JP2017187686A - Voice selection device and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To select, from a plurality of complementary voices, a complementary voice to be added to a program voice when it is presented, the complementary voice distinguishable from the program voice.SOLUTION: A feature amount calculation unit 11-n of a voice selection device 1 calculates an acoustic feature amount by reading voice waveform data of a program voice from a program voice DB 10-n. A feature amount calculation unit 21-m calculates an acoustic feature amount by reading voice waveform data of a complementary voice from a complementary voice DB 20-m. A similarity calculation unit 22-m calculates similarities between 1st-Nth program voices and an mth complementary voice. A similarity addition unit 23-m calculates the sum total of the similarities between the mth complementary voice and the program voices of 1st-Nth program speakers. A selection unit 24 specifies the sum total of smallest similarities in the sum total of similarities for every 1st-Mth complementary voices, and selects a complementary voice DB 20 corresponding to the sum total of smallest similarities.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a sound selection device and a program for selecting commentary sound to be added to program sound from a plurality of commentary sounds when commentary sound is added to program sound to generate commentary-added program sound.

  2. Description of the Related Art Conventionally, in the production of commentary broadcast programs in television broadcasts, a commentary manuscript about the scene description or caption content for the visually impaired is created separately from the script or script of the program. The explanation manuscript is a manuscript that the narrator reads out as a commentary voice in a section (pause section) of silence or background sound only so as not to overlap with a section of speech (program sound) including speech sounds such as dialogue or narration.

  At the time of recording the commentary voice, it is necessary to adjust the voice start timing and voice speed, which requires much time and cost including rehearsal. In order to solve this problem, a technique for producing audio of a commentary broadcast program in a short time and at a low cost is disclosed (for example, see Patent Document 1).

  In this technology, program sound and text related to the contents of the program are input, and commentary sound is generated from the text by speech synthesis. Then, the pause section is detected from the program voice, the commentary voice is converted to the speech speed so as to match the pause section length, and the commentary voice after the speech speed conversion is added to the pause section.

Japanese Patent No. 4594908

  However, in the technique of Patent Document 1, there is a case where the pause section cannot be correctly detected from the program sound, and the commentary sound cannot be inserted at a suitable timing and speaking speed, and as a result, a suitable commentary sound is provided. There was a problem that could not.

  In order to solve this problem, it is assumed that program audio with commentary in a state where program audio and commentary audio are superimposed is generated. However, when the program sound and the commentary sound are similar, it is difficult to distinguish the commentary sound information from the generated commentary program sound.

  In this way, it is possible to distinguish TV broadcast program audio when commentary audio for supplementing the information of the program audio (hereinafter referred to as complementary audio) is added to generate program audio with explanation. There is a problem in that it may not be possible to provide appropriate supplementary speech properly. No method has been proposed to solve this problem.

  Accordingly, the present invention has been made to solve the above-mentioned problems, and the object thereof is complementary speech when a supplementary audio is added to a program audio for presentation, and the complementary audio is easy to distinguish from the program audio. Is to provide a voice selection device and program capable of selecting from a plurality of complementary voices.

  In order to solve the above-mentioned problem, an audio selection device according to claim 1 is an audio selection device that selects a complementary audio from a plurality of complementary audios when adding a complementary audio to a program audio for presentation. A program audio DB (database) storing a predetermined number of program audio data, a complementary audio DB storing two or more predetermined numbers of complementary audio data, and the predetermined number of program audios stored in the program audio DB A feature amount calculation unit that calculates an acoustic feature amount for each of the data, and calculates an acoustic feature amount for each of the predetermined number of complementary speech data stored in the complementary speech DB, and the feature amount calculation unit. About each of the calculated acoustic feature amount for each of the predetermined number of program audio data and each of the predetermined number of complementary audio data calculated by the feature amount calculation unit. A similarity calculation unit that calculates a similarity between the audio feature quantity and an acoustic feature quantity for each of the predetermined number of program audio data calculated by the similarity calculation unit for each of the complementary audio data And the similarity between the acoustic feature amount of the complementary speech data and a similarity addition unit for obtaining a sum, and a minimum sum among the sums for each of the complementary speech data obtained by the similarity addition unit And a selection unit that selects the complementary audio data corresponding to the minimum sum from the predetermined number of complementary audio data.

