data in the domain is destroyed after the usage period.

However, once the data is destroyed, it cannot be retrained even if a more effective model structure is proposed in the future. It is desirable to be able to conceal the voice data in the domain and store it in a privacy-protected state.

Generally, there is a method of adding specific noise. For example, Patent Document 1 describes a method for defining conditions for deteriorating data at the time of data disclosure in order to reduce the risk of the data owner and the data analyst when the data owner discloses the data to the data analyst. Is disclosed. Further, Non-Patent Documents 1 and 2 disclose a method of concealing the procedure of the calculation method.

Japanese Unexamined Patent Publication No. 2018-128913

P. Smaragdis and M. Shashanka, “A framework for secure speech recognition,” IEEE Transactions on Audio, Speech, Language Processing, 15, 1404--1413 (2007). M.A. Pathak, B. Raj, S. Rane, and P. Smaragdis, “Privacy-preserving speech processing,” IEEE Signal Processing Magazine, 62--74 (2013).

The method of adding noise to the data described in Patent Document 1 described above is effective when the data provider and the data user are different and the data provider does not need data protection. However, when the data provider and the data user are the same, there is no point in protecting the data if the added noise is known.

Further, the methods described in Non-Patent Documents 1 and 2 require a large amount of calculation as compared with the case where concealment is not performed, and it is necessary to change the operation protocol when changing the model. Further, the methods described in Non-Patent Documents 1 and 2 cannot be used for data protection.
Therefore, in view of the above background, an object of the present invention is to provide a technique capable of concealing and storing the contents of utterance data used for learning an acoustic model.

The storage device of the present invention is a storage device that stores utterance data composed of data related to voice and text corresponding to the voice, and is used for learning an acoustic model, and corresponds to the voice and the voice. An input unit for inputting a plurality of original utterance data consisting of text, and a fragment for dividing the original utterance data to generate a plurality of fragment data consisting of a voice containing one or more phrases and text corresponding to the voice. The data generation unit, the fragment data combination unit that randomly combines a plurality of fragment data generated from the original speech data to generate a plurality of speech data of a predetermined length, and the fragment data combination unit. It is provided with a storage unit for storing a plurality of utterance data generated in the above.

The contents of the original utterance data are generated by dividing the plurality of original utterance data input in this way to generate a plurality of fragment data, and randomly combining the plurality of fragment data to generate new utterance data. Can be obscured. Further, since the fragment data includes one or more clauses, the time series of the acoustic features is stored, and the generated utterance data can be used for learning the acoustic model. Here, the "bunsetsu" is a sentence delimiter unit composed of one or more independent words and zero or more attached words.

The storage device of the present invention includes an acoustic feature amount generation unit that generates an acoustic feature amount from the voice input from the input unit, and the storage unit stores the acoustic feature amount of the utterance data as data related to the voice. You may. By storing the acoustic features as data related to speech in this way, it is possible to make it difficult to estimate the combined position of the fragment data constituting the utterance data based on the background noise and the like.

The storage device of the present invention includes a feature amount normalization unit that normalizes the acoustic feature amount generated by the acoustic feature amount generation unit, and the storage unit normalizes the speech data as data related to the voice. The generated acoustic features may be stored. By normalizing the acoustic features in this way, it is possible to make it more difficult to estimate the connection position of the fragment data constituting the utterance data.

In the storage device of the present invention, the input unit receives input of a speaker identifier of the utterance data together with a plurality of the original utterance data, and the storage unit receives the utterance data generated by the fragment data combination unit. May be stored in association with the speaker's identifier. By storing the speaker identifier of the newly generated utterance data in this way, the use of the stored data is expanded.

In the storage device of the present invention, the input unit accepts input of a speaker identifier of the utterance data together with a plurality of the original utterance data, and the fragment data combining unit is generated based on the speaker identifier. Fragment data obtained from a plurality of speakers may be included in each utterance data to be made. By including the fragment data obtained from the utterance data of a plurality of speakers in the newly generated utterance data in this way, it is possible to make it difficult to identify the speaker from the utterance data.

