JP2017223848A - Speaker recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To recognize a speaker with higher accuracy even when there is disturbance such as noise or an utterance length is short.SOLUTION: With respect to a speech inputted from a speech input unit 11, an analysis unit 13 extracts a feature quantity called i-vector using a large-scale speech database 12 having the speeches or speech models of a large number of unspecified speakers or registered speakers. A similarity calculation unit 14 calculates the similarity of speech models of a large number of unspecified speakers and speech models of registered speakers in the large-scale speech database 12 with the feature quantity of the inputted speech. An order calculation unit 15 calculates the order of magnitude in all speaker models of the similarity of the feature quantity of the inputted speech with the registered speaker model insisted by an input speaker. A determination unit 16 determines that the speaker is the person in question when the order calculated in the order calculation process is within a previously determined order.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a speaker recognition device for recognizing a speaker based on an acquired voice signal.

  In conventional speaker recognition, generally, voices for registration are collected in advance, feature quantities are extracted by analyzing the collected voices, and newly acquired voices of unknown speakers and registered speakers are acquired. There is a method for determining whether or not the person is the person based on the similarity of the feature quantity. In addition, there is a technique for speaker recognition by ranking similarities with the speech of multiple speakers.

  The speaker recognition device described in Patent Document 1 analyzes the voice of an input speaker, extracts feature amounts, obtains the similarity with all registered speakers using a tree structure, The degree of similarity with all the registered speakers is ranked, and if the degree of similarity with the person claimed by the input speaker is within a predetermined order, the person is determined to be the person.

  Also, in Patent Document 1, a conventional speaker recognition method and speaker identification device based only on a general similarity are described, and the speaker recognition method based on the ranking is more various than the speaker recognition method based only on a similarity. He says he is robust against disturbances.

  Non-Patent Document 1 newly proposes a speaker-specific feature amount called i-vector and a method for obtaining the feature amount as a highly accurate feature amount for speaker recognition.

Japanese Patent No. 2991288

Dehak, Najim, et al. "Front-end factor analysis for speaker verification." Audio, Speech, and Language Processing, IEEE Transactions on 19.4 (2011): 788-798.

  In conventional speaker recognition, there are cases where disturbances such as noise are added to the speech, and cases where the target speech is extremely short as a problem of reduced accuracy.

  In Patent Document 1, robustness against disturbance is shown by ranking the similarity with each speaker, but a specific method for the feature amount is not clearly shown, and a detailed method for determining a threshold is not provided. Since it is undecided, the grounds and the more reliable method that the method based on the ranking is more accurate than the method based only on the similarity when a disturbance occurs are not shown. In addition, although robustness against disturbances has been discussed with respect to factors that reduce accuracy, utterances with a short word level are not mentioned.

  An object of the present invention is to provide a speaker recognition method and a speaker recognition device that extract a feature quantity named i-vector from a speech as a speaker-specific model and that has high accuracy based on experimental results and the like. To do.

  In the speaker recognition method according to one aspect of the present invention, the speech of an unknown speaker is input using a large-scale speech database having speech and speech models of a large number of unspecified majority speakers or registered speakers in advance. Speech input processing, analysis processing for extracting feature quantities called i-vectors for the input speech using the large-scale database, speech models of unspecified majority speakers in the large-scale speech database, and registered speech The similarity calculation process for calculating the similarity between the speaker voice model and the feature amount of the input speech, and the similarity between the feature amount of the input speech and the registered speaker model claimed by the input speaker are all A rank calculation process for calculating the highest rank in the speaker model, and a person who claims that the rank is determined by the rank calculation process is within a predetermined rank. Provided by that determination process.

  With this configuration, the voice signal of an unknown speaker is acquired. An i-vector, which is a unique feature amount, is extracted from the acquired audio signal. The method of extracting i-vector as a speaker-specific feature amount needs to use a general speech feature amount distribution obtained from the speech of many speakers, and may use information in a large-scale speech database. it can. Since i-vector is expressed as a numerical string of several hundreds, the similarity can be easily calculated. In addition, it is robust against disturbances and is less affected by noise and device differences during input. The degree of similarity between the extracted feature quantity of the unknown speaker and the speech model of a large-scale unspecified majority speaker registered in the large-scale speech database or the model of the registered speaker is calculated. The calculated similarities are ranked in descending order, and if the similarity with the registered speaker claimed by the unknown speaker is within a predetermined rank, the person is determined to be the person concerned. For speaker voices in a large-scale voice database, it is possible to select in advance the collected voice conditions such as presence / absence of noise, utterance length, and utterance contents. For the speaker voices to be compared, stable speech ranking can be achieved even if disturbances to unknown speaker voices or utterances with extremely short word levels occur due to the use of voices with low noise and sufficient utterance length. Can be expected.

