JP6500375B2 - Voice processing apparatus, voice processing method, and program - Google Patents

Description

  The present invention relates to a voice processing apparatus, a voice processing method, and a program for recognizing attribute information such as a speaker's individuality and a spoken language from a voice signal.

  A speech processing apparatus is known which extracts, from a speech signal, acoustic features representing individuality for specifying a speaker who has made speech, and acoustic features representing a language that the speech conveys. Also, a speaker recognition device that estimates a speaker from a speech signal using these acoustic features, and a language recognition device that estimates a language are known.

  In a speaker recognition apparatus using this kind of speech processing apparatus, the speech processing apparatus evaluates the similarity between the acoustic feature extracted from the speech signal and the speaker model that expresses the speaker dependency of the appearance tendency of the acoustic feature. , Select a speaker based on the evaluation. The speaker recognition apparatus selects, for example, a speaker of a speaker model evaluated as having the highest similarity. At this time, if distortion occurs in the acoustic feature due to a partial loss of sound type or noise mixing in the voice signal input to the speaker recognition device, the acoustic feature of the speaker model There may be differences between them, which may reduce the accuracy of speaker recognition. The following prior art documents describe techniques for suppressing the degradation of the accuracy of speaker recognition.

  The prior art documents describe techniques for adjusting the criteria for determining the speaker recognition based on the characteristics of the speech signal input to the speaker recognition device.

  In Non-Patent Document 1, the duration of the speech signal is measured as a characteristic representing the quantity of the speech signal, and according to the value, the speaker recognition result is given a reliability indicating the reliability of the result. Techniques for suppressing speaker recognition errors are described. Further, in Patent Document 1, the ratio of voiced sound to unvoiced sound included in the voice signal and the ratio of repetitive utterance sections included in the voice signal are calculated as characteristics representing the diversity of the voice signal, and the value is used as a speaker recognition result There is disclosed a technique for suppressing the degradation of the speaker recognition accuracy by shifting the determination threshold of the speaker recognition by using it as the reliability of the.

  In Patent Document 2, a distance between each voice quality is calculated using a plurality of voice quality features of a DB (database) and weights calculated from inputs. A technique is described that calculates coordinates of each voice quality on the voice quality space using the distance of the voice quality, and displays speaker attribute information for the voice quality to the coordinates on the voice quality space calculated by the display unit. Patent Document 3 is an apparatus for speaker authentication that determines the legitimacy of the speaker based on the feature amount of the speaker's voice, that is, whether or not the user is a legitimate user registered in advance. A technique is described which sequentially executes the calculation of feature vectors and the update of the update mixture model for each unit interval in which the registration intervals necessary for determining the mixture model are divided.

JP, 2007-156422, A JP 2008-233759 A International Publication No. 2008/149547

W. M. Campbell, D. A. Reynolds, J.A. P. Campbell, and K. J. Brady, “Estimating and Evaluating Confidence for Forensic Speaker Recognition,” in Proceedings of IEEE International Conference on Acoustics, Speech, and Signal Processing, 2005.

  However, in the technique described in the prior art document, there is a problem that the reduction in the speaker recognition accuracy may not be sufficiently suppressed because the above-mentioned reliability is not properly obtained.

  According to the technique described in Non-Patent Document 1, the longer the duration of the speech signal input to the speaker recognition apparatus, the smaller the loss or bias of the type of sound included in the speech signal, and the confidence in the speaker recognition result. It is assumed that the sex is high. However, since the technology does not explicitly evaluate the type of sound included in the audio signal, it is not appropriate when the same word is repeated even if the duration of the audio signal is large.

  According to the technique described in Patent Document 1, the determination threshold for speaker recognition is set to a different value according to the magnitude of the diversity of sounds included in the speech signal input to the speaker recognition device. However, the same technology calculates only the characteristics of the speech signal input to the speaker recognition device, and adjusts the speaker recognition criteria based on a predetermined uniform standard. Therefore, when there is a difference between the characteristic representing the diversity of the speech signal input to the speaker recognition device and the characteristic representing the diversity of the speech signal of each speaker used in learning each speaker model , There is a possibility that the judgment criteria of speaker recognition can not be adjusted appropriately. Such a problem may occur when, for example, the training speech signal of the speaker model can not be obtained sufficiently.

  The technology described in Patent Document 2 calculates the distance between each voice quality using a plurality of voice quality features of a DB (database) and weights calculated from the input. Then, the technology described in Patent Document 2 calculates the coordinates of each voice quality on the voice quality space using the distance of the voice quality, and converts the coordinates on the voice quality space calculated by the display unit to the speaker attribute information for the voice quality. Is displayed, but the reliability of the displayed coordinates is not calculated. In addition, the technology described in Patent Document 3 sequentially executes the calculation of feature vectors and the update of the update mixture model for each unit interval obtained by dividing the registration intervals necessary for determining the mixture model. The reliability of the mixed model to be updated is not calculated.

  An object of the present invention is to provide a speech processing apparatus which suppresses the degradation of the accuracy of speaker recognition by solving the above problems and appropriately determining the reliability of the speaker recognition result.

  An audio processing device according to one aspect of the present invention includes acoustic diversity calculation means for calculating acoustic diversity representing a degree of variation regarding the type of the audio signal based on an audio signal representing audio, and the acoustic diversity of the audio signal. Calculating the acoustic reliability indicating the degree to which the audio signal is suitable for recognizing specific attribute information from the audio signal based on the degree of difference between the acoustic diversity of the other audio signal as a reference And a reliability calculation means.

  The sound processing method according to one aspect of the present invention calculates an acoustic diversity degree indicating a degree of variation regarding the type of sound included in the sound signal based on the sound signal representing the sound, and the sound diversity of the sound signal Based on the degree of difference between the acoustic diversity of another audio signal as a reference, acoustic reliability is calculated that represents the degree to which the audio signal is suitable for recognizing specific attribute information from the audio signal.

  A program according to one aspect of the present invention includes an acoustic diversity calculation process of calculating acoustic diversity representing the degree of variation regarding the type of sound included in the audio signal based on an audio signal representing audio; Based on the degree of acoustic diversity and the degree of difference between the acoustic diversity of another audio signal as a reference, the audio reliability indicating the degree to which the audio signal is suitable for recognizing specific attribute information from the audio signal is calculated The computer is made to execute the sound reliability calculation processing to be performed.

  The present invention can suppress the degradation of the accuracy of speaker recognition by appropriately determining the reliability of the speaker recognition result.

It is a block diagram showing composition of speech processing unit 100 in a 1st embodiment of the present invention. It is a flowchart which shows operation | movement of the speech processing unit 100 in the 1st Embodiment of this invention. It is a figure explaining the relationship between the continuation time length of an audio | voice signal, and the acoustic reliability which the acoustic reliability calculation part 14 calculates. It is a block diagram which shows the structure of the speaker recognition apparatus 200 in the 2nd Embodiment of this invention. It is a flowchart which shows operation | movement of the speaker recognition apparatus 200 in the 2nd Embodiment of this invention. It is a block diagram which shows the structure of the speaker model learning apparatus 300 in the 3rd Embodiment of this invention. It is a flowchart which shows operation | movement of the speaker model learning apparatus 300 in the 3rd Embodiment of this invention. It is a block diagram which shows the structure of the speech processing unit 400 in the 4th Embodiment of this invention.

