JP7216348B2 - Speech processing device, speech processing method, and speech processing program - Google Patents

Description

The present invention relates to an audio processing device, an audio processing method, and an audio processing program.

2. Description of the Related Art A speech processing apparatus is known that calculates speaker characteristics representing individuality for identifying a speaker who has emitted a speech from a speech signal. Also known is a speaker recognition apparatus that estimates a speaker who has emitted a speech signal by using this speaker feature.

A speaker recognition apparatus using this type of speech processing apparatus uses first speaker features extracted from a first speech signal and second speech features extracted from a second speech signal to identify a speaker. Evaluate the degree of similarity with person characteristics. Then, the speaker recognition device determines whether or not the speaker of the two speech signals is the same based on the similarity evaluation result.

Non-Patent Document 1 describes a technique for extracting speaker features from a speech signal. The speaker feature extraction technique described in Non-Patent Document 1 calculates speech statistics of a speech signal using a speech model. Then, the speaker feature extraction technique described in Non-Patent Document 1 processes the speech statistic based on the factor analysis technique and calculates a speaker feature vector represented by a predetermined number of elements. That is, in Non-Patent Document 1, a speaker feature vector is used as a speaker feature representing the speaker's individuality.

Najim Dehak, Patrick Kenny, Reda Dehak, Pierre Dumouchel, and Pierre Ouellet, “Front-End Factor Analysis for Speaker Verification,” IEEE Transaction on Audio, Speech and Language Processing, Vol. 19, No. 4, pp. 788-798 , 2011.

However, the technique described in Non-Patent Document 1 has a problem that the accuracy of speaker recognition using extracted speaker features is not sufficient.

The technique described in Non-Patent Document 1 performs predetermined statistical processing on a speech signal input to a speaker feature extraction device to calculate a speaker feature vector. Specifically, the technology described in Non-Patent Document 1 calculates a speaker feature vector by uniformly performing statistical processing on the entire speech signal input to the speaker feature extraction device. Therefore, the technology described in Non-Patent Document 1 can be used to extract speech from the entire speech signal even if a partial section of the speech signal includes a signal that is not suitable as a basis for calculating the individuality of the speaker. Since the speaker feature vector is calculated, the accuracy of speaker recognition may be impaired. Specifically, when a partial segment of a speech signal contains, for example, unclear utterances of the speaker, sounds different from the speaking voice such as coughing or laughter of the speaker, noise, etc., it is difficult to recognize the speaker. Accuracy may be lost.

SUMMARY OF THE INVENTION The present invention has been made in view of the above problems, and an object of the present invention is to provide a speech processing apparatus, a speech processing method, and a speech processing program that further improve the accuracy of speaker recognition.

A speech processing apparatus according to a first aspect of the present invention comprises reception means for accepting input of a plurality of speech signals representing speech ; quality estimating means for calculating two types of quality of the generated voice; and information processing for calculating a recognition feature amount for recognizing specific attribute information from the plurality of voice signals based on the quality of the plurality of voice signals. and speech statistic calculation means for calculating a speech statistic representing the frequency of occurrence of types of sounds contained in the plurality of speech signals, wherein the information processing means calculates the speech statistics of the plurality of speech signals. and the quality of the plurality of speech signals, the recognition feature amount is calculated.

A speech processing method according to a second aspect of the present invention accepts input of a plurality of speech signals representing speech, and selects speech that is correct for speaker recognition and speech that causes an error in speaker recognition in the plurality of speech signals. calculating two types of quality, calculating a recognition feature amount for recognizing specific attribute information from the plurality of audio signals based on the quality of the plurality of audio signals, and calculating the recognition feature amount for recognizing specific attribute information from the plurality of audio signals; A speech statistic representing the degree of appearance of the type of sound is calculated, and the recognition feature quantity is calculated based on the speech statistic of the plurality of speech signals and the quality of the plurality of speech signals.

A speech processing program according to a third aspect of the present invention provides a computer with a process of receiving input of a plurality of speech signals representing speech, a speech correct for speaker recognition in the plurality of speech signals, and processing of the speaker recognition. a process of calculating two types of quality of speech that causes an error; and a process of calculating a recognition feature amount for recognizing specific attribute information from the plurality of speech signals based on the qualities of the plurality of speech signals. , based on a process of calculating a speech statistic representing the appearance of types of sounds included in the plurality of speech signals, and the speech statistics of the plurality of speech signals and the quality of the plurality of speech signals; , and a process of calculating the recognition feature amount.