The storage device of the present invention applies a speaker identification technique to a plurality of the original speech data to obtain a speaker feature amount calculation unit for obtaining a speaker feature amount, and the story based on the speaker feature amount. A clustering unit for clustering people may be provided, and the fragment data combining unit may include fragment data obtained from speech data of a plurality of speakers included in the same cluster in each generated speech data. With this configuration, utterance data in which the voices of a plurality of speakers having similar features are combined is generated, so that it becomes more difficult to identify the speaker from the utterance data. As the feature amount of the speaker, for example, an i-vector created from factor analysis or an x-vector / d-vector obtained from the output of the intermediate layer of the speaker identification network can be used.

The storage device of the present invention includes an acoustic feature amount generation unit that generates an acoustic feature amount from the voice, and a feature amount normalization unit that normalizes the acoustic feature amount of speech data of a plurality of speakers included in the same cluster. To be equipped. By normalizing the acoustic features in this way, it is possible to make it more difficult to identify the speaker from the utterance data.

In the storage device of the present invention, the fragment data generation unit performs morphological analysis and syntactic analysis on the input text to obtain a clause, and based on the break position of the clause, fragment data including one or more clauses. May be generated. By using morphological analysis and parsing for the text in this way, it is possible to appropriately obtain the phrase of the utterance data.

In the storage device of the present invention, the fragment data generation unit detects a non-pronounced section in the input voice, sets the non-pronounced section as a phrase delimiter position, and sets one or more phrases based on the phrase delimiter position. Fragment data containing may be generated. Since the unpronounced section of the voice data often coincides with the phrase delimiter, the fragment data can be generated by using the non-pronounced section as the delimiter position of the fragment data. In addition, by dividing into non-pronunciation sections, it is possible to reduce the influence of phonological changes due to vocalization deformation caused by the connection of words.

In the storage device of the present invention, the fragment data combining unit may set the predetermined length of the utterance data to be generated based on the distribution of the lengths of the plurality of original utterance data. The fragment data generated from the utterance data loses the phoneme information that was continuous in the original utterance data. Specifically, in utterance data without a non-pronounced section, phonemes are continuous from the beginning to the end of the utterance sentence, so it is the utterance sentence that contains NULL indicating that there is no phoneme before or after. There are only two places, the beginning and the end. If this utterance data is divided into N, for example, there will be NULL before and after each fragment data, so that 2 × N (pieces) of NULL will be generated. In the present invention, since the utterance data of a predetermined length is generated based on the distribution of the lengths of the plurality of input utterance data, the number of NULL indicating that there is no phoneme can be made substantially the same as the original utterance data. Here, "based on the distribution of the lengths of the plurality of original utterance data" means an embodiment in which the average length of the plurality of utterance data is used, or an embodiment having the same distribution as the original plurality of utterance data. It may be.

The storage device of the present invention includes a language model generation unit that generates a language model based on the text stored in the storage unit, and a language that applies the language model to the text of the original utterance data to perform language prediction. It may be provided with a prediction unit. With this configuration, the degree of concealment of the generated utterance data can be quantitatively determined. The degree of concealment is low if the original utterance data can be predicted accurately by the language model generated from the text of the generated utterance data, and conversely, the degree of concealment if the original utterance data cannot be predicted. Can be judged to be high. For example, the reciprocal of the perplexity of the language model can be used for the prediction accuracy.

The storage device of the present invention is a language model generation unit that generates a language model based on the text of the original utterance data, and a language that applies the language model to the text stored in the storage unit to perform language prediction. It may be provided with a prediction unit. With this configuration, the degree of concealment of the generated utterance data can be quantitatively determined. If the generated utterance data can be predicted accurately by the language model generated from the text of the original utterance data, the degree of concealment is low, and conversely, if the generated utterance data cannot be predicted, concealment is performed. It can be judged that the degree is high.