  Hereinafter, embodiments of the speech processing device and the like and the speaker recognition device will be described with reference to the drawings. In addition, since the component which attached | subjected the same code | symbol in each embodiment performs the same operation | movement, description for the second time may be abbreviate | omitted.

First Embodiment
FIG. 1 is a block diagram of the speech processing apparatus 100 according to the first embodiment. The voice processing device 100 includes a voice model storage unit 11, an acoustic diversity calculation unit 12, an acoustic diversity storage unit 13, and an acoustic reliability calculation unit 14.

  The acoustic diversity calculation unit 12 receives an audio signal. Here, acceptance means, for example, reception from an external device, and delivery of processing results from another processing device or another program.

  The voice model storage unit 11 stores one or more voice models. The speech model has information for calculating, for the speech signal, numerical information indicating the degree of matching with the model. For example, when the speech model is a Gaussian mixture model (GMM: Gaussian Mixture Model), the speech processing apparatus 100 can calculate the appearance probability of the speech signal based on the average, the variance, and the mixing coefficient of the Gaussian mixture distribution.

  The speech model stored in the speech model storage unit 11 is a speech model trained using a training speech signal according to a general optimization criterion such as a maximum likelihood criterion. Unlike the training of an acoustic model generally used in speech recognition, the training of a speech model does not require linguistic information (such as words) representing the content of a training speech signal. The voice model storage unit 11 may store, for example, two or more voice models in which training voice signals are learned separately, such as gender of a speaker (male and female) and recording environment separately (indoor and outdoor). Good.

  The acoustic diversity calculation unit 12 receives an audio signal, refers to one or more audio models stored in the audio model storage unit 11, and calculates an audio diversity representing the type of sound included in the audio signal. Do. Then, the acoustic diversity calculation unit 12 outputs the processing result together with the received audio signal. The type of sound is represented, for example, as a grouping of sounds obtained by automatically grouping or clustering audio signals based on the degree of similarity, that is, a class of sounds. Here, the output means, for example, transmission to an external device, and delivery of the processing result to another processing device or another program. Further, the output is a concept including, for example, display on a display, projection using a projector, printing on a printer, and the like.

An example of a method of calculating the acoustic multiplicity V (x) of the audio signal x by the acoustic diversity computation unit 12 will be described. For example, when the speech model is a Gaussian mixture, each element distribution of the Gaussian mixture represents different sounds. Therefore, the acoustic diversity calculation unit 12 obtains, for the audio signal x, the posterior probability of each element distribution of the mixed Gaussian distribution that is an audio model. The posterior probability P _i (x) of the i-th element distribution of the mixed Gaussian distribution can be calculated by the following equation.

Here, N represents the probability density function of the Gaussian distribution, θ _i is the parameter (average and variance) of the i-th element distribution of the mixed Gaussian distribution, and w _i is the mixing coefficient of the i-th element distribution of the mixed Gaussian distribution Represent. This P _i (x) indicates how much the speech signal x belongs to the i-th element distribution of the mixed Gaussian distribution. The acoustic multiplicity V (x) is configured as a vector having this P _i (x) as an element. For example, when the mixing number of the mixture Gaussian distribution, which is a speech model, is four, the acoustic diversity is expressed as V (x) = (P ₁ (x), P ₂ (x), P ₃ (x), P ₄ (x It defines as).

In addition, another example of a method for the acoustic diversity calculation unit 12 to calculate the acoustic diversity V (x) of the audio signal x will be described. For example, when the speech model is a mixed Gaussian distribution, acoustic diversity calculating unit 12, the time series _{{x 1, x 2, ...} , x T} brief audio signal a sound signal x is divided into a short time voice For each of the signals, an element distribution number i for which the appearance probability is maximum is determined by Argmax _i (x _t ) = N (x _t | θ _i ). Assuming that the number of times the i-th element distribution of the mixed Gaussian distribution is selected is C _i (x), this can be interpreted as representing how much the speech signal x belongs to the i-th element distribution of the mixed Gaussian distribution . The acoustic diversity V (x) is configured as a vector having this C _i (x) or C _i (x) / Σ _j C _j (x) as an element. For example, when the mixing number of the mixture Gaussian distribution, which is a speech model, is four, the acoustic multiplicity is V (x) = (C ₁ (x), C ₂ (x), C ₃ (x), C ₄ (x It defines as).

  The acoustic diversity calculation unit 12 may calculate the acoustic diversity of the audio signal obtained by dividing the received audio signal. The acoustic diversity calculation unit 12 may, for example, segment the received audio signal into audio signals of a predetermined time, and calculate the acoustic diversity of each segmented audio signal as a result of the segmentation. The acoustic diversity calculation unit 12 may also calculate the acoustic diversity of the audio signal received at that time when the duration of the audio signal exceeds a predetermined value in synchronization with the reception of the audio signal. Good.

  When referring to two or more speech models stored in the speech model storage unit 11, the acoustic diversity calculation unit 12 performs weighted addition on the two or more acoustic diversity calculated based on the respective speech models. You may make it the acoustic diversity.

  The acoustic diversity calculation unit 12 calculates the likelihood for the element distribution of the speech model of the speech signal based on the speech model that is configured as a mixed model with the probability distribution as the element distribution and represents the appearance probability distribution of the speech signal The acoustic diversity of the speech signal may be calculated based on the likelihood or a value obtained by normalizing the likelihood with the sum of the likelihoods of all the element distributions.

  As described above, the acoustic diversity calculation unit 12 calculates the acoustic diversity based on the speech model stored in the speech model storage unit 11. Since a speech model is represented by one or a small number of mixed Gaussian distributions, it is more likely to be a speech model used in general speech recognition than a language information, for example, one having a mixed Gaussian distribution different for each phoneme, etc. There are few parameters to configure. For the same reason, the amount of calculation required to calculate the acoustic diversity is also less than in general speech recognition. From this, the acoustic diversity calculation unit 12 can calculate the acoustic diversity faster than the technology described in Patent Document 1 that uses language information.

The acoustic diversity storage unit 13 stores information indicating one or more acoustic diversity. The acoustic diversity storage unit 13 stores, in advance, the acoustic diversity calculated for the one or more audio signals in the same manner as the above-described acoustic diversity calculation unit 12. For example, when learning the speaker model included in the speaker recognition apparatus, the sound diversity storage unit 13 associates the sound diversity calculated for each of the audio signals used for learning of each speaker model with each speaker. To memorize. Speech processing apparatus 100 includes, for example, three speakers (A, B, C) audio signal _x A _of, x B, when learning the three speaker models using _{x C,} of the respective audio signals calculating an acoustic diversity _stores {a, V (x a) }, {B, V (x B)}, the information represented as _{{C, V (x C)} }.

  The sound reliability calculation unit 14 receives the sound signal output from the sound diversity calculation unit 12 and the sound diversity thereof, and refers to the sound diversity stored in the sound diversity storage unit 13 to obtain the sound signal The audio reliability is calculated, and the processing result is output together with the audio signal and the audio diversity output from the audio diversity calculation unit 12.