  Therefore, more accurate speaker recognition can be performed by ranking unspecified number of speakers and registered speakers based on similarities with features robust to disturbances using a large-scale speech database.

  In the above speaker recognition method, the same determination is made in advance using a development speaker as a registered speaker in advance, and the probability that the person is rejected and the probability that the impersonator is accepted are set for each threshold. The order in which the error rate is calculated may be determined as the order for determining that the person is the person.

  With this configuration, the probability of rejecting the person for each predetermined rank in the development speaker and the two types of recognition error rates for accepting the impersonator are calculated. The order in which the speaker recognition accuracy is highest among the development speakers is determined as the threshold value.

  Therefore, since the rank determined by the development speaker voice can be used in the determination of the unknown speaker, the speaker can be recognized with higher accuracy.

  According to the present invention, a speaker can be recognized with higher accuracy even in a bad environment in speaker recognition such as disturbance such as noise or extremely short utterance.

It is a figure which shows the structure of the speaker identification device in Embodiment 1 of this invention. It is a figure which shows the structure of the speaker identification device in Embodiment 2 of this invention. It is a figure which shows the graph for threshold value ranking determination in Embodiment 3 of this invention.

  Embodiments of the present invention will be described below with reference to the accompanying drawings. The following embodiments are examples embodying the present invention, and do not limit the technical scope of the present invention.

(Embodiment 1)
FIG. 1 is a diagram illustrating the configuration of the speaker recognition apparatus according to the first embodiment. The speaker recognition device is built in, for example, a television, a smart phone, or a car navigation device.

  The speaker recognition apparatus shown in FIG. 1 includes a voice input unit 11, a large-scale voice database 12, an analysis unit 13, a similarity calculation unit 14, a rank calculation unit 15, and a determination unit 16.

  The voice input unit 11 is composed of, for example, a microphone, collects unknown speaker voice, converts the collected voice into a voice signal, and outputs the voice signal.

  The large-scale voice database 12 is a storage device arranged on the cloud, for example, and holds voice data or a voice model. In the large-scale speech database, the unspecified majority speaker speech database that holds the speech or speech model of an unspecified number of speakers that do not include registered speakers, and the input unknown speakers are determined. Although a registered speaker voice database that holds a registered speaker's voice or a voice model is included, the present invention is not limited to this configuration, and a number of speaker's voices may be held.

  The analysis unit 13 analyzes the voice signal input from the voice input unit 11 and calculates the feature amount of the voice uttered by the unknown speaker. Here, a feature value w obtained by an equation M = m + Tw called i-vector is calculated as a speaker-specific feature value. M in this equation is a feature amount indicating an individual speaker to be input. For example, MFCC (Mel Frequency Cepstial Coefficient) is a method of expressing a numerical sequence obtained by analyzing a frequency spectrum of speech by overlapping normal distributions. Some GMM (Gaussian Mixture Model), GMM super vectors, and the like are used. For m, a feature value obtained in the same manner as M from a large number of speaker voices is used. The GMM in m is called UBM (Universal Background Model). T is a basis vector that can cover the feature amount space of a general speaker obtained by M. w is a feature value used in the present invention. Detailed extraction methods and the like of each are described in Non-Patent Document 1 and related documents, and thus will be omitted. The speaker voice used to generate the UBM is considered to have higher accuracy as the environment, speaker characteristics, utterance content, and the like are more diverse and large in order to indicate general feature amounts of the voice data. Therefore, the analysis unit 13 extracts feature amounts using a large number of speaker voices in the large-scale voice database 12.

  The similarity calculation unit 14 compares the feature amount w of the unknown speaker calculated by the analysis unit 13 with all or some of the speech models on the large-scale speech database 12, and calculates the similarity. Since the feature quantity and the speech model are a numerical sequence of about several hundreds, the similarity can be easily calculated by Cosse distance scoring shown in Non-Patent Document 1, for example. The case distance scoring is close to 1 when the degree of similarity is high, and close to -1 when the degree of similarity is low. Further, the similarity calculation method is not limited to the above.

  The rank calculation unit 15 ranks the similarities obtained by the similarity calculation unit 14 in descending order, and claims that the input features of the unknown speaker and the unknown speaker on the large-scale speech database are the person himself / herself. The number of similarities with the registered speaker model to be calculated is calculated from among all the similarities calculated by the similarity calculating unit 14.