In the storage device of the present invention, when the prediction accuracy by the language prediction unit is higher than a predetermined threshold value, the fragment data generation unit may shorten the length of the fragment data. With this configuration, when the prediction accuracy is high, that is, when the utterance data is not sufficiently concealed, the utterance data can be concealed by making the fragment data finer and concealing it. The predetermined threshold value may be an absolute value of prediction accuracy or a relative value. When using relative values, prepare a language model generated from the original utterance data and a language model generated from the generated utterance data, and combine the original text or fragment data with each language model. Make predictions and compare their prediction accuracy. When the difference in prediction accuracy is larger than a predetermined threshold value, it can be determined that the utterance data is sufficiently concealed.

The storage method of the present invention is a method of storing utterance data including data related to voice and text corresponding to the voice in a storage device, which is used for learning an acoustic model, and is stored in the voice and the voice. A step of inputting a plurality of original utterance data consisting of corresponding texts and the original utterance data are divided to generate a plurality of fragment data consisting of a voice containing one or more phrases and a text corresponding to the voice. The storage device includes a step, a step of randomly combining a plurality of fragment data generated from the plurality of original utterance data to generate a plurality of utterance data having a predetermined length, and a plurality of generated utterance data. It has a step to save to.

The program of the present invention is utterance data composed of data related to voice and text corresponding to the voice, and is a program for storing utterance data used for learning an acoustic model in a storage device, and is a program for storing voice in a computer. A step of inputting a plurality of original utterance data consisting of text corresponding to the voice and the original utterance data, and a plurality of fragments consisting of a voice containing one or more phrases and text corresponding to the voice by dividing the original utterance data. A step of generating data, a step of randomly combining a plurality of fragment data generated from the plurality of original utterance data to generate a plurality of utterance data of a predetermined length, and a plurality of generated utterance data. Is executed in the storage device.

According to the present invention, it is possible to save the contents of the original utterance data in a state where the contents are unknown.

It is a figure which shows the structure of the storage apparatus of 1st Embodiment. It is a figure which shows the structure of the storage apparatus of the 2nd Embodiment. It is a figure which shows the structure of the storage apparatus of the 3rd Embodiment. It is a figure which shows the structure of the storage apparatus of 4th Embodiment. It is a figure which shows the structure of the storage apparatus of 5th Embodiment. It is a figure which shows the structure of the storage apparatus of 6th Embodiment. (A) It is a conceptual diagram which shows that the speaker vector v1 -v9 is obtained based on the utterance data of the speaker 1-speaker 9. (B) It is a figure which shows the example which clustered the speaker 1 to the speaker 9 based on the speaker vector v1 to v9. It is a figure which shows the structure of the storage apparatus of 7th Embodiment. It is a figure which shows the structure of the storage apparatus of 8th Embodiment.

Hereinafter, the storage device according to the embodiment of the present invention will be described with reference to the drawings.
(First Embodiment)
FIG. 1 is a diagram showing a configuration of a storage device 1 according to the first embodiment. The storage device 1 of the first embodiment is a storage device 1 that stores utterance data including data related to voice and text corresponding to the voice, and is used for learning an acoustic model.

The storage device 1 is predetermined by randomly combining the input unit 10 for inputting the original utterance data, the fragment data generation unit 11 for dividing the original utterance data to generate a plurality of fragment data, and the plurality of fragment data. It includes a fragment data combining unit 12 that generates a plurality of utterance data having a length of the above, and a storage unit 13 that stores a plurality of utterance data generated by the fragment data combining unit 12.

Hereinafter, the operation of the storage device 1 and the details of each configuration of the storage device 1 will be described. The storage device 1 receives the input of the original utterance data from the input unit 10. The original utterance data input from the input unit 10 is, for example, utterance data in the domain acquired in the target domain, and is composed of a voice and a text corresponding to the voice. The text corresponding to the voice is, for example, a text transcribed from the voice.