An example of a method of calculating the sound reliability R (x) of the audio signal x by the sound reliability calculation unit 14 will be described. For example, it is assumed that the acoustic multiplicity V (x) of the audio signal x is a multinomial distribution (vector) having P _i (x) which is the calculation result of the above [Equation 1] as an element. Further, it is assumed that V _A = V (x _A ), which is one of one or more acoustic diversity stored in the acoustic diversity storage unit 13, has the same multinomial distribution. For example, V _A is the acoustic diversity of the speaker model training speech signal of speaker A. The acoustic reliability is configured as the difference between these two multinomial distributions V (x) and _VA . The acoustic reliability calculation unit 14 determines the acoustic reliability using an arbitrary scale that indicates the degree of difference between the two multinomial distributions V (x) and _VA , for example, the amount of information of the Kullback-Leibler and the degree of cosine similarity. For example, the Kullack-Leibler information amount is a measure for determining the difference between two probability distributions. For example, when calculating the sound reliability according to the following equation, the value takes a positive value, and the smaller the difference between the two multinomial distributions, the larger the value.

Alternatively, the sound reliability calculation unit 14 may calculate the entropy of the sound multiplicity V (x) of the audio signal x to obtain the sound reliability. In this case, acoustic reliability R (x) = - when _{Σ i V i (x) logV} i (x) is a positive value, the type of sound contained in the audio signal x is uniform, the largest Take a value.

  The sound reliability calculation unit 14 may calculate the sound reliability of the sound signal obtained by dividing the received sound signal. The sound reliability calculation unit 14 may divide the audio signal in the same manner as the sound diversity calculation unit 12 or may divide the sound signal differently.

  The acoustic reliability calculation unit 14 outputs two or more results calculated based on the respective acoustic diversity degrees when referring to two or more acoustic diversity degrees stored in the acoustic diversity degree storage unit 13 You may

  As described above, it is possible to express the diversity representing the degree of the variation of the type of sound included in the audio signal by the acoustic diversity calculated by the acoustic diversity calculation unit 12. Thereby, for example, when there are two audio signals of the same length, one is a repetition of the same word, and the other is not, using the sound diversity, the types of sound included in each of the two audio signals are It can represent different things. That is, since the degree of the variation of the kind of sound is small in the sound signal of the same word, the degree of acoustic diversity is low. On the other hand, since the degree of the variation of the kind of sound is large in the sound signal which is not the repetition of the same word, the degree of acoustic diversity increases. The sound reliability calculation unit 14 calculates high sound quality of the sound signal suitable for speaker recognition by calculating the sound reliability of the sound signal with less deviation or omission in the sound type to be high based on the sound diversity degree. Confidence can be given.

  Note that the technology described in Patent Document 1 uses the ratio of voiced sound to unvoiced sound included in voice and the ratio of repeated utterance sections as a characteristic indicating the diversity of sounds included in a voice signal. However, in order to calculate these characteristics, it is necessary to estimate symbols corresponding to speech (such as voiced speech, unvoiced speech, and words). Since the speech recognition technology used for such symbol estimation generally requires computation time, there is a problem that it is not suitable for high-speed speaker recognition processing.

  On the other hand, the sound reliability calculation unit 14 according to the first embodiment is the sound diversity degree of the received voice signal calculated by the sound diversity degree calculation unit 12 and the sound diversity degree storage unit 13 that are stored. Based on the acoustic diversity of the audio signal, a measure representing the difference between the two is calculated, and the acoustic reliability is determined based on the calculation result. The sound reliability calculation unit 14 calculates sound reliability based on, for example, the difference in acoustic diversity between the sound signal received by the speaker recognition device and the sound signal used for learning the speaker model. When evaluating the similarity between the speech signal received by the speaker recognition device and the speaker model, rather than evaluating the similarity based on sounds of types not included in the training speech signal of the speaker model The degree of similarity is more reliable if the degree of similarity is evaluated based on the type of sound included in the training speech signal of the speaker model. The acoustic reliability calculation unit 14 can not acquire the speaker model learning voice signal of a sufficient length, and even if the type of sound included in the learning voice signal is uneven or missing, the acoustic reliability can be obtained. Can be calculated accurately.

  The sound reliability calculation unit 14 may calculate the sound reliability of the segmented audio signal obtained by subjecting the audio signal to any segmentation. Thereby, the speech processing apparatus 100 according to the first embodiment can calculate the acoustic reliability more accurately than the calculation method based on the duration time of the speech signal, which is the technology described in Non-Patent Document 1. The acoustic reliability will be described with reference to FIG. The graph of FIG. 3 shows the relationship between the duration of the audio signal and the acoustic diversity of the audio signal at that time. The graph (1) shown in FIG. 3 is based on the technique described in Non-Patent Document 1, and shows that the sound reliability is 1 when the duration time of the audio signal is 30. In the technique described in Non-Patent Document 1, any audio signal is similar to the graph of (1) shown in FIG. On the other hand, in the technique according to one aspect of the present invention, the appearance of the graph is different depending on the variety of sounds included in the audio signal. When the diversity of the audio signal is large, the acoustic reliability takes a high value even at a stage where the duration is small, for example, when the duration is 20 as shown in the graph (2) shown in FIG. The degree may be one. In addition, when the diversity of the head portion of the audio signal is small, the acoustic reliability does not have a high value until the duration becomes large. For example, as in the graph of (3) shown in FIG. The sound reliability may be 1 at 40. As described above, the speech processing apparatus 100 according to the first embodiment calculates the acoustic reliability of the speech signal based on the acoustic diversity, so that the acoustic reliability can be more accurately obtained than the technology described in Non-Patent Document 1. It can be asked.

  Although the voice model storage unit 11 and the acoustic multiplicity storage unit 13 are preferably non-volatile recording media, they can also be realized as volatile recording media.

  The process in which the speech model is stored in the speech model storage unit 11 does not matter. For example, the speech model may be stored in the speech model storage unit 11 via the recording medium, and the speech model storage unit 11 may store the speech model transmitted via the communication line or the like. Alternatively, the speech model storage unit 11 may store the speech model input through the input device. The same applies to the acoustic diversity degree storage unit 13.

  The sound diversity calculation unit 12 and the sound reliability calculation unit 14 can be realized by, for example, an arithmetic device or a memory. The processing procedure of the acoustic diversity calculation unit 12 or the like is realized by, for example, software, and the software is stored in a storage medium such as a ROM. Also, it may be realized by hardware (dedicated circuit).

(Operation of Voice Processing Device 100 in First Embodiment)
Next, the operation of the speech processing apparatus 100 according to the first embodiment will be described with reference to the flowchart of FIG. FIG. 2 is a flowchart showing the operation of the speech processing apparatus 100.