  In addition, in the audio selection device according to claim 2, in the audio selection device according to claim 1, the feature amount calculation unit performs frame units of a predetermined length for each of the program audio data and the complementary audio data. Voice data is cut out, a frequency characteristic is obtained for each voice data of the frame unit, and based on the frequency characteristic, a static coefficient composed of a Mel frequency cepstrum coefficient and logarithmic energy, and a primary regression coefficient of the static coefficient and 2 A spectral feature amount including a next regression coefficient is obtained, and an EM algorithm is used based on the spectral feature amount to calculate a GMM parameter composed of a mixture weight for the number of mixtures and a Gaussian distribution for the number of mixtures, and from the GMM parameter An average vector of the Gaussian distribution is extracted, and the average vector is combined by the number of the mixtures. Seek vector, on the basis of the GMM supervectors, calculates the i vector which is the acoustic feature, and wherein the.

  According to a third aspect of the present invention, there is provided the voice selecting device according to the first aspect, wherein the feature amount calculation unit is configured to perform frame units of a predetermined length for each of the program voice data and the complementary voice data. Cut out audio data, set basic period candidates for each frame-based audio data, determine the degree of periodicity of the basic period candidates, extract the basic period from the basic period candidates, and based on the basic period , A logarithmic fundamental frequency and a pitch feature amount including a primary regression coefficient and a quadratic regression coefficient of the logarithmic fundamental frequency are obtained, and an EM algorithm is used based on the pitch feature amount to mix the mixture weight and the mixture number. A GMM parameter composed of a Gaussian distribution of minutes, an average vector of the Gaussian distribution is extracted from the GMM parameter, and the average vector is extracted from the mixed vector. Seeking GMM supervector bound for the number of, on the basis of the GMM supervectors, calculates the i vector which is the acoustic feature, and wherein the.

  According to a fourth aspect of the present invention, there is provided the voice selecting device according to the first aspect, wherein the feature amount calculation unit is configured to perform frame units of a predetermined length for each of the program voice data and the complementary voice data. Voice data is cut out, a frequency characteristic is obtained for each voice data of the frame unit, and based on the frequency characteristic, a static coefficient composed of a Mel frequency cepstrum coefficient and logarithmic energy, and a primary regression coefficient of the static coefficient and 2 A spectral feature amount including a next regression coefficient is obtained, and an EM algorithm is used based on the spectral feature amount to calculate a GMM parameter composed of a mixture weight for the number of mixtures and a Gaussian distribution for the number of mixtures, and from the GMM parameter An average vector of the Gaussian distribution is extracted, and the average vector is combined by the number of the mixtures. A vector is obtained, a first i vector that is the acoustic feature amount is calculated based on the GMM super vector, a basic period candidate is set for each frame-based audio data, and the periodicity of the basic period candidate is determined. A basic period is extracted from the basic period candidates, and a pitch feature amount including a logarithmic fundamental frequency and a primary regression coefficient and a quadratic regression coefficient of the logarithmic fundamental frequency is obtained based on the fundamental period, Based on the pitch feature amount, an EM algorithm is used to calculate a GMM parameter composed of a mixture weight for the number of mixtures and a Gaussian distribution for the number of mixtures, and an average vector of the Gaussian distribution is extracted from the GMM parameter, A GMM supervector obtained by combining vectors by the number of the mixture is obtained, and the acoustic characteristics are obtained based on the GMM supervector. A second i vector as a quantity is calculated, and the similarity calculation unit calculates a first i vector for each of the predetermined number of program audio data calculated by the feature quantity calculation unit and the feature quantity calculation. A similarity is calculated between each of the predetermined number of complementary audio data calculated by the first i-vector and each of the predetermined number of program audio data calculated by the feature amount calculating unit. A similarity between the second i vector of the second i vector and the second i vector for each of the predetermined number of complementary speech data calculated by the feature amount calculation unit, and the similarity addition unit includes: For each complementary audio data, a first i vector for each of the predetermined number of program audio data and a first i vector of the complementary audio data calculated by the similarity calculation unit The second i for each of the predetermined number of program audio data calculated by the similarity calculator for each of the complementary audio data is obtained by adding the similarity between Adding the similarity between the vector and the second i vector of the complementary speech data, obtaining a second addition result, weighting and adding the first addition result and the second addition result, It is characterized in that the sum is obtained.