According to the present invention, it is possible to provide a speech processing device, a speech processing method, and a program that improve the accuracy of speaker recognition.

1 is a block diagram showing the configuration of a speech processing device according to a first embodiment of the present invention; FIG. 4 is a flow chart showing the operation flow of the speech processing device according to the first embodiment of the present invention; FIG. 4 is a block diagram showing the configuration of a speech processing device according to a second embodiment of the present invention; FIG. 9 is a flow chart showing the operation flow of the speech processing device according to the second embodiment of the present invention; FIG. 11 is a block diagram showing the configuration of a speech processing device according to a third embodiment of the present invention; FIG. FIG. 11 is a block diagram showing the configuration of a speech processing device according to another embodiment of the present invention;

Hereinafter, embodiments of a speech processing device and the like and a speaker feature extraction device will be described with reference to the drawings. It should be noted that, since components denoted by the same reference numerals in the embodiments perform the same operations, re-explanation may be omitted.

<First embodiment>
FIG. 1 is a block diagram showing the configuration of a speech processing device according to the first embodiment of the present invention.

The speech processing device 100 includes a contribution estimation unit 11 and a speaker feature calculation unit 12 .

The contribution estimation unit 11 receives an audio signal representing audio from the outside. Further, based on the received audio signal, the contribution estimating unit 11 calculates a contribution that numerically expresses the degree of quality of the partial section of the audio signal.

The speaker feature calculator 12 uses the contribution of the partial section of the speech signal calculated by the contribution estimator 11 as the weight of the partial section to calculate the recognition feature amount for recognizing specific attribute information from the speech signal. Calculate

Here, the specific attribute information is information indicating the speaker who issued the voice signal, the language composing the voice signal, the emotional expression included in the voice signal, the personality of the speaker estimated from the voice signal, and the like. .

The operation flow of the speech processing device 100 will be described with reference to FIG. FIG. 2 is a flow chart showing the operation flow of the speech processing device according to the first embodiment of the present invention.

First, the contribution estimating unit 11 calculates the contribution of the partial section of the audio signal based on the externally received audio signal (step S101). Next, the contribution estimation unit 11 outputs the calculated contribution of the partial section of the speech signal to the speaker feature calculation unit 12 .

Next, the speaker feature calculation unit 12 calculates a recognition feature quantity based on the degree of contribution received from the contribution degree estimation unit 11 (step S102).

<Second embodiment>
FIG. 3 is a block diagram of a speech processing device 200 according to the second embodiment. The speech processing device 200 includes a contribution estimation unit 11 , a speaker feature calculation unit 12 , a speech section detection unit 21 and a speech statistics calculation unit 22 . Moreover, the speech processing device 200 may further include a contribution storage unit 23 and a contribution learning unit 24 .

The voice segment detection unit 21 receives voice signals from the outside. Further, the voice segment detection unit 21 detects and segments voice segments included in the received voice signal. At this time, the voice segment detection unit 21 may segment the voice signal into segments of a fixed length or segments of different lengths. For example, the voice segment detection unit 21 may determine that a segment of the voice signal in which the volume continues for a certain period of time and is smaller than a predetermined value is silent, and the segments before and after the segment may be determined as different voice segments and segmented. Then, the speech interval detection unit 21 outputs the segmented speech signal, which is the segmentation result (the processing result of the speech interval detection unit 21 ), to the contribution estimation unit 11 and the speech statistic calculation unit 22 . Here, receiving an audio signal means, for example, receiving an audio signal from an external device or another processing device, or passing a processing result of audio signal processing from another program. Also, the output means, for example, transmission to an external device or other processing device, or delivery of the processing result of the speech segment detection unit 21 to another program.

The speech statistic calculator 22 receives the segmented speech signal from the speech section detector 21 . Based on the received segmented audio signal, the audio statistic calculator 22 calculates audio statistics representing the types of sounds included in the segmented audio signal. Here, the type of sound is, for example, a group of sounds obtained by clustering phonemes, words, and speech signals determined by a language based on the degree of similarity. Then, the speech statistic calculator 22 outputs the speech statistic to the speaker feature calculator 12 . A speech statistic calculated for a certain speech signal is hereinafter referred to as a speech statistic of the speech signal.