  The voice processing apparatus 100 receives one or more voice signals from the outside, and provides the voice diversity calculation unit 12 (step S101). The acoustic diversity calculation unit 12 calculates the acoustic diversity with reference to the one or more speech models stored in the speech model storage unit 11 for each of the received one or more speech signals (step S102). The acoustic reliability calculation unit 14 refers to one or more received audio signals and their acoustic diversity, and one or more acoustic diversity stored in the acoustic diversity storage unit 13 as necessary. The acoustic reliability is calculated (step S103). The speech processing apparatus 100 ends the series of processing when reception of the speech signal from the outside is completed (Yes in step S104). When the reception of the external audio signal has not ended (No in step S104), the speech processing device 100 returns the process to step S101.

  This is the end of the operation of the speech processing device 100 according to the first embodiment.

(Effects of the first embodiment)
As described above, according to the speech processing apparatus 100 of the first embodiment, a high-speed calculation method using no speech information, in which sound reliability is calculated for speech signals based on speech diversity. Use Therefore, in the case where there is a bias or omission in the types of sounds included in the audio signal, the audio processing apparatus 100 has a large difference between the audio diversity of the audio signal and the audio diversity of the speaker model learning audio signal. The acoustic reliability of the speech signal can be underestimated. Thus, the speech processing apparatus 100 according to the first embodiment can suppress the output of the speaker recognition result for the speech signal not suitable for the speaker recognition, and can suppress the reduction in the speaker recognition accuracy.

Second Embodiment
FIG. 4 is a block diagram of the speaker recognition apparatus 200 in the second embodiment. The speaker recognition apparatus 200 includes a voice section detection unit 21, a voice processing unit 22, a speaker model storage unit 23, a speaker recognition calculation unit 24, and a speaker recognition output unit 25.

  The speaker recognition device 200 in the present embodiment is an example of an attribute recognition device that recognizes specific attribute information from a voice signal, and the present embodiment is applicable to all attribute recognition devices. Another example of the attribute recognition device is, for example, a language recognition device. The speaker recognizer recognizes information indicative of the speaker who originated the speech signal. The language recognition device recognizes information indicating the language that the speech signal conveys. That is, the present embodiment is applicable to a speaker recognition apparatus and a language recognition apparatus.

  The voice segment detection unit 21 receives a voice signal, detects a voice segment included in the voice signal, segments the voice segment, and outputs a segmented voice signal as a processing result. Here, acceptance means, for example, reception from an external device, and delivery of processing results from another processing device or another program. For example, the voice section detection unit 21 determines that a section in which the sound volume is smaller than a predetermined value continuously for a predetermined time in the voice signal is silent, and determines that the sections before and after that section are different voice sections. Good.

  The voice processing unit 22 receives one or more voice signals output from the voice section detection unit 21, performs voice processing corresponding to the voice processing apparatus 100 according to the first embodiment, and determines the acoustic reliability of the voice signal. Calculate and output the acoustic reliability of the processing result. The configuration and operation of the audio processing unit 22 may be similar to the configuration and operation of the audio processing device 100 in the first embodiment. For example, the audio processing unit 22 may be the audio processing device 100.

  The speaker model storage unit 23 stores one or more speaker models. The speaker model has information for giving the speech signal numerical information (score) indicating the degree of matching with the model. For example, when the speaker model has a mixed Gaussian distribution, the speaker recognition apparatus 200 can calculate the appearance probability of the speech signal as a score based on the average, the variance, and the mixing coefficient of the mixed Gaussian distribution. Here, the mixed Gaussian distribution corresponding to each speaker is optimized according to a general standard such as a maximum likelihood standard or a maximum a posteriori probability standard using a speech signal to which speaker IDs are given as teacher labels. Use Here, the speaker ID is an identifier for identifying a speaker.

  The speaker recognition calculation unit 24 receives the voice signal output from the voice section detection unit 21, refers to one or more speaker models stored in the speaker model storage unit 23, and receives the voice signal and each voice signal. The degree of matching with the speaker model is calculated, and the speaker recognition result is calculated and output to the speaker recognition output unit 25.

  The speaker recognition result output from the speaker recognition calculation unit 24 is, for example, a speaker arranged in the order of scores calculated based on each speaker model when the speaker recognition apparatus 200 aims to identify the speakers. It is in the form of a list of IDs. Further, the speaker recognition result output from the speaker recognition calculation unit 24 may be, for example, a score calculated based on the speaker model to be matched when the speaker recognition apparatus 200 aims to perform speaker verification. It is judgment information on the basis of whether or not collation is possible.

  The speaker recognition output unit 25 receives the speaker recognition result output from the speaker recognition calculation unit 24 and the sound reliability output from the voice processing unit 22, and changes the speaker recognition result as necessary. Output to the outside. Here, the output means, for example, transmission to an external device, and delivery of the processing result to another processing device or another program. The term "output" is a concept including display on a display, projection using a projector, printing on a printer, and the like.

  An example of a method of creating a speaker recognition result output by the speaker recognition output unit 25 will be described. For example, when the speaker recognition apparatus 200 aims to identify a speaker, the speaker recognition result received from the speaker recognition calculation unit 24 is represented by the speaker ID list as described above. The speaker recognition output unit 25 determines the speaker recognition result based on the acoustic diversity of the speech signal and the acoustic diversity of the learning speech signal of the speaker model of the speaker ID for each speaker ID. Refer to the calculated sound reliability. Then, the speaker recognition output unit 25 removes the speaker ID from the speaker recognition result, for example, when the sound reliability is lower than a predetermined value. Alternatively, the speaker recognition output unit 25 includes, for example, a value calculated based on recalculation such as an operation of adding a score to the result of speaker recognition by multiplying the sound reliability by a constant coefficient. The speaker recognition output unit 25 sets the speaker recognition result including the score, that is, the speaker recognition score, as a new speaker recognition result. Alternatively, the speaker recognition output unit 25 outputs the new speaker recognition result.

  In addition, another example of a method of creating a speaker recognition result output by the speaker recognition output unit 25 will be described. For example, when the purpose of the speaker recognition apparatus 200 is to perform speaker verification, the received speaker recognition result is determination information as to whether or not the verification is possible as described above. The speaker recognition output unit 25 calculates the sound reliability based on the sound recognition degree of the speaker recognition result based on the sound diversity degree and the sound diversity degree of the learning speech signal of the speaker model of the speaker to be compared. refer. For example, when the acoustic reliability is lower than a predetermined value, the speaker recognition output unit 25 adds the acoustic reliability to the verification availability information as a speaker recognition result. Alternatively, for example, when the sound reliability is lower than a predetermined value, the speaker recognition output unit 25 uses information indicating that the matching is impossible as the speaker recognition result, instead of the matching availability information.

  The speaker model storage unit 23 is preferably a non-volatile storage medium, but can be realized also as a volatile storage medium.

  The process for storing the speaker model in the speaker model storage unit 23 does not matter. For example, the speaker model may be stored in the speaker model storage unit 23 through the recording medium, and the speaker model transmitted through the communication line or the like is stored in the speaker model storage unit 23 The speaker model may be stored in the speaker model storage unit 23 or the speaker model input via the input device.

  The voice section detection unit 21, the voice processing unit 22, the speaker recognition calculation unit 24, and the speaker recognition output 25 can be realized by, for example, an arithmetic processing unit or a memory. The processing procedure of the voice section detection unit 21 or the like is realized by, for example, software, and the software is stored in a storage medium such as a ROM. Also, it may be realized by hardware (dedicated circuit).