Next, the fragment data generation unit 11 divides the input original utterance data and generates a plurality of fragment data. When dividing the original utterance data, the fragment data generation unit 11 divides the phrase into the minimum unit. That is, the fragment data is divided so as to include at least one clause. As a method for the fragment data generation unit 11 to divide the utterance data and generate the fragment data, for example, the text of the utterance data is subjected to morphological analysis, the utterance data is divided into words, and then the parsing is performed. A possible method is to cut at the part that belongs to the same syntax tree. Further, instead of parsing, the part where the particle such as a particle appears may be cut as a delimiter position.

Further, the fragment data generation unit 11 detects a non-pronounced section (also referred to as “pause”) in the input voice, sets the non-pronounced section as a phrase break position, and uses one or more clauses based on the phrase break position. Fragment data containing may be generated. Specifically, the fragment data generation unit 11 aligns the phonemes, identifies the non-pronounced section as a break in the phrase, and divides the utterance data. Further, the break of the clause may be determined by the OR / AND of the division portion obtained by the above method.

Subsequently, the fragment data combining unit 12 randomly combines the fragment data generated by the fragment data generating unit 11 to generate new utterance data. Here, if only the fragment data obtained from one original utterance data is randomly combined, the order is only changed in the original utterance data, so that the original utterance data may be inferred. high. The fragment data combining unit 12 conceals the contents of the original utterance data by randomly combining the fragment data obtained from the plurality of original utterance data.

The length of the utterance data newly generated by the fragment data combining unit 12 is set based on the lengths of the plurality of original utterance data. For example, the average value of the lengths of a plurality of original utterance data may be obtained, and the average value may be rounded off. Alternatively, the length distribution (mean, variance) of a plurality of original utterance data may be obtained, and the length of the new utterance data may be set so as to have the same distribution.
Finally, the new utterance data generated by the fragment data combining unit 12 is stored in the storage unit 13.

The storage device 1 of the present embodiment divides a plurality of original utterance data into fragment data, randomly combines the plurality of fragment data to generate new utterance data, and stores the new utterance data. The utterance data can be saved without knowing the contents of the original utterance data. Here, since the fragment data contains one or more clauses and the time series of the acoustic features is retained even in the new utterance data, the acoustic model can be learned using the saved utterance data. It can be carried out.

Further, since the storage device 1 of the present embodiment sets the lengths of a plurality of newly generated utterance data so as to be the same as the average or distribution of the lengths of the original utterance data, it is newly generated. The number of first and last characters contained in the plurality of utterance data to be generated can be made equal to the number of first and last characters contained in the plurality of original utterance data. As a result, the stored utterance data can be used for learning similar to the original utterance data.

(Second Embodiment)
FIG. 2 is a diagram showing a configuration of a storage device 2 of a second embodiment, which is a modification of the first embodiment. The basic configuration of the storage device 2 of the second embodiment is the same as that of the first embodiment, but the storage device 2 of the second embodiment is based on the voice input from the input unit 10. It is provided with an acoustic feature amount generation unit 14 that generates an acoustic feature amount. The acoustic feature amount generation unit 14 inputs the acoustic feature amount generated from the voice to the fragment data generation unit 11. The fragment data generation unit 11 obtains a position for dividing the utterance data in order to generate the fragment data based on the text of the original utterance data or a non-sounding section, and divides the acoustic feature data at the dividing position to divide the fragment data. Generate.

The fragment data generation unit 11 passes the fragment data (consisting of an acoustic feature amount and the corresponding text) to the fragment data combination unit 12. The fragment data combining unit 12 randomly combines the fragment data to generate new utterance data and stores it in the storage unit 13, similarly to the storage device 2 of the first embodiment. The utterance data stored by the storage device 2 of the present embodiment includes acoustic features as data related to voice.