An example of a method of calculating the speech statistic by the speech statistic calculator 22 will be described. Specifically, based on the segmented speech signal received from the speech segment detection unit 21, the speech statistic calculation unit 22 calculates acoustic features represented by the result of frequency analysis processing of the segmented speech signal. , and output the calculated result. For example, the speech statistic calculator 22 converts the segmented speech signal received from the speech section detector 21 into a short-time frame time series. Then, the speech statistic calculation unit 22 frequency-analyzes each frame in the short-time frame time series, and outputs the processing result as an acoustic feature. In this case, the speech statistic calculation unit 22 generates, for example, frames of 25 millisecond intervals every 10 milliseconds as a short-time frame time series. The speech statistic calculation unit 22, for example, as acoustic features that are frequency analysis results, frequency filter bank features obtained by Fast Fourier Transform (FFT) and filter bank processing, and in addition discrete cosine transform Mel-Frequency Cepstrum Coefficients (MFCC) features obtained by the processing are calculated.

Then, the speech statistic calculation unit 22 calculates the time series of numerical information representing the types of sounds using the time series of the acoustic features and the speech model that stores the correspondence relationship between the acoustic features and the types of sounds. For example, when the speech model is a Gaussian mixture model (GMM), the speech statistic calculation unit 22 calculates each element distribution based on the mean, variance, and mixture coefficient of each element distribution of the Gaussian mixture model. Calculate the posterior probability of Here, the posterior probability of each element distribution is the appearance of each type of sound included in the speech signal. Further, for example, when the speech model is a neural network, the speech statistic calculation unit 22 calculates the degree of appearance of the types of sounds contained in the speech signal based on the acoustic features and the weighting coefficients of the neural network. Calculate

The contribution storage unit 23 stores one or more contribution estimators. The contribution estimator is configured to operate to classify the speech signal into a plurality of classes according to signal quality. The contribution estimator outputs, for example, numerical information representing the quality of the speech signal. The types of signal quality are, for example, speech/non-speech/silence. Further, the types of signal quality are, for example, speech that is correct in speaker recognition and speech that causes an error in speaker recognition.

Specifically, the contribution storage unit 23 stores parameters held by the contribution estimator. For example, if the contribution estimator is a neural network, the contribution storage unit 23 stores a set of parameters, such as the number of nodes that make up the network and weighting coefficients for connections between nodes.

In addition, in FIG. 3, an example in which the degree-of-contribution storage unit 23 is built in the speech processing device 200 has been described, but the present invention is not limited to this. Contribution degree storage unit 23 may be realized by a storage device provided outside speech processing device 200 .

The contribution estimation unit 11 receives the segmented speech signal from the speech section detection unit 21 . The contribution estimation unit 11 uses the contribution estimator stored in the contribution storage unit 23 to calculate numerical information representing the quality of the segmented speech signal. Like the speech statistic calculation unit 22, the contribution estimation unit 11 converts the segmented speech signal into a short-time frame time series, calculates the acoustic features of each frame, and calculates the time series of the acoustic features. Subsequently, the contribution estimator 11 calculates a numerical value representing the quality of each frame using the acoustic features of each frame and the parameters of the contribution estimator. Hereinafter, a numerical value representing the signal quality calculated for a certain audio signal is referred to as the contribution of the audio signal.

Specifically, for example, if the contribution estimator is a neural network, the contribution estimating unit 11 calculates the contribution of the acoustic feature based on the acoustic feature and the weighting factor of the neural network. For example, the contribution estimator is a neural network whose output layer corresponds to two signal quality types: "correct speaker recognition signals" and "false speaker recognition signals". and At this time, the contribution estimator calculates the probability that the acoustic feature is a signal that corrects speaker recognition and the probability that the acoustic feature is a signal that causes an error in speaker recognition. It outputs the probability that the signal is the correct signal for person recognition. Further, the contribution estimating unit 11 may determine whether a partial section of the speech signal is speech or not and calculate the probability that it is speech before executing speaker recognition.

The speaker feature calculation unit 12 receives the speech statistics output by the speech statistics calculation unit 22 and the contributions output by the contribution estimation unit 11 . The speaker feature calculator 12 uses the speech statistic and the degree of contribution to calculate a recognition feature for recognizing specific attribute information from the speech signal.