(Operation of Speaker Recognition Device 200 in Second Embodiment)
Next, the operation of the speaker recognition apparatus 200 will be described using the flowchart of FIG. FIG. 5 is a flowchart showing the operation of the speaker recognition apparatus 200.

  The voice section detection unit 21 receives a voice signal from the outside, performs segmentation by voice section detection on the received voice signal, and generates one or more segmented voice signals as the voice processing unit 22 and the speaker recognition calculation unit 24. (Step S201). The voice processing unit 22 performs the voice processing (the processing of steps S101 to S104) of the voice processing apparatus 100 according to the first embodiment on the received one or more segmented voice signals to calculate the sound reliability, This is provided to the speaker recognition output unit 25 (step S202). The speaker recognition calculation unit 24 executes speaker recognition calculation with reference to the one or more speaker models stored in the speaker model storage unit 23 for the received one or more segmented speech signals. Then, the speaker recognition result is provided to the speaker recognition output unit 25 (step S203). The speaker recognition output unit 25 receives the speaker recognition result and the sound reliability regarding the speech signal, changes the speaker recognition result as necessary, outputs the result to the outside, and ends the series of processing ( Step S204).

  As described above, in the speaker recognition apparatus 200 according to the present embodiment, after the speaker recognition calculation unit 24 performs the speaker recognition calculation, the speaker recognition output unit 25 generates the speaker recognition result based on the acoustic reliability. It works like changing it.

(Effect of the second embodiment)
As described above, according to the speaker recognition apparatus 200 of the second embodiment, the input is performed using the speech processing apparatus of the first embodiment (in the present embodiment, the speech processing unit 22). The sound reliability is calculated based on the voice diversity degree for the voice signal, and when there is a bias or omission in the type of sound included in the voice signal, the acoustic diversity of the voice signal and the sound of the speaker model learning voice signal If the difference between the diversity is large, the acoustic reliability of the audio signal can be underestimated. Thus, the speaker recognition apparatus 200 according to the second embodiment can suppress the output of the speaker recognition result for the voice signal not suitable for the speaker recognition, and can suppress the reduction in the speaker recognition accuracy.

  In the technique described in Non-Patent Document 1, the longer the speech signal is, the longer it takes to calculate the speaker recognition time and the response time required to output the speaker recognition result. There is a problem that it is not suitable. On the other hand, in the speaker recognition apparatus 200 of the second embodiment, for example, when there is a bias or omission in the type of sound included in the audio signal, the acoustic diversity of the audio signal and the audio of the audio signal for speaker model learning. If the difference between the multiplicity is large, the acoustic reliability of the speech signal is underestimated. Therefore, the speaker recognition apparatus 200 can suppress the output of the speaker recognition result for the voice signal which is not suitable for the speaker recognition, and is suitable even in a use scene requiring high speed processing.

(Modification of the second embodiment)
A configuration similar to that of the second embodiment is to provide the speaker recognition calculation unit 24 with the sound reliability calculated by the speech processing unit 22. In this configuration, whether or not the speaker recognition calculation unit 24 performs the speaker recognition calculation based on the sound reliability, for example, starts the speaker recognition calculation when the sound reliability exceeds a predetermined value. Operate to determine Therefore, the speaker recognition apparatus 200 in the similar form of the present embodiment has an advantage that the calculation for the speaker recognition can be omitted.

Third Embodiment
FIG. 6 is a block diagram of a speaker model learning device 300 in the third embodiment. The speaker model learning device 300 includes a voice section detection unit 31, a voice processing unit 32, a voice model storage unit 33, a speaker model storage unit 34, a speaker model learning unit 35, and a voice signal input request unit 36.

  The voice section detection unit 31 receives a voice signal via the input unit and detects a voice section included in the voice signal, as in the voice section detection unit 21 of the speaker recognition apparatus 200 according to the second embodiment. It segments and outputs a segmented speech signal which is the processing result. The configuration and operation of the voice activity detection unit 31 are the same as the configuration and operation of the voice activity detection unit 21 in the second embodiment.

  The voice processing unit 32 receives one or more voice signals output from the voice section detection unit 31, performs voice processing corresponding to the voice processing apparatus 100 according to the first embodiment, and detects the acoustic diversity of the voice signal. It is stored in the acoustic diversity storage unit in the speech processing apparatus 100. Further, the sound reliability of the audio signal is calculated, and the sound reliability of the processing result is output. The configuration and operation of the audio processing unit 32 are the same as the configuration and operation of the audio processing unit 100 in the first embodiment and the audio processing unit 22 in the second embodiment. Also, the audio processing unit 32 is, for example, the audio processing device 100.

  The speech model storage unit 33 stores one or more speech models, as in the speech model storage unit 11 of the first embodiment. The speech model is, for example, a mixed Gaussian distribution. The configuration and operation of the speech model storage unit 33 are the same as the configuration and operation of the speech model storage unit 11 in the first embodiment.

  The speaker model storage unit 34 stores one or more speaker models, as in the speaker model storage unit 23 of the second embodiment. The speaker model is, for example, a mixed Gaussian distribution. The configuration and operation of the speaker model storage unit 34 are the same as the configuration and operation of the speaker model 23 in the second embodiment.

  The speaker model learning unit 35 receives one or more voice signals, refers to the voice model stored in the voice model storage unit 33, creates a speaker model using both of them, and creates a speaker model storage unit. Store in 34 When both the speech model and the speaker model are mixed Gaussian distributions, the speaker model uses the speech model as the initial model and the teacher label of the speech signal as the speaker ID, and the general rule such as maximum likelihood criterion or posterior probability maximization criterion. Use one that is optimized according to the standard.

  The speech signal input request unit 36 receives the acoustic reliability calculated by the speech processing unit 32 for the speech signal input to the speaker model learning device 300, and calculates request information representing the necessity of the speech signal input, Output to the outside. Here, the output means, for example, transmission to an external device, and delivery of the processing result to another processing device or another program. The term "output" is a concept including display on a display, projection using a projector, printing on a printer, and the like.

  For example, when the sound reliability received from the sound processing unit 32 is higher than a predetermined value, the sound signal input requesting unit 36 outputs request information indicating that input of the sound signal is stopped. The voice signal input request unit 36 outputs request information indicating stop of voice signal input and presents the speaker that the voice signal input may be stopped, so that the speaker can input more voice signals than necessary. Reduce the burden of

  For example, when the sound reliability received from the sound processing unit 32 is lower than a predetermined value, the sound signal input request unit 36 outputs request information indicating input continuation of the sound signal. The speech signal input request unit 36 outputs the request information indicating the continuation of the input of the speech signal and presents a display prompting the speaker to continue the input of the speech signal, whereby the speaker stops the input of the speech signal quickly. Can be suppressed.

  The voice section detection unit 31 receives the request information output from the voice signal input request unit 36. The voice section detection unit 31 may stop the process of detecting the voice section, for example, when the request information indicating the stop of the input of the voice signal is received. With this configuration, the speaker model learning device 300 can acquire an audio signal suitable for speaker model learning with a minimum necessary duration, and can therefore create an appropriate speaker model with a small amount of calculation.