The storage device 2 of the second embodiment has been described above. The storage device 2 of the second embodiment can store the utterance data in a state where the contents are concealed, as in the first embodiment. Further, by storing the acoustic feature amount as the data related to the voice, it is possible to make it difficult to estimate the combined position of the fragment data constituting the utterance data based on the background noise or the like.

(Third Embodiment)
FIG. 3 is a diagram showing a configuration of a storage device 3 of a third embodiment, which is a modification of the first embodiment. The basic configuration of the storage device 3 of the third embodiment is the same as that of the second embodiment, but the storage device 3 of the third embodiment is generated by the acoustic feature amount generation unit 14. The feature quantity normalization unit 15 for normalizing the generated acoustic feature quantity is provided. As the normalized features, for example, a linearly transformed feature by the feature space maximum likelihood linear regression or a bottleneck feature of a neural network can be used.

The feature amount normalization unit 15 inputs the normalized acoustic feature amount to the fragment data generation unit 11. The fragment data generation unit 11 obtains a position for dividing the utterance data in order to generate the fragment data based on the text of the original utterance data or a non-sounding section, and divides the normalized acoustic feature data at the dividing position. To generate fragment data.

The fragment data generation unit 11 passes the fragment data (consisting of the normalized acoustic features and the corresponding text) to the fragment data combination unit 12. Similar to the storage device 3 of the first embodiment, the fragment data combining unit 12 randomly combines the fragment data to generate new utterance data and stores it in the storage unit 13. The utterance data stored by the storage device 3 of the present embodiment includes a normalized acoustic feature amount as data related to voice.

The storage device 3 of the third embodiment has been described above. The storage device 3 of the third embodiment can store the utterance data in a state where the contents are concealed, as in the first embodiment. Further, by storing the normalized acoustic features as the data related to the voice, it is possible to further make it more difficult to estimate the combined position of the fragment data constituting the utterance data based on the background noise and the like.

(Fourth Embodiment)
FIG. 4 is a diagram showing a configuration of a fourth embodiment obtained by modifying the first embodiment. The basic configuration of the storage device 4 of the fourth embodiment is the same as that of the third embodiment, but in the storage device 4 of the fourth embodiment, the input unit 10 uses the original utterance data. The input of utterance data including the speaker ID that identifies the speaker is accepted. In this embodiment, it is assumed that one speaker is a plurality of utterance data input at one time.

Similar to the storage device 4 of the third embodiment, the storage device 4 of the fourth embodiment generates a plurality of fragment data from the input plurality of utterance data, and randomly generates the generated plurality of fragment data. To generate new utterance data, and store it in the storage unit 13. In the fourth embodiment, the speaker ID is added to the newly generated utterance data. The fragment data combining unit 12 generates new utterance data and stores it in the storage unit 13.

The storage device 4 of the fourth embodiment has been described above. The storage device 4 of the fourth embodiment can store the utterance data in a state where the contents are concealed, as in the first embodiment. Further, since the speaker ID is added to the saved utterance data, when using the saved data, for example, learning is performed using only the utterance data of the same speaker, or conversely, a plurality of different speech data are used. The use of data is expanding, such as learning using the speaker's utterance data.

(Fifth Embodiment)
FIG. 5 is a diagram showing the configuration of the storage device 5 according to the fifth embodiment. In the storage device 5 of the first to fourth embodiments, the device for concealing the utterance content and storing the utterance data has been described, but in the fifth embodiment, in addition to concealing the utterance content, the utterance is spoken. A device that makes it difficult to identify a person and stores utterance data in consideration of anonymity will be described.

The basic configuration of the storage device 5 of the fifth embodiment is the same as that of the storage device 4 of the fourth embodiment. The storage device 5 of the fifth embodiment includes utterance data by a plurality of speakers in a plurality of utterance data input at one time, and a speaker ID that identifies the speaker is attached to each utterance data. ing.