  Furthermore, a voice selection program according to a fifth aspect causes a computer to function as the voice selection device according to any one of the first to fourth aspects.

  As described above, according to the present invention, it is possible to select, from a plurality of complementary voices, a complementary voice that is easy to distinguish from the program voice when the supplementary voice is added to the program voice for presentation. It becomes possible. Therefore, even when the selected complementary sound is added to the program sound and the program sound and the complementary sound are presented at the same timing, a person who listens to these sounds can easily distinguish the program sound and the complementary sound. And supplementary speech that is easy to distinguish.

It is a block diagram which shows the structural example of the audio | voice selection apparatus by embodiment of this invention. 6 is a flowchart illustrating a processing example of a feature amount calculation unit according to the first embodiment. It is a figure explaining the GMM parameter (lambda) calculated by the process of step S213. It is a figure explaining the GMM super vector M calculated by the process of step S214. 10 is a flowchart illustrating a processing example of a feature amount calculation unit according to the second embodiment. 10 is a flowchart illustrating a processing example of speech frame section determination as pre-processing of processing by the feature amount calculation unit according to the second embodiment. It is a figure explaining the example which calculates | requires the basic period of a silent interval and an unvoiced sound area from the basic period of the front and back voiced sound area.

  Hereinafter, embodiments for carrying out the present invention will be described in detail with reference to the drawings. The present invention calculates acoustic feature quantities of one or more program sounds and two or more complementary sounds, calculates a similarity between one or more program sounds for each of the two or more complementary sounds, The complementary speech having the lowest similarity is selected from two or more complementary speeches.

  As a result, a complementary sound having an acoustic feature that is not similar to the program sound is selected. Therefore, even when the program sound and the complementary sound are presented at the same time, a person who listens to these sounds can easily distinguish between the program sound and the complementary sound, and can obtain a complementary sound that is easy to distinguish. .

[Voice selection device]
First, a voice selection device according to an embodiment of the present invention will be described. FIG. 1 is a block diagram illustrating a configuration example of a voice selection device according to an embodiment of the present invention. The audio selection device 1 includes program audio DBs (databases) 10-1 to 10-N, feature amount calculation units 11-1 to 11-N, complementary audio DBs 20-1 to 20-M, and a feature amount calculation unit 21-1. 21-M, similarity calculation units 22-1 to 22-M, similarity addition units 23-1 to 23-M, and a selection unit 24.

  N is an integer equal to or greater than 1, and corresponds to the number of speakers (program audio speakers) about the program audio stored in the program audio DBs 10-1 to 10-N. M is an integer equal to or greater than 2, and corresponds to the number of speakers (complementary speech speakers) regarding the supplementary speech stored in the supplementary speech DBs 20-1 to 20-M. Let n = 1,..., N and m = 1,.

  The program audio DB 10-n is a database in which audio waveform data (program audio data) of program audio by a program audio speaker is stored. The audio waveform data of the program audio is sampled at a sampling frequency of 16 kHz and a conversion bit number of 16 bits.

  The feature amount calculation unit 11-n reads out the audio waveform data of the program audio by the nth program audio speaker from the corresponding program audio DB 10-n, and program audio based on the audio waveform data of the program audio The acoustic feature amount (acoustic feature amount) is calculated. Then, the feature amount calculation unit 11-n outputs the acoustic feature amount of the program sound by the nth program sound speaker to the similarity calculation units 22-1 to 22-M.