  In addition, the speaker model learning unit 35 may not start speaker model learning, for example, when the voice section detection unit 31 receives the request information indicating the input continuation of the voice signal. With this configuration, the speaker model learning device 300 can reduce the amount of calculation due to the speaker model learning unit 35 performing the speaker model learning process with the audio signal not suitable for the speaker model learning.

  Although the voice model storage unit 33 and the speaker model storage unit 34 are preferably non-volatile recording media, they can also be realized as volatile recording media.

  There is no limitation on the process in which the speaker model is stored in the speech model storage unit 33. For example, a voice model may be stored in the voice model storage unit 33 via a recording medium, and a voice model transmitted via a communication line or the like is stored in the voice model storage unit 33. Alternatively, the speech model storage unit 33 may store the speech model input through the input device. The same applies to the speaker model storage unit 34.

  The voice section detection unit 31, the voice processing unit 32, the speaker model learning unit 35, and the voice signal input request unit 36 can be realized by, for example, an arithmetic processing unit, a memory, or the like. The processing procedure of the voice section detection unit 31 or the like is realized by, for example, software, and the software is stored in a storage medium such as a ROM. Also, it may be realized by hardware (dedicated circuit).

(Operation of Speaker Model Learning Device 300 in Third Embodiment)
Next, the operation of the speaker model learning device 300 will be described using the flowchart of FIG. FIG. 7 is a flowchart showing the operation of the speaker model learning device 300.

  The voice section detection unit 31 receives a voice signal from the outside, performs segmentation on the received voice signal by detecting a voice section, and outputs one or more segmented voice signals (step S301). The voice processing unit 32 performs voice processing (processing of steps S101 to S104) of the voice processing apparatus 100 according to the first embodiment on the received voice signal to calculate acoustic diversity, and the voice processing unit 32 It stores in the acoustic diversity storage unit, and calculates and outputs acoustic reliability (step S302). The audio signal input request unit 36 receives the sound reliability and outputs request information indicating that the input of the audio signal is requested (step S303). If the voice activity detection unit 31 receives (detects) request information indicating that the voice activity detection requested by the voice signal input request unit 36 is to be stopped (Yes in step S304), the process returns to step S301. When the speech segment detection unit 31 does not receive the request information indicating that the speech segment detection requested by the speech signal input request unit 36 is to be stopped (No in step S304), the speaker model learning unit 35 receives the request. For the speech signal, the speech model stored in the speech model storage unit 33 is referenced to learn the speaker model, and the speaker model is stored in the speaker model storage unit 34 (step S304). .

  In addition, the speaker model learning device 300 detects that the speech input to the speech segment detection unit 31 is completed, or that the speech segment detection unit 31 receives the request information indicating that the detection of the speech segment is stopped. When the processing has been completed up to step S304 for the audio signal received at that time, the series of processing ends.

(Effect of the third embodiment)
As described above, according to the speaker model learning device 300 of the third embodiment, by using the voice processing device of the first embodiment, the input voice signal can be converted to voice diversity degree. The sound reliability based on the sound can be calculated, and the sound signal input can be requested when the types of sounds included in the sound signal are biased or missing. As a result, the speaker model can be learned with a voice signal suitable for speaker recognition, and a voice processing device capable of suppressing a reduction in speaker recognition accuracy can be configured.

Fourth Embodiment
The configuration of the speech processing apparatus 400 according to the fourth embodiment will be described with reference to the drawings. FIG. 8 is a block diagram showing the configuration of the speech processing apparatus 400 in the fourth embodiment.

  The speech processing device 400 in the fourth embodiment includes an acoustic diversity calculation unit 12 and an acoustic reliability calculation unit 14. The acoustic diversity calculation unit 12 calculates an acoustic diversity representing the degree of variation regarding the type of the audio signal based on the audio signal representing the audio. The sound reliability calculation unit 14 calculates specific attribute information from the sound signal based on the degree of difference between the sound diversity of the sound signal calculated by the sound diversity calculation unit 12 and the sound diversity of another sound signal as a reference. The sound reliability is calculated to indicate the degree to which the audio signal is suitable for recognizing.

  The voice processing apparatus 400 having the above configuration calculates an acoustic diversity degree indicating the degree of variation regarding the type of the voice signal based on the voice signal representing the voice. The voice processing device 400 recognizes specific attribute information from the audio signal based on the degree of difference between the audio diversity of the audio signal calculated by the acoustic diversity calculation unit 12 and the audio diversity of another audio signal as a reference. The acoustic reliability that represents the degree to which the audio signal is suitable as the target is calculated. Therefore, the speech processing apparatus 400 can suppress the decrease in the accuracy of the speaker recognition by appropriately determining the reliability of the speaker recognition result.

  The sound diversity calculation unit 12 and the sound reliability calculation unit 14 in the first to third embodiments are the sound diversity calculation unit 12 and the sound reliability calculation unit 14 in the fourth embodiment, and are equivalent to each other. Including the features of

  As mentioned above, although this invention was demonstrated using embodiment, this invention is not limited to the said embodiment. 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 is needless to say that the present invention is not limited to the above embodiments, and various modifications are possible, which are also included in the scope of the present invention.

  Some or all of the above embodiments may be described as in the following appendices, but is not limited to the following.

[Supplementary Note 1]
Acoustic diversity calculation means for calculating an acoustic diversity degree representing a degree of variation regarding the type of sound of the audio signal based on an audio signal representing audio;
Indicates the degree to which the audio signal is suitable for recognizing specific attribute information from the audio signal based on the degree of difference between the acoustic diversity of the audio signal and the acoustic diversity of another audio signal as a reference And a sound reliability calculation means for calculating sound reliability.

[Supplementary Note 2]
The acoustic reliability calculation means
The difference between the acoustic diversity of the audio signal and the acoustic diversity of another audio signal is calculated based on a scale that represents a distance between two probability distributions. Voice processing device.

[Supplementary Note 3]
The acoustic diversity calculation means
The likelihood to the element distribution of the speech model of the speech signal is calculated based on the speech model that is configured as a mixed model in which the probability distribution is an element distribution and represents the appearance probability distribution of the speech signal, the likelihood or the likelihood The speech processing apparatus according to any one of the above 1 to 3, wherein the degree of acoustic multiplicity of the speech signal is calculated based on a value obtained by normalizing the degree with the sum of the likelihoods of all element distributions.

[Supplementary Note 4]
The acoustic reliability calculation means
The sound processing apparatus according to any one of appendices 1 to 3, wherein the sound reliability of the sound signal is calculated based on a signal obtained by dividing the sound signal.

[Supplementary Note 5]
The acoustic reliability calculation means
Supplementary note 1 to 4 wherein the degree of difference between the degree of acoustic diversity of the audio signal and the degree of acoustic diversity of the other audio signal is calculated based on the Kullback-Leibler information amount or the cosine similarity. A voice processing device according to any one of the preceding claims.