An example of a method for calculating the feature vector F(x) based on the i-vector as the recognition feature quantity of the speech signal x by the speaker feature calculator 12 will be described. Note that the feature vector F(x) calculated by the speaker feature calculation unit 12 may be any vector that can be calculated by performing a predetermined operation on the speech signal x, and i-vector is one example.

The speaker feature calculator 12 receives from the speech statistic calculator 22 the acoustic posterior probability Pt(x) and the acoustic feature At(x) calculated for each short-time frame as information on the statistic of the speech signal x. (t={1 . . . T}, where T is a natural number equal to or greater than 1). Further, the speaker feature calculator 12 receives, for example, the contribution Ct(x) calculated for each short-time frame from the contribution estimator 11 as information about the contribution of the speech signal x. The speaker feature calculation unit 12 multiplies each element of the acoustic posterior probability Pt(x) by the degree of contribution Ct(x) as shown in the following equation (1), and obtains the result as Qt(x). calculate.

Using acoustic posterior probability Qt(x) weighted by the degree of contribution and acoustic feature At(x), speaker feature calculation unit 12 calculates the zero-order statistic of speech signal x based on the following equation (2): S0(x) is calculated, and the primary statistic S1(x) is calculated based on the equation (3).

The speaker feature calculator 12 then calculates F(x), which is the i-vector of the audio signal x, based on the following equation (4).

In equations (1) to (4), C is the number of elements of statistics S0(x) and S1(x), D is the number of elements (number of dimensions) of acoustic feature At(x), and mc is the number of Mean vector of acoustic features of the c-th region, I is a unit matrix, and 0 is a zero matrix. T is the parameter for the i-vector computation and Σ is the covariance matrix of the acoustic features in the acoustic feature space.

When the speaker feature calculator 12 calculates the feature vector F(x) in the above-described procedure, at all times t (t={1 . . . T}, where T is a natural number equal to or greater than 1) of the speech signal x, the If the contribution Ct(x) is 1, it is equivalent to the i-vector calculation procedure described in Non-Patent Document 1. In this embodiment, the speaker feature calculator 12 uses the contribution Ct(x) estimated by the contribution estimator 11 according to the time t of the speech signal x to obtain i- A feature vector F(x) different from vector can be calculated.

In this way, in the speech processing device 200, the speaker feature calculator 12 calculates the feature vector F( By calculating x), a feature vector corresponding to the quality of the speech signal can be output.

The contribution learning unit 24 learns a contribution estimator that can be stored in the contribution storage unit 23 using the training speech signal. For example, when the contribution estimator is a neural network, the contribution learning unit 24 optimizes parameters such as connection weight coefficients between nodes that constitute it, according to general optimization criteria. The training speech signal used by the contribution learning unit 24 is a collection of a plurality of speech signals, and each speech signal is associated with one of the quality types of signals output by the contribution estimation unit 11. It is a thing.

Below, a contribution estimator whose input is acoustic features and whose output is the quality of two types of signals, ``speech that correctly recognizes a speaker'' and ``speech that causes an error in speaker recognition'', is used as a contribution learning unit. An example of how T.24 learns is described.

(a) First, the contribution learning unit 24 uses a plurality of speech signals with speaker labels to learn a classifier capable of identifying speaker labels of speech signals. (b) Next, the contribution learning unit 24 converts each of the plurality of speaker-labeled speech signals into a time series of acoustic features calculated for each short-time frame, and uses the discriminator learned in (a) as is used to identify the speaker label for each frame. (c) Next, the contribution level learning unit 24 selects frames having the same speaker label assigned in advance and the speaker label identified by the classifier among the speaker labels of each identified frame as " Frames that are not correct for speaker recognition are defined as "speech that causes errors in speaker recognition". (d) Then, the contribution learning unit 24 learns the contribution estimator using "speech that corrects for speaker recognition" and "speech that causes an error in speaker recognition" as training speech signals.

As described above, in the speech processing device 200 according to the present embodiment, the contribution estimating unit 11 can calculate the contribution of the speech signal as an index representing the quality corresponding to the partial section of the speech signal. The speaker feature calculation unit 12 also calculates feature vectors based on the acoustic statistics and contribution of the speech signal. This makes it possible to output a feature vector for the speech signal, taking into consideration the quality of each partial section of the speech signal. That is, the speech processing apparatus 200 according to the present embodiment can calculate speaker features suitable for increasing the accuracy of speaker recognition.