  The complementary speech DB 20-m is a database in which speech waveform data (complementary speech data) of complementary speech by a certain complementary speech speaker is stored. The audio waveform data of the complementary audio is sampled at a sampling frequency of 16 kHz and a conversion bit number of 16 bits, similarly to the audio waveform data of the program audio. The audio waveform data of the complementary audio may be, for example, actual audio data recorded for addition to the program audio, or audio data created by voice synthesis (not actual complementary audio data). Or voice data included in a voice database used for voice synthesis.

  The feature amount calculating unit 21-m reads out the speech waveform data of the supplementary speech by the m-th supplementary speech speaker from the corresponding supplementary speech DB 20-m, and based on the speech waveform data of the supplementary speech, Is calculated. Then, the feature amount calculation unit 21-m outputs the acoustic feature amount of the complementary speech by the m-th complementary speech speaker to the corresponding similarity calculation unit 22-m.

  The similarity calculation unit 22-m inputs the acoustic feature amount of the program sound by the first to Nth program sound speakers from the feature amount calculation units 11-1 to 11-N, and the corresponding feature amount calculation unit 21. The acoustic feature amount of the complementary speech by the m-th complementary speech speaker from -m is input.

  The similarity calculation unit 22-m calculates the similarity between the acoustic feature quantity of the program sound by the first program voice speaker and the acoustic feature quantity of the complementary speech by the m-th complementary speech speaker. . Similarly, the similarity calculation unit 22-m similarly uses the acoustic feature amount of the program sound by the 2nd to Nth program speech speakers and the acoustic feature amount of the supplementary speech by the mth complementary speech speaker. The similarity between is calculated. Then, the similarity calculation unit 22-m determines the respective similarities between the program audio by the first to Nth program audio speakers and the complementary audio by the mth complementary audio speaker, and the corresponding similarity. The result is output to the adder 23-m.

前記式（１）の右辺の分子は、ｗ_in及びｗ_cmの内積を示し、その分母は、ｗ_in及びｗ_cmにおけるそれぞれのノルムの乗算を示す。 Here, the acoustic features of a program audio according to the n-th program audio speaker and w _in the acoustic features of complementary speech in the m-th complementary audio speaker and w _cm, similarity cosine similarity cos (W _in , w _cm ). The cosine similarity cos (w _in , w _cm ) between the program audio from the nth program audio speaker and the complementary audio from the mth complementary audio speaker is calculated by the following equation.

Molecule on the right side of the equation (1) represents the inner product of w _in and w _cm, the denominator indicates the multiplication of the respective norms in w _in and w _cm.

  The similarity adding unit 23-m receives each of the program audio from the first to Nth program audio speakers and the complementary audio from the mth complementary audio speaker from the corresponding similarity calculation unit 22-m. Enter the similarity. And the similarity addition part 23-m calculates | requires the sum total of a similarity degree by adding each similarity degree about the complementary audio | voice by the mth complementary audio | voice speaker. The similarity adding unit 23-m adds the sum of similarities between the supplementary speech by the mth complementary speech speaker and the program speech by the first to Nth program speech speakers (the mth supplementary speech). The sum of the similarities of complementary speech by the speaker is output to the selection unit 24.

Here, the similarity cos (w _in, w _cm) for complementing the voice according to the m-th complementary audio speaker when the sum of the s _m, the sum s _m is calculated by the following equation.

  The selection unit 24 inputs the sum of the similarities from the similarity adding units 23-1 to 23 -M, and specifies the minimum sum of the similarities among the sums of these similarities. Then, the selection unit 24 selects the complementary speech DB 20 (complementary speech speaker) corresponding to the sum of the minimum similarities among the supplementary speech DBs 20-1 to 20-M (among M supplementary speech speakers). And select information is output.

Here, minimum and complementary speech DB20 correspond to the similarity of the sum s _m (complementary audio speakers) complementary voice DB20-c (complementary audio speaker c), any selection information c (of 1~M The selection information c is selected by the following formula.