In the fifth embodiment, when the fragment data combining unit 12 generates new utterance data, the fragment data of a plurality of different speakers is always used to generate new utterance data. That is, the fragment data combining unit 12 includes fragment data obtained from a plurality of speakers in each generated utterance data. The fragment data combining unit 12 does not assign a speaker ID to the generated utterance data, unlike the fourth embodiment.

The storage device 5 of the fifth embodiment has been described above. The storage device 5 of the fifth embodiment can store the utterance data in a state where the contents are concealed, as in the first to fourth embodiments. Further, since each of the stored utterance data includes fragment data obtained from a plurality of different speakers, it becomes difficult to identify the speaker from the utterance data. If fragment data of k speakers is included in one utterance, k-anonymization can be achieved in which the candidates for speakers of the utterance data cannot be narrowed down to k or less.

(Sixth Embodiment)
FIG. 6 is a diagram showing a configuration of a storage device 6 according to a sixth embodiment, which is a modification of the second embodiment. The basic configuration of the storage device 6 of the sixth embodiment is the same as that of the first embodiment, but the storage device 6 of the sixth embodiment is the storage device of the fifth embodiment. In addition to the configuration of 5, a speaker feature amount calculation unit 16 and a clustering unit 17 are provided.

The speaker feature amount calculation unit 16 has a function of calculating the feature amount of the speaker from the voice of the utterance data by using the speaker identification technique. In the present embodiment, the speaker's embedded vector (distributed expression) is used as the speaker's feature quantity. For this vector, for example, i-vector, which is a quantity created by factor analysis, or x-vector / d-vector obtained from the output of the middle layer of the speaker identification network can be used. FIG. 7A is a conceptual diagram showing that the speaker vectors v1 to v9 are obtained based on the utterance data of the speakers 1 to 9.

The clustering unit 17 clusters the speakers of the original utterance data based on the embedded vector obtained by the speaker feature amount calculation unit 16, so that speakers having similar acoustic features are included in the same cluster.

FIG. 7B is a diagram showing an example of clustering speakers 1 to 9 based on speaker vectors v1 to v9. The speaker 1, the speaker 4, and the speaker 5 are clustered in the same cluster c1 because the features of the speakers of the vectors v1, v4, and v5 are similar.

The clustering unit 17 passes the cluster ID and the speaker ID of the speaker included in the cluster to the fragment data combining unit 12. The fragment data combining unit 12 uses the fragment data of a plurality of speakers when the fragment data combining unit 12 generates new utterance data, and at this time, the fragment data of a plurality of speakers included in the same cluster is used. Use.

The storage device 6 of the sixth embodiment has been described above. Since the storage device 6 of the sixth embodiment includes fragment data of a plurality of speakers having similar acoustic features in one utterance data, it can be difficult to identify the speaker from the utterance data. .. Moreover, since the acoustic features are normalized, it is possible to make it more difficult to identify the speaker from the utterance data.

In the storage device 6 of the sixth embodiment, the cluster ID and the speaker ID of the speaker included in the cluster are passed from the clustering unit 17 to the feature amount normalization unit 15, and the feature amount normalization unit 15 is the cluster. The data may be used to normalize features among speakers in the same cluster. By using the same cluster as the unit of normalization, the difference between the acoustic features of each speaker and the normalized acoustic features can be reduced.

(7th Embodiment)
FIG. 8 is a diagram showing the configuration of the storage device 7 according to the seventh embodiment. The storage device 7 of the seventh embodiment quantitatively evaluates the degree of concealment of the utterance data stored in the storage unit 13, and generates utterance data based on the evaluation result.

The storage device 7 of the seventh embodiment is generated by the language model generation unit 18 based on the speech data stored in the storage device 7, in addition to the configuration of the storage device 7 of the second embodiment. It includes a language model storage unit 19 that stores a language model, and a language prediction unit 20 that performs language prediction based on the language model.