Note that the contribution storage unit 23 in the speech processing device 200 according to the present embodiment is preferably a non-volatile recording medium, but can also be realized with a volatile recording medium. Also, the process by which the contribution estimator is stored in the contribution storage unit 23 is not particularly limited. For example, the contribution estimator may be stored in the contribution storage unit 23 via a recording medium, or the contribution estimator transmitted via a communication line or the like may be stored in the contribution storage unit 23. . Alternatively, the contribution estimator input via the input device may be stored in the contribution storage unit 23 .

(Operation of Second Embodiment)
Next, the operation of the speech processing device 200 according to the second embodiment will be described using the flowchart of FIG. FIG. 4 is a flow chart showing an example of the operation of the speech processing device 200. As shown in FIG.

The speech processing device 200 receives one or more speech signals from the outside and provides them to the speech segment detection unit 21 . Specifically, the speech segment detection unit 21 segments the received speech signal and outputs the segmented speech signal to the contribution estimation unit 11 and the speech statistics calculation unit 22 (step S201).

The speech statistic calculator 22 performs short-time frame analysis processing on each of the received one or more segmented speech signals, and calculates time series of acoustic features and speech statistics (step S202).

The contribution estimating unit 11 performs short-time analysis frame processing on each of the received one or more segmented speech signals, and calculates a contribution time series (step S203).

The speaker feature calculator 12 calculates and outputs a speaker recognition feature based on one or more received acoustic features, speech statistics, and time series of contribution. (Step S204). After receiving the audio signal from the outside, the audio processing device 200 ends the series of processes.

(Effect of Second Embodiment)
As described above, according to the speech processing device 200 of the present embodiment, it is possible to improve the accuracy of speaker recognition using the speaker features calculated by the speech processing device 200 . This is because, in the speech processing apparatus 200, the contribution estimating unit 11 calculates the quality of the speech signal as the contribution, and the speaker feature calculating unit 12 calculates the feature vector considering the contribution. This is because feature vectors are output with weights placed on subintervals with high .

In this way, the speech processing apparatus 200 according to the present embodiment calculates a feature vector for the speech signal, taking into account the degree of contribution corresponding to the quality of each partial section. As a result, even if the partial section of the speech signal contains unclear utterances of the speaker, sounds different from the speaking voice such as coughing or laughter of the speaker, noise, etc., recognition suitable for speaker recognition is possible. A feature amount can be obtained.

<Third Embodiment>
FIG. 5 is a block diagram showing an example configuration of a speech processing device according to the third embodiment of the present invention.

As shown in FIG. 5 , the speech processing device 300 includes a contribution estimation unit 11 , a speaker feature calculation unit 12 and an attribute recognition unit 13 . The voice processing device 300 is a voice processing device capable of recognizing attribute information.

The contribution estimating unit 11 and the speaker feature calculating unit 12 are the same as those in the first and second embodiments, so descriptions thereof will be omitted.

The attribute recognition unit 13 receives recognition feature amounts for recognizing attribute information from the speaker feature calculation unit 12 . The attribute recognition unit 13 recognizes, based on the recognition feature amount, the speaker who issued the voice signal, the language constituting the voice signal, the emotional expression included in the voice signal, the character of the speaker estimated from the voice signal, and the like. . Specifically, the attribute recognition unit 13, for example, refers to a storage device (not shown) that stores comparison speech data for comparing recognition feature amounts. In this case, the attribute recognition unit 13 can recognize the attribute information by calculating the recognition feature amount and the degree of similarity of the comparison audio data.

<Specific example of the third embodiment>
Next, a specific application example of the speech processing device 300 according to the third embodiment of the present invention will be described.

The speaker features calculated by the speech processing device 300 according to the third embodiment of the present invention can be used for speaker recognition to estimate the speaker of the speech signal. For example, from a first speaker feature calculated from a first speech signal and a second speaker feature calculated from a second speech signal, cosine similarity Calculate degrees. For example, when the purpose is to verify the speaker, it is possible to output determination information as to whether or not verification is possible based on the degree of similarity. Further, when the purpose is to identify a speaker, a plurality of second speech signals may be prepared for the first speech signal, the degree of similarity of each may be obtained, and a set having a large value may be output. .