  As described above, according to the audio selection device 1 of the embodiment of the present invention, the selection unit 24 includes the program audio and the audio of the complementary audio DBs 20-1 to 20 -M (among M complementary audio speakers). Selects the complementary speech DB 20-c (complementary speech speaker c) having the most similar acoustic features. The selected complementary audio DB 20-c is used when generating the program audio with explanation by adding the complementary audio to the program audio. As a result, even if the program sound and the complementary sound are presented at the same timing as a result of adding the complementary sound to the program sound, a person who listens to these sounds can easily distinguish the program sound and the complementary sound. And supplementary speech that is easy to distinguish.

  Hereinafter, the voice selection device 1 according to the embodiment of the present invention will be specifically described with reference to Examples 1 to 3. The feature quantity calculation units 11-1 to 11-N and 21-1 to 21-M are collectively referred to as feature quantity calculation units 11 and 21.

In the first to third embodiments, as a process in which the feature amount calculation units 11 and 21 calculate the acoustic feature amount, an i-vector technique used in speaker recognition or speaker verification is used. For details of i-vector, refer to the following documents.
[Non-Patent Document 1]
N. Dehak, P. Kenny, R. Dehak, P. Dumouchel and P. Ouellet, “Front-end factor analysis for speaker verification”, IEEE Trans. Audio Speech Lang. Process., 19, 788-798 (2011)

[Example 1]
First, Example 1 will be described. The first embodiment is an example in which complementary audio that is easy to distinguish from program audio is selected from the viewpoint of voice quality. Specifically, in the first embodiment, the sound using the static coefficient including the mel frequency cepstrum coefficient (MFCC) and the logarithmic energy (E) and the spectral feature amount including the primary regression coefficient and the secondary regression coefficient thereof. One complementary voice is selected from a plurality of complementary voices based on the feature amount.

The feature quantity calculation units 11 and 21 combine the average vectors constituting the mixed Gaussian distribution model (GMM) based on the spectral feature quantity as many as the number of mixtures as the acoustic feature quantity, obtain the GMM super vector, and calculate the i vector. For the calculation method of the spectral feature amount, refer to the following documents.
[Non-Patent Document 2]
The HTK Book (for HTK Version 3.4) Cambridge University Engineering Department

  FIG. 2 is a flowchart illustrating a processing example of the feature amount calculation units 11 and 21 according to the first embodiment. The feature quantity calculators 11 and 21 read the speaker's voice waveform data from the program voice DB 10 or the complementary voice DB 20 (step S201), and the voice data (speech) of the frame having a window width of 25 ms and a shift width of 10 ms from the voice waveform data. A frame is cut out (step S202).

  The feature quantity calculation units 11 and 21 perform high frequency emphasis (pre-emphasis) on the speech frame with a pre-emphasis coefficient of 0.97 (step S203). Then, the feature amount calculation units 11 and 21 multiply the speech frame after high frequency emphasis by a window function of a Hamming window having a window width of 25 ms (step S204), and perform a discrete Fourier transform (FFT) with 1024 FFT points. The frequency characteristic is obtained (step S205).

  The feature quantity calculation units 11 and 21 obtain 26-channel filter bank coefficients by multiplying the frequency characteristics by the mel filter bank (step S206). And the feature-value calculation parts 11 and 21 calculate a 12-dimensional mel frequency cepstrum coefficient (MFCC) by performing a discrete cosine transform (DCT) with respect to a filter bank coefficient (step S207).

  The feature amount calculation units 11 and 21 shift from step S202 to calculate logarithmic energy (E) for the audio frame (step S208).

The feature quantity calculation units 11 and 21 set a 13-dimensional static coefficient obtained by combining the 12-dimensional mel frequency cepstrum coefficient (MFCC) and the logarithmic energy (E) (step S209). Then, the feature quantity calculation units 11 and 21 calculate, for these static coefficients, primary differences ΔMFCC, ΔE that are primary regression coefficients and secondary differences Δ ² MFCC, Δ ² E that are secondary regression coefficients. (Step S210, Step S211). The feature quantity calculation units 11 and 21 set the mel frequency cepstrum coefficient (MFCC), logarithmic energy (E), primary difference ΔMFCC, ΔE, and secondary difference Δ ² MFCC, secondary difference Δ ² E as spectral feature quantities. (Step S212).