[Supplementary Note 6]
It further comprises an audio signal input request means for outputting to the outside request information indicating the degree to which an audio signal input is required,
The voice signal input request means
It is determined based on the sound reliability of the audio signal whether or not to output request information to the outside, and whether or not to stop input of the audio signal. Voice processing device.

[Supplementary Note 7]
Acoustic diversity degree calculation means for calculating an acoustic diversity degree that represents the degree of variation regarding the type of sound included in the audio signal, based on a tonality signal representing an audio;
An audio processing device comprising: attribute recognition means for recognizing specific attribute information from the audio signal when the acoustic diversity degree satisfies a predetermined condition.

[Supplementary Note 8]
Acoustic diversity degree calculation means for calculating an acoustic diversity degree that indicates the degree of variation regarding the type of sound included in the audio signal based on an audio signal representing an audio;
Indicates the degree to which the audio signal is suitable for recognizing specific attribute information from the audio signal based on the degree of difference between the acoustic diversity of the audio signal and the acoustic diversity of another audio signal as a reference Sound reliability calculation means for calculating sound reliability;
And an attribute recognition unit that recognizes specific attribute information from the voice signal;
The voice processing apparatus according to claim 1, wherein the attribute recognition unit starts attribute recognition processing when the acoustic reliability of the voice signal satisfies a predetermined condition.

[Supplementary Note 9]
Acoustic diversity degree calculation means for calculating an acoustic diversity degree that indicates the degree of variation regarding the type of sound included in the audio signal based on an audio signal representing an audio;
Indicates the degree to which the audio signal is suitable for recognizing specific attribute information from the audio signal based on the degree of difference between the acoustic diversity of the audio signal and the acoustic diversity of another audio signal as a reference Sound reliability calculation means for calculating sound reliability;
Attribute recognition means for recognizing specific attribute information from the voice signal;
A recognition result output unit that outputs a recognition result of the attribute recognition process;
The recognition result output unit
Supplementary note 8: The recognition result is calculated as a score that is numerical information indicating the matching degree of the voice signal with the voice model based on the sound reliability, or the recognition result is output. The voice processing device according to claim 1.

[Supplementary Note 10]
The voice processing apparatus according to claim 8 or 9, wherein the specific attribute information is information representing a speaker who issued a voice signal or a language that composes the voice signal.

[Supplementary Note 11]
Based on an audio signal representing an audio, an acoustic diversity representing the degree of variation regarding the type of sound included in the audio signal is calculated;
Indicates the degree to which the audio signal is suitable for recognizing specific attribute information from the audio signal based on the degree of difference between the acoustic diversity of the audio signal and the acoustic diversity of another audio signal as a reference A voice processing method characterized by calculating acoustic reliability.

[Supplementary Note 12]
The difference between the acoustic diversity of the audio signal and the acoustic diversity of another audio signal is calculated based on a scale that represents a distance between two probability distributions. Speech processing method.

[Supplementary Note 13]
The likelihood to the element distribution of the speech model of the speech signal is calculated based on the speech model that is configured as a mixed model in which the probability distribution is an element distribution and represents the appearance probability distribution of the speech signal, the likelihood or the likelihood 11. The speech processing method according to appendix 11 or 12, wherein an acoustic diversity of the speech signal is calculated based on a value obtained by normalizing the degree with the sum of the likelihoods of all element distributions.

[Supplementary Note 14]
The sound processing method according to any one of appendices 11 to 13, wherein the sound reliability of the sound signal is calculated based on a signal obtained by dividing the sound signal.

[Supplementary Note 15]
The difference degree between the acoustic diversity degree of the audio signal and the acoustic diversity degree of the other audio signal is calculated based on the Kullback-Leibler information amount or the cosine similarity. The voice processing method according to any one of the above.

[Supplementary Note 16]
Output request information representing the degree to which audio signal input is required,
It is determined based on the sound reliability of the audio signal whether or not to output request information to the outside, and whether or not to stop input of the audio signal. Voice processing method.

[Supplementary Note 17]
Based on a tonality signal representing speech, an acoustic diversity indicating the degree of variation regarding the type of sound included in the speech signal is calculated;
Recognizing specific attribute information from the audio signal when the acoustic diversity satisfies a predetermined condition;
A voice processing device characterized by

[Supplementary Note 18]
Based on an audio signal representing an audio, an acoustic diversity representing the degree of variation regarding the type of sound included in the audio signal is calculated;
Indicates the degree to which the audio signal is suitable for recognizing specific attribute information from the audio signal based on the degree of difference between the acoustic diversity of the audio signal and the acoustic diversity of another audio signal as a reference Calculate acoustic reliability,
Recognize specific attribute information from the voice signal,
A voice processing method comprising: determining whether to start attribute recognition processing when the acoustic reliability of the voice signal satisfies a predetermined condition.

[Supplementary Note 19]
Based on an audio signal representing an audio, an acoustic diversity representing the degree of variation regarding the type of sound included in the audio signal is calculated;
Indicates the degree to which the audio signal is suitable for recognizing specific attribute information from the audio signal based on the degree of difference between the acoustic diversity of the audio signal and the acoustic diversity of another audio signal as a reference Calculate acoustic reliability,
Recognize specific attribute information from the voice signal,
Outputting the recognition result of the attribute recognition process;
The recognition result is calculated as a score which is numerical information indicating the matching degree of the voice signal with the voice model based on the sound reliability, or the recognition result is output. The voice processing method described in.

[Supplementary Note 20]
The voice processing method according to appendix 18 or 19, wherein the specific attribute information is information representing a speaker who issued a voice signal or a language constituting the voice signal.

[Supplementary Note 21]
Acoustic diversity calculation processing for calculating acoustic diversity representing the degree of variation regarding the type of sound included in the audio signal based on an audio signal representing audio;
Indicates the degree to which the audio signal is suitable for recognizing specific attribute information from the audio signal based on the degree of difference between the acoustic diversity of the audio signal and the acoustic diversity of another audio signal as a reference A program for causing a computer to execute sound reliability calculation processing for calculating sound reliability.

[Supplementary Note 22]
The sound reliability calculation process is
The difference between the acoustic diversity of the audio signal and the acoustic diversity of another audio signal is calculated based on a scale that represents a distance between two probability distributions. program.

[Supplementary Note 23]
The acoustic diversity calculation process
The likelihood to the element distribution of the speech model of the speech signal is calculated based on the speech model that is configured as a mixed model in which the probability distribution is an element distribution and represents the appearance probability distribution of the speech signal, the likelihood or the likelihood The program according to Appendix 21 or 22, wherein the degree of acoustic diversity of the audio signal is calculated based on a value obtained by normalizing the degree with the sum of the likelihoods of all element distributions.

[Supplementary Note 24]
The sound reliability calculation process is
24. The program according to any one of appendices 21 to 23, wherein acoustic reliability of the audio signal is calculated based on a signal obtained by dividing the audio signal.

[Supplementary Note 25]
The sound reliability calculation process is
The difference between the acoustic diversity of the audio signal and the acoustic diversity of the other audio signal is calculated based on the Kullback-Leibler information amount or the cosine similarity. The program described in any one.