A speech processing device 300 according to the third embodiment of the present invention is an example of a feature calculation device that calculates a recognition feature amount for recognizing specific attribute information from a speech signal. The speech processing device 300 can be used as a speaker feature extraction device when the specific attribute is the speaker who emitted the speech signal. In addition, the speech processing device 300 includes a mechanism for adapting, for example, a speech signal of a sentence utterance to the speaking style characteristics of the speaker based on the speaker information estimated using the speaker characteristics. It can also be used as part of a device. Further, the information indicating the speaker may be information indicating the gender of the speaker, or information indicating the age or age group of the speaker.

The speech processing device 300 according to the third embodiment of the present invention can be used as a language feature calculation device when a specific attribute is used as information indicating the language (language constituting the speech signal) conveyed by the speech signal. In addition, the speech processing device 300 is also used as part of a speech translation device having a mechanism for selecting a language to be translated based on language information estimated using the language feature, for example, for a speech signal of sentence utterance. It is possible.

The speech processing device 300 according to the third embodiment of the present invention can be used as an emotion feature calculation device when the specific attribute is information indicating the speaker's emotion at the time of speaking. In addition, the speech processing device 300 has a mechanism for identifying a speech signal corresponding to a specific emotion, based on emotion information estimated using the emotion feature, for example, for speech signals of a large number of accumulated utterances. It can also be used as part of a voice search device or voice display device. This emotional information includes, for example, information indicating emotional expression, information indicating the personality of the speaker, and the like.

As described above, the specific attribute information in the present embodiment includes information such as the speaker who issued the audio signal, the language composing the audio signal, the emotional expression included in the audio signal, and the character of the speaker estimated from the audio signal. It is information representing at least one of them.

(Description of hardware configuration)
Although the present invention has been described using the embodiments, the present invention is not limited to the above embodiments. Various changes that can be understood by those skilled in the art can be made to the configuration and details of the present invention within the scope of the present invention. That is, it goes without saying that the present invention is not limited to the above embodiments, and that various modifications are possible and are also included within the scope of the present invention.

As described above, the speech processing device or the like according to one aspect of the present invention has the effect of being able to extract a feature vector in consideration of the quality of the speech signal and improve the accuracy of speaker recognition. etc. and as a speaker recognizer. It should be noted that, in the present invention, acquisition and use of information relating to the user shall be done legally.

<Other embodiments>
The audio processing device may be implemented by hardware or by software. Also, the audio processing device may be realized by a combination of hardware and software.

FIG. 6 is a block diagram showing an example of an information processing device (computer) that constitutes the speech processing device.

As shown in FIG. 6, the information processing apparatus 400 includes a control unit (CPU: Central Processing Unit) 410, a storage unit 420, a ROM (Read Only Memory) 430, a RAM (Random Access Memory) 440, and a communication interface. 450 and a user interface 460 .

The control unit (CPU) 410 expands the programs stored in the storage unit 420 or the ROM 430 into the RAM 440 and executes them, thereby realizing various functions of the speech processing device and the speaker recognition device. Also, the control unit (CPU) 410 may include an internal buffer that can temporarily store data and the like.

The storage unit 420 is a large-capacity storage medium capable of holding various data, and can be realized by a storage medium such as an HDD (Hard Disc Drive) and an SSD (Solid State Drive). Further, when the information processing device 400 is connected to a communication network via the communication interface 450, the storage unit 420 may be a cloud storage existing on the communication network. Further, the storage unit 420 may hold a program readable by the control unit (CPU) 410 .

The ROM 430 is a non-volatile storage device that can be configured with a flash memory or the like that has a smaller capacity than the storage unit 420 . Further, the ROM 430 may hold a program readable by the control unit (CPU) 410 . At least one of the storage unit 420 and the ROM 430 may hold a program readable by the control unit (CPU) 410 .

The program readable by the control unit (CPU) 410 may be supplied to the information processing apparatus 400 in a state of being non-temporarily stored in various computer-readable storage media. Such storage media are, for example, magnetic tapes, magnetic disks, magneto-optical disks, CD-ROMs, CD-Rs, CD-R/Ws, and semiconductor memories.

The RAM 440 is a semiconductor memory such as DRAM (Dynamic Random Access Memory) and SRAM (Static Random Access Memory), and can be used as an internal buffer for temporarily storing data and the like.