Thus, for each voice frame, 12 mel frequency cepstrum coefficients (MFCC), 1 logarithmic energy (E), 12 primary differences ΔMFCC, 1 primary difference ΔE, 12 secondary differences A spectral feature quantity consisting of Δ ² MFCC and D _F (= 39) coefficients that are one secondary difference Δ ² E is obtained.

The feature quantity calculators 11 and 21 use the EM (Expectation Maximization) algorithm to calculate the speech from the spectrum feature quantity (coefficients in all voice frames) composed of D _F (= 39) coefficients calculated for each voice frame. The GMM parameter λ relating to the entire data of the person's voice waveform is calculated (step S213). For details of the method of calculating the GMM parameter λ using the EM algorithm, refer to the following documents.
[Non-Patent Document 3]
REFERENCE MANUAL for Speech Signal Processing Toolkit Ver. 3.9

The GMM parameter λ is constituted by a mixture weight C (= 512) mixture weights and a mixture number C Gaussian distribution as shown in the following equation. Let the mixing weight be W. Gaussian distribution is represented by the mean vector mu, and the vector sigma ² consisting of D _F-number of variance values consisting of D _F number of the mean.

FIG. 3 is a diagram for explaining the GMM parameter λ calculated by the process of step S213. As described above, the GMM parameter λ is calculated from the spectral feature amount (coefficient in all voice frames) including _DF (= 39) coefficients for each voice frame using the EM algorithm in the process of step S213. The

As shown in FIG. 3, the GMM parameter λ is composed of a mixture weight W (0) and a Gaussian distribution for the 0th in the number of mixture C. In this case, the Gaussian distribution has an average vector μ ₀ (0),..., Μ ₀ (D _F −1) composed of _DF average values, and a vector σ ₀ ² composed of _DF variance values ( 0),..., Σ ₀ ² (D _F −1).

Similarly, the GMM parameter λ is composed of a mixture weight W (C−1) and a Gaussian distribution for the (C−1) th in the number of mixture C. In this case, the Gaussian distribution has an average vector μ _C-1 (0),..., Μ _C-1 (D _F −1) consisting of _DF average values, and a vector consisting of _DF variance values. σ _C-1 ² (0),..., σ _C-1 ² (D _F −1).

Returning to FIG. 2, the feature quantity calculation units 11 and 21 obtain the GMM super vector M from the GMM parameter λ after step S213 (step S214). Specifically, the feature amount calculation unit 11 and 21, the number of mixture C number mixture weight and number of mixture C-number of Gaussian distribution (D _F mean vector of pieces of the mean value mu, and D _F-number of variance Only the average vector μ is extracted by the GMM parameter λ composed of the following vector σ ² ). Then, the feature amount calculation units 11 and 21 combine the average vector μ composed of the _DF average values by the number C of mixtures to obtain the GMM super vector M. The GMM super vector M is a C · D _F dimensional real vector and is expressed as follows.

FIG. 4 is a diagram for explaining the GMM super vector M calculated by the process of step S214. As shown in FIG. 4, GMM supervector M is the mean vector mu ₀ consist _{D F-number} of the mean value for the 0th _{(0), ···, μ 0} (D F -1), ··· and, the (C-1) th mean vectors μ _C-1 (0) consisting of _{D F-number} of mean values for, ..., constituted by _{μ C-1 (D F -1} ).

Returning to FIG. 2, after step S214, the feature quantity calculation units 11 and 21 use the technique described in Non-Patent Document 1 described above, based on the GMM supervector M, and satisfy the following equation. The i vector as a quantity: w is calculated (step S215).

The i vector: w is a D _T- dimensional real vector and is expressed as follows.

Here, m is a GMM super vector learned using a large amount of unspecified speaker's speech data, and T is a low-rank rectangular matrix (D _T << C · D _F ). The rectangular matrix T is a C · D _F × D _T- dimensional real vector and is expressed as follows.