The language model generation unit 18 generates a language model based on a plurality of utterance data stored in the storage unit 13. As the language model to be generated, for example, an n-gram language model, a language model based on a recurrent neural network, or the like can be used.

The language prediction unit 20 uses the generated language model to perform language prediction on the text of the input utterance data, and obtains the prediction accuracy of the language prediction. As the prediction accuracy, for example, the reciprocal of perplexity can be used. When the reciprocal of the perplexity obtained by the language prediction unit 20 is equal to or greater than a certain threshold value, it can be determined that the input utterance data is not sufficiently concealed. As the method for determining the prediction accuracy by the language prediction unit 20, another method can be adopted. For example, a language model is generated based on the text of the original utterance data, the prediction accuracy of the original utterance data is obtained using the generated language model, and the prediction accuracy is compared with the language model generated from the stored utterance data. You may. When the difference in prediction accuracy between the two language models is less than or equal to the threshold value, it can be determined that the input utterance data is not sufficiently concealed.

If the concealment is not sufficiently performed, the language prediction unit 20 notifies the fragment data generation unit 11 to that effect. Upon receiving this notification, the fragment data generation unit 11 shortens the length of the fragment data to be generated. For example, when it is determined that the concealment is not sufficient when the fragment data consisting of about 3 clauses is originally generated, the fragment data generation unit 11 generates the fragment data of 2 clauses or less. By using shorter fragmented data, the original utterance data is more fragmented, so that the degree of concealment can be increased.

The storage device 7 of the seventh embodiment has been described above. The storage device 7 of the seventh embodiment can quantitatively determine the degree of concealment of the utterance data stored in the storage unit 13. Then, when the degree of concealment is low, it is possible to generate more subdivided fragment data, generate new utterance data, and store the utterance data maintaining the concealment level.

(8th Embodiment)
FIG. 9 is a diagram showing a configuration of a storage device 8 according to an eighth embodiment, which is a modification of the seventh embodiment. The basic configuration of the storage device 8 of the eighth embodiment is the same as that of the seventh embodiment, but the storage device 8 of the eighth embodiment is based on the input original utterance data. The difference is that a language model is generated and the generated language model is applied to the utterance data stored in the storage unit 13 to perform language prediction. With this configuration as well, the degree of concealment of the utterance data stored in the storage unit 13 can be evaluated.

Although the storage device of the present embodiment has been described above, an example of the hardware of the storage device described above is a computer provided with a CPU, RAM, ROM, hard disk, display, keyboard, mouse, communication interface, and the like. The above-mentioned storage device is realized by storing a program having a module that realizes each of the above-mentioned functions in a RAM or ROM and executing the program by a CPU. Such programs are also included in the scope of the present invention.

The present invention is useful as a technique for storing utterance data used for learning an acoustic model.

1-8 Preservation device 10 Input unit 11 Fragment data generation unit 12 Fragment data combination unit 13 Preservation unit 14 Acoustic feature amount generation unit 15 Feature amount normalization unit 16 Speaker feature amount calculation unit 17 Clustering unit 18 Language model generation unit 19 languages Model storage 20 Language prediction unit

Claims (15)