The communication interface 450 is an interface that connects the information processing device 400 and a communication network via wire or wireless.

The user interface 460 is, for example, a display unit such as a display, and an input unit such as a keyboard, mouse, and touch panel.

Some or all of the above-described embodiments may be described in the following additional remarks, but are not limited to the following.

[Appendix 1]
receiving means for receiving input of a plurality of audio signals representing speech;
and information processing means for calculating a recognition feature amount for recognizing specific attribute information from the plurality of audio signals based on the quality of the plurality of audio signals.

[Appendix 2]
Further comprising a speech statistic calculation means for calculating a speech statistic representing a ratio of types of sounds contained in the plurality of speech signals,
The speech processing device according to supplementary note 1, wherein the information processing means calculates the recognition feature amount based on the speech statistics of the plurality of speech signals and the quality of the plurality of speech signals.

[Appendix 3]
Said quality is
A value representing the likelihood of speech calculated by identifying whether a portion of the plurality of speech signals is speech, and a value calculated by identifying whether a portion of the plurality of speech signals is correct speech for speaker recognition. at least one of a value representing the likelihood of correct speaker recognition, and a value representing the likelihood of speaker recognition error calculated by identifying whether or not a portion of the plurality of speech signals is speech that causes speaker recognition errors. 3. The audio processing device according to appendix 1 or 2, which is one of the above.

[Appendix 4]
3. The audio processing device according to appendix 3, further comprising quality estimating means for calculating the quality of the plurality of audio signals using a neural network.

[Appendix 5]
The information processing means is
5. The speech processing device according to appendix 3 or 4, wherein an i-vector is calculated as the recognition feature quantity.

[Appendix 6]
6. The speech processing device according to any one of appendices 1 to 5, comprising attribute recognition means for recognizing the attribute information based on the recognition feature amount.

[Appendix 7]
The specific attribute information is
Supplementary notes 1 to 6, which is information representing at least one of the speaker who issued the audio signal, the language that composes the audio signal, the emotional expression contained in the audio signal, and the personality of the speaker estimated from the audio signal. The audio processing device according to any one of the above.

[Appendix 8]
accepts input of a plurality of audio signals representing speech;
A speech processing method, comprising: calculating a recognition feature amount for recognizing specific attribute information from the plurality of speech signals based on the quality of the plurality of speech signals.

[Appendix 9]
further calculating an audio statistic representing the ratio of types of sounds included in the plurality of audio signals;
9. The speech processing method according to appendix 8, wherein the recognition feature quantity is calculated based on the speech statistic of the plurality of speech signals and the quality of the plurality of speech signals.

[Appendix 10]
Said quality is
A value representing the likelihood of speech calculated by identifying whether a portion of the plurality of speech signals is speech, and a value calculated by identifying whether a portion of the plurality of speech signals is correct speech for speaker recognition. at least one of a value representing the likelihood of correct speaker recognition, and a value representing the likelihood of speaker recognition error calculated by identifying whether or not a portion of the plurality of speech signals is speech that causes speaker recognition errors. 10. The audio processing method according to appendix 8 or 9, which is one of

[Appendix 11]
11. The audio processing method of claim 10, wherein a neural network is used to calculate the quality of the plurality of audio signals.

[Appendix 12]
12. The speech processing method according to appendix 10 or 11, wherein an i-vector is calculated as the recognition feature quantity.

[Appendix 13]
13. The speech processing method according to any one of appendices 8 to 12, wherein the attribute information is recognized based on the recognition feature amount.

[Appendix 14]
The specific attribute information is
Supplementary notes 8 to 13, which is information representing at least one of the speaker who emitted the audio signal, the language that composes the audio signal, the emotional expression contained in the audio signal, and the character of the speaker estimated from the audio signal. 1. A speech processing method according to any one of the preceding claims.

[Appendix 15]
to the computer,
a process of accepting input of a plurality of audio signals representing speech;
and calculating a recognition feature amount for recognizing specific attribute information from the plurality of audio signals based on the quality of the plurality of audio signals.

[Appendix 16]
to the computer;
a process of further calculating an audio statistic representing the ratio of types of sounds included in the plurality of audio signals;
16. The speech processing program according to appendix 15, causing execution of a process of calculating the recognition feature quantity based on the speech statistics of the plurality of speech signals and the quality of the plurality of speech signals.