  The determination unit 16 determines whether or not the unknown speaker is a registered speaker based on the rank calculated by the rank calculation unit 15. If the calculated rank is higher than a predetermined rank, it is determined that the speaker is a registered speaker to be claimed.

(Embodiment 2)
FIG. 2 is a diagram illustrating a configuration of the speaker recognition device according to the second embodiment. 2, the same components as those in FIG. 1 are denoted by the same reference numerals, and description thereof is omitted.

  In the speaker recognition apparatus according to the second embodiment, the process at the time of threshold determination is performed before the process at the time of speaker recognition. The process in the first embodiment is a process at the time of speaker recognition.

  In the threshold value determination process according to the second embodiment, the voice input unit 11 receives a development speaker voice, converts the input voice into a voice signal, and outputs the voice signal. The speaker voice for development has a known voice speaker, and may be a speaker different from the registered speaker, or may overlap with the registered speaker.

  The analysis unit 13, the similarity calculation unit 14, and the rank calculation unit 15 perform the processing described in the first embodiment using the development speaker voice and the model on the large-scale voice database 12, and calculate the rank.

  The threshold value determination unit 17 calculates the recognition error rate of the development speaker voice from the rank calculated by the rank calculation unit 15, thereby determining the rank that is an appropriate threshold value. For example, the recognition error rate of the speaker voice for development is calculated when the threshold ranking is 100th. The error rate includes a probability that the voice that should be the person is determined as the other person (person rejection rate) and a probability that the voice that is supposed to be the impersonator is determined as the person (other person acceptance rate). One person A is selected from the speaker voices for development, and the rest is regarded as an impersonator. The utterance of A is input, and the degree of similarity between the model of speaker A and the speaker whose degree of similarity is calculated is calculated as in the first embodiment. . In addition, the utterances of other scammers other than A are input, the ranking based on the similarity is calculated in the same manner, and if the similarity with the model of speaker A is within 100th, the scammer is determined to be A. It will be accepted by others. As described above, the two types of error rates are calculated in increments of 10 to 100, for example, 100 and above are calculated in increments of 100, and an appropriate threshold at which the error rate is lowest when the two types of recognition error rates intersect. Is determined as a ranking. FIG. 3 is a graph in which two types of error rates are calculated by the above method when a short utterance of 10 women is input as a development speaker. The vertical axis indicates the error rate, the horizontal axis indicates the threshold, the broken line indicates the principal rejection rate, and the solid line indicates the other person acceptance rate. In FIG. 3, about 200th place is determined as an appropriate threshold value. The determined rank order is used for determination by the determination unit 16 in the process at the time of speaker recognition.

  In the process at the time of speaker recognition in the second embodiment, the determination unit 16 determines whether the unknown speaker input by using the threshold ranking calculated by the threshold determination unit 17 is a speaker claimed.

  The speaker recognition method and the speaker recognition apparatus according to the present invention can identify a speaker with higher accuracy even when there is a disturbance such as noise or a short utterance length by using voice data on a large-scale database. Therefore, the present invention is useful as a speaker recognition method and a speaker recognition device for recognizing a speaker based on an acquired voice signal.

DESCRIPTION OF SYMBOLS 11 Audio | voice input part 12 Large-scale audio | voice database 13 Analysis part 14 Similarity calculation part 15 Ranking calculation part 16 Judgment part 17 Threshold value determination part

Claims (2)

A speaker recognition device that performs speaker recognition using a large-scale speech database having speech or speech models of unspecified majority speakers or registered speakers,
An audio input unit for inputting audio;
An analysis unit that extracts a feature amount called i-vector using the large-scale speech database for the input speech;
A similarity calculator that calculates the similarity between the speech model of the unspecified majority speakers and the registered speaker speech model of the large-scale speech database and the feature amount of the input speech;
A rank calculation unit for calculating the highest degree of similarity between the input speech feature quantity and the registered speaker model claimed by the input speaker in the entire speaker model;
A speaker recognition device comprising: a determination unit that determines that the user is an individual when the rank obtained by the rank calculation processing is within a predetermined rank.   Make a similar determination in advance using a development speaker as a registered speaker in advance, calculate the probability of rejecting the person and the probability of accepting the impersonator for each threshold order, and rank with the lowest error rate The speaker recognition device according to claim 1, further comprising: a threshold value determination unit that determines as a ranking for determining that the person is the person.