An utterance data consisting of voice data and text corresponding to the voice, and a storage device for storing utterance data used for learning an acoustic model.
An input unit for inputting multiple original utterance data consisting of a voice and text corresponding to the voice,
A fragment data generation unit that divides the original utterance data and generates a plurality of fragment data consisting of a voice containing one or more phrases and a text corresponding to the voice.
A fragment data combining unit that randomly combines a plurality of fragment data generated from the plurality of original utterance data to generate a plurality of utterance data having a predetermined length.
A storage unit that stores a plurality of utterance data generated by the fragment data combination unit, and a storage unit.
A storage device equipped with. It is provided with an acoustic feature amount generation unit that generates an acoustic feature amount from the voice input from the input unit.
The storage device according to claim 1, wherein the storage unit stores the acoustic feature amount of the utterance data as data related to the voice. It is provided with a feature amount normalization unit that normalizes the acoustic feature amount generated by the acoustic feature amount generation unit.
The storage device according to claim 2, wherein the storage unit stores a normalized acoustic feature amount of the utterance data as data related to the voice. The input unit accepts the input of the speaker identifier of the utterance data together with the plurality of original utterance data.
The storage device according to any one of claims 1 to 3, wherein the storage unit stores the utterance data generated by the fragment data combination unit in association with the identifier of the speaker. The input unit accepts the input of the speaker identifier of the utterance data together with the plurality of original utterance data.
The storage device according to claim 1, wherein the fragment data combining unit includes fragment data obtained from a plurality of speakers in each utterance data generated based on the identifier of the speaker. A speaker feature calculation unit that obtains speaker features by applying speaker identification technology to a plurality of the original utterance data, and a speaker feature calculation unit.
A clustering unit that clusters the speaker based on the feature amount of the speaker,
With
The storage device according to claim 5, wherein the fragment data combining unit includes fragment data obtained from utterance data of a plurality of speakers included in the same cluster in each utterance data generated. An acoustic feature amount generation unit that generates an acoustic feature amount from the voice,
A feature normalization unit that normalizes the acoustic features of the utterance data of multiple speakers included in the same cluster,
The storage device according to claim 6. From claim 1, the fragment data generation unit performs morphological analysis and syntactic analysis on the input text to obtain a clause, and generates fragment data including one or more clauses based on the break position of the clause. 7. The storage device according to any one of 7. The fragment data generation unit detects a non-pronounced section in the input voice, and uses the non-pronounced section as a phrase delimiter position to generate fragment data including one or more phrases based on the phrase delimiter position. Item 2. The storage device according to any one of Items 1 to 8. The storage according to any one of claims 1 to 9, wherein the fragment data combining unit sets the predetermined length of the utterance data to be generated based on the distribution of the lengths of the plurality of original utterance data. apparatus. A language model generation unit that generates a language model based on the text stored in the storage unit,
A language prediction unit that applies the language model to the text of the original utterance data to perform language prediction, and
The storage device according to any one of claims 1 to 10. A language model generator that generates a language model based on the text of the original utterance data,
A language prediction unit that applies the language model to the text stored in the storage unit to perform language prediction,
The storage device according to any one of claims 1 to 10. When the prediction accuracy by the language prediction unit is higher than a predetermined threshold value,
The storage device according to claim 11 or 12, wherein the fragment data generation unit shortens the length of the fragment data. It is utterance data consisting of data related to voice and text corresponding to the voice, and is a method of storing utterance data used for learning an acoustic model in a storage device.
Steps to input multiple original utterance data consisting of voice and text corresponding to the voice,
A step of dividing the original utterance data to generate a plurality of fragmentary data consisting of a voice containing one or more phrases and a text corresponding to the voice.
A step of randomly combining a plurality of fragment data generated from the plurality of original utterance data to generate a plurality of utterance data having a predetermined length, and
A step of storing a plurality of generated utterance data in the storage device, and
Preservation method with. It is a utterance data consisting of data related to voice and text corresponding to the voice, and is a program for storing utterance data used for learning an acoustic model in a storage device, and is stored in a computer.
Steps to input multiple original utterance data consisting of voice and text corresponding to the voice,
A step of dividing the original utterance data to generate a plurality of fragmentary data consisting of a voice containing one or more phrases and a text corresponding to the voice.
A step of randomly combining a plurality of fragment data generated from the plurality of original utterance data to generate a plurality of utterance data having a predetermined length, and
A step of storing a plurality of generated utterance data in the storage device, and
A program that executes.

Images