[Appendix 17]
Said quality is
A value representing the likelihood of speech calculated by identifying whether a portion of the plurality of speech signals is speech, and a value calculated by identifying whether a portion of the plurality of speech signals is correct speech for speaker recognition. At least a value representing the likelihood of correct speaker recognition and a value representing the likelihood of speaker recognition error calculated by identifying whether or not a portion of the plurality of speech signals is speech that causes speaker recognition errors. 17. The audio processing program according to appendix 15 or 16, which is any one.

[Appendix 18]
to the computer;
18. The audio processing program according to appendix 17, causing execution of processing for calculating the quality of the plurality of audio signals using a neural network.

[Appendix 19]
to the computer;
19. The speech processing program according to appendix 17 or 18, which executes a process of calculating an i-vector as the recognition feature quantity.

[Appendix 20]
to the computer;
20. The speech processing program according to any one of appendices 15 to 19, which executes a process of recognizing the attribute information based on the recognition feature amount.

[Appendix 21]
The specific attribute information is
Supplementary notes 15 to 20, which is information representing at least one of the speaker who issued the audio signal, the language that composes the audio signal, the emotional expression contained in the audio signal, and the personality of the speaker estimated from the audio signal. The audio processing program according to any one of the preceding claims.

REFERENCE SIGNS LIST 11 contribution estimation unit 12 speaker feature calculation unit 13 attribute recognition unit 21 speech segment detection unit 22 speech statistic calculation unit 23 contribution storage unit 24 ... Contribution degree learning unit 100, 200, 300 ... Speech processing device 400 ... Information processing device 410 ... Control unit (CPU)
420... Storage unit 430... ROM
440 RAM
450... communication interface 460... user interface

Claims (8)

receiving means for receiving input of a plurality of audio signals representing speech;
quality estimating means for calculating two types of quality of the speech that is correct for speaker recognition and the speech that causes an error in speaker recognition, in the plurality of speech signals;
information processing means for calculating a recognition feature amount for recognizing specific attribute information from the plurality of audio signals based on the quality of the plurality of audio signals;
an audio statistic calculation means for calculating an audio statistic representing the degree of occurrence of types of sounds included in the plurality of audio signals;
The speech processing device, wherein the information processing means calculates the recognition feature amount based on the speech statistics of the plurality of speech signals and the quality of the plurality of speech signals. 2. The audio processing device according to claim 1, wherein said quality is a value representing the likelihood of speech calculated by identifying whether a portion of said plurality of audio signals is speech. The information processing means is
3. The speech processing apparatus according to claim 2, wherein an i-vector is calculated as said recognition feature amount. 4. The speech processing apparatus according to claim 1, further comprising attribute recognition means for recognizing said attribute information based on said recognition feature quantity. The specific attribute information is
Claims 1 to 4, wherein the information represents at least one of the speaker who issued the voice signal, the language composing the voice signal, the emotional expression included in the voice signal, and the personality of the speaker estimated from the voice signal. The audio processing device according to any one of . accepts input of a plurality of audio signals representing speech;
calculating two types of quality of speech that is correct for speaker recognition and speech that causes an error in speaker recognition, in the plurality of speech signals;
calculating a recognition feature amount for recognizing specific attribute information from the plurality of audio signals based on the quality of the plurality of audio signals;
calculating an audio statistic representing the frequency of occurrence of types of sounds included in the plurality of audio signals;
The speech processing method, wherein the recognition feature quantity is calculated based on the speech statistic of the plurality of speech signals and the quality of the plurality of speech signals. Said quality is
7. The audio processing method according to claim 6, wherein said value is a value representing likelihood of speech calculated by identifying whether a portion of said plurality of audio signals is speech. to the computer,
a process of accepting input of a plurality of audio signals representing speech;
a process of calculating two types of quality of speech that is correct for speaker recognition and speech that causes an error in speaker recognition in the plurality of speech signals;
a process of calculating a recognition feature amount for recognizing specific attribute information from the plurality of audio signals based on the quality of the plurality of audio signals;
a process of calculating an audio statistic representing the frequency of occurrence of types of sounds included in the plurality of audio signals;
A speech processing program for executing a process of calculating the recognition feature amount based on the speech statistics of the plurality of speech signals and the quality of the plurality of speech signals.

Images