[Supplementary Note 26]
The system further comprises an audio signal input request process of externally outputting request information indicating the degree to which the audio signal input is required,
The voice signal input request process is
It is determined based on the sound reliability of the audio signal whether or not to output request information to the outside, and whether or not to stop input of the audio signal. Programs.

[Supplementary Note 27]
Acoustic diversity calculation processing for calculating acoustic diversity representing the degree of variation regarding the type of sound included in the audio signal based on a sound signal representing audio;
An attribute recognition process for recognizing specific attribute information from the audio signal when the acoustic diversity degree satisfies a predetermined condition;
A voice processing apparatus characterized by causing a computer to execute.

[Supplementary Note 28]
Acoustic diversity calculation processing for calculating acoustic diversity representing the degree of variation regarding the type of sound included in the audio signal based on an audio signal representing audio;
Indicates the degree to which the audio signal is suitable for recognizing specific attribute information from the audio signal based on the degree of difference between the acoustic diversity of the audio signal and the acoustic diversity of another audio signal as a reference Sound reliability calculation processing for calculating sound reliability;
And attribute recognition processing for recognizing specific attribute information from the voice signal,
The program is characterized in that the attribute recognition process causes a computer to start the attribute recognition process when the sound reliability of the audio signal satisfies a predetermined condition.

[Supplementary Note 29]
Acoustic diversity calculation processing for calculating acoustic diversity representing the degree of variation regarding the type of sound included in the audio signal based on an audio signal representing audio;
Indicates the degree to which the audio signal is suitable for recognizing specific attribute information from the audio signal based on the degree of difference between the acoustic diversity of the audio signal and the acoustic diversity of another audio signal as a reference Sound reliability calculation processing for calculating sound reliability;
Attribute recognition processing for recognizing specific attribute information from the voice signal;
It further comprises a recognition result output process for outputting a recognition result of the attribute recognition process,
The recognition result output process is
The recognition result is calculated as a score, which is numerical information indicating the matching degree of the voice signal with the voice model, based on the sound reliability, or the recognition result is output. The program described in.

[Supplementary Note 30]
The program according to Appendix 28 or 29, wherein the specific attribute information is information representing a speaker who issued an audio signal or a language that configures the audio signal.

  In addition, when acquiring and using the information regarding a user in each aspect of this invention, this shall be performed legally.

DESCRIPTION OF SYMBOLS 10 information processing apparatus 11 speech model storage part 12 acoustic diversity degree calculation part 13 acoustic diversity degree storage part 14 acoustic reliability calculation part 21 speech area detection part 22 speech processing part 23 speaker model storage part 24 speaker recognition calculation part 25 talk Speaker recognition output unit 31 voice section detection unit 32 voice processing unit 33 voice model storage unit 34 speaker model storage unit 35 speaker model learning unit 36 voice signal input request unit 41 acoustic diversity calculation unit 42 acoustic reliability calculation unit 100 voice Processing unit 200 Speaker recognition unit 300 Speaker model learning unit 400 Speech processing unit

Claims (10)

Acoustic diversity degree calculation means for calculating an acoustic diversity degree that indicates the degree of variation regarding the type of sound included in the audio signal based on an audio signal representing an audio;
Based on the difference between the acoustic diversity of the voice signal and the acoustic diversity calculated from other voice signals used for learning a speaker model for recognizing specific attribute information , An audio reliability calculation means for calculating an audio reliability representing the degree to which the audio signal is suitable for recognizing attribute information. The acoustic reliability calculation means
And the acoustic diversity of the audio signal, the degree of difference between the acoustic diversity of the other audio signals, be calculated based on the measure of the distance between two probability distributions to claim 1, wherein A voice processing device as described. The acoustic diversity calculation means
The likelihood to the element distribution of the speech model of the speech signal is calculated based on the speech model that is configured as a mixed model in which the probability distribution is an element distribution and represents the appearance probability distribution of the speech signal, the likelihood or the likelihood The speech processing apparatus according to claim 1 or 2, wherein an acoustic diversity degree of the speech signal is calculated based on a value obtained by normalizing the degree with the sum of the likelihoods of all element distributions. The audio signal is segmented at a predetermined time,
The acoustic reliability calculation means
The audio processing apparatus according to any one of claims 1 to 3, wherein an acoustic reliability is calculated for each of the divided audio signals. The acoustic reliability calculation means
A difference degree between the acoustic diversity degree of the audio signal and the acoustic diversity degree of the other audio signal is calculated based on the Kullback-Leibler information amount or the cosine similarity. The voice processing device according to any one of 4. It further comprises an audio signal input request means for outputting to the outside request information indicating the degree to which an audio signal input is required,
The voice signal input request means
The voice processing device according to any one of claims 1 to 5, wherein it is determined whether or not to output request information to the outside based on the sound reliability of the voice signal. Acoustic diversity degree calculation means for calculating an acoustic diversity degree that indicates the degree of variation regarding the type of sound included in the audio signal based on an audio signal representing an audio;
Based on the difference between the acoustic diversity of the voice signal and the acoustic diversity calculated from other voice signals used for learning a speaker model for recognizing specific attribute information , Acoustic reliability calculation means for calculating an acoustic reliability that represents the degree to which the audio signal is suitable for recognizing attribute information;
Attribute recognition means for recognizing specific attribute information from the voice signal;
A recognition result output unit that outputs a recognition result of the attribute recognition unit ;
The recognition result output unit
Outputting the recognition result calculated as a score representing the degree of matching of the audio signal with specific attribute information based on the sound reliability , or outputting the recognition result according to the sound reliability A voice processing device characterized by Acoustic diversity degree calculation means for calculating an acoustic diversity degree that indicates the degree of variation regarding the type of sound included in the audio signal based on an audio signal representing an audio;
Identification from the audio signal based on the degree of difference between the acoustic diversity of the audio signal and the acoustic diversity calculated from other audio signals used for learning a speaker model for recognizing specific attribute information Acoustic reliability calculation means for calculating an acoustic reliability that represents the degree to which the audio signal is suitable for recognizing the attribute information of
And an attribute recognition unit that recognizes specific attribute information from the voice signal;
The voice processing apparatus according to claim 1, wherein the attribute recognition unit starts attribute recognition processing when the acoustic reliability of the voice signal satisfies a predetermined condition. Based on an audio signal representing an audio, an acoustic diversity representing the degree of variation regarding the type of sound included in the audio signal is calculated;
Based on the difference between the acoustic diversity of the voice signal and the acoustic diversity calculated from other voice signals used for learning a speaker model for recognizing specific attribute information , A voice processing method, comprising: calculating an acoustic reliability indicating a degree to which the voice signal is suitable as a target for recognizing attribute information. Acoustic diversity calculation processing for calculating acoustic diversity representing the degree of variation regarding the type of sound included in the audio signal based on an audio signal representing audio;
Based on the difference between the acoustic diversity of the voice signal and the acoustic diversity calculated from other voice signals used for learning a speaker model for recognizing specific attribute information , A program for causing a computer to execute sound reliability calculation processing for calculating sound reliability indicating the degree to which the audio signal is suitable for recognizing attribute information.

Images