共分散行列Ｉは、Ｄ_T×Ｄ_T次元の実数のベクトルであり、以下のように表される。

w follows a Gaussian distribution N (w; 0, I) in which the mean vector is 0 and the covariance matrix is the unit matrix I. The average vector 0 is a _DT- dimensional real vector and is expressed as follows.

The covariance matrix I is a D _T × D _T dimensional real vector and is expressed as follows.

  Note that the feature quantity calculation units 11 and 21 perform acoustic fluctuations within the same speaker by processing such as LDA (Linear Discrimination Analysis) and WCCN (Within-Class Covariance Normalization) on the calculated i vector: w. to correct. The same applies to Examples 2 and 3 described later.

  The processes of the similarity calculation units 22-1 to 22-M, the similarity addition units 23-1 to 23-M, and the selection unit 24 are the same as those in FIG.

  As described above, the feature amount calculation units 11 and 21 according to the first embodiment configure a mixed Gaussian distribution model (GMM) based on spectral feature amounts with respect to audio waveform data of audio read from the program audio DB 10 or the complementary audio DB 20. The GMM super vector M is obtained by combining the average vectors μ by the number of mixtures C. Then, the feature amount calculation units 11 and 21 calculate an i vector that is an acoustic feature amount using the spectrum feature amount based on the GMM super vector M.

  The selection unit 24 at the latter stage is based on the i vector calculated by the feature quantity calculation units 11 and 21, and the program audio, among the complementary audio DBs 20-1 to 20 -M (among M complementary audio speakers), A complementary speech DB 20-c (complementary speech speaker c) having an acoustic feature that is least similar is selected.

  Here, the complementary speech DB 20-c (complementary speech speaker c) is selected with the acoustic feature amount calculated from the spectral feature amount as an index, and the frequency feature of the speech is reflected in the spectral feature amount. Voice quality is determined by the frequency component of the voice.

  Therefore, even if the program sound and the complementary sound are simultaneously presented as a result of adding the complementary sound to the program sound, the person who listens to these sounds can easily distinguish the program sound and the complementary sound. Therefore, it is possible to obtain complementary speech that can easily distinguish the voice quality of the speaker.

[Example 2]
Next, Example 2 will be described. The second embodiment is an example in which a complementary voice that is easy to distinguish from a program voice is selected from the viewpoint of loudness. Specifically, in the second embodiment, based on the logarithmic fundamental frequency (LF0) and the acoustic feature amount using the pitch feature amount including the primary regression coefficient and the quadratic regression coefficient, one of the complementary speeches is used. Select complementary audio.

The feature quantity calculation units 11 and 21 combine the average vectors constituting the mixed Gaussian distribution model (GMM) based on the pitch feature quantity as many as the number of mixtures as the acoustic feature quantity to obtain the GMM super vector, and calculate the i vector. For the calculation method of the pitch feature value, refer to the following document.
[Non-Patent Document 4]
Tsuki, Kiyoyama, Miyasaka, “Speech Basic Period Extraction Method Using Autocorrelation Function Obtained from Multiple Window Widths”, IEICE Transactions A Vol, J80-A No.9 pp.1341-1350 1997 September [Non-Patent Document 5]
Kiyoyama, Imai, Mishima, Miyagi, Miyasaka, “Development of high-quality real-time speech rate conversion system”, IEICE Transactions D-II Vol, J84-D-II No.6 pp.918-926 2001 6 Moon

  FIG. 5 is a flowchart illustrating a processing example of the feature amount calculation units 11 and 21 according to the second embodiment. The feature quantity calculators 11 and 21 read voice waveform data of the voice from the program voice DB 10 or the complementary voice DB 20 (step S501). Then, the feature amount calculation units 11 and 21 perform low-pass filtering on the voice waveform data at a cutoff frequency of 1 kHz and perform ¼ decimation (step S502). Then, the feature amount calculators 11 and 21 cut out voice data (voice frames) of a voice waveform frame with a predetermined window width from the low-pass filtered and decimated voice waveform data (step S503).

  The feature amount calculation units 11 and 21 calculate an autocorrelation function for each cut out audio frame, and obtain a plurality of maximum points in a specified range. Then, the feature amount calculation units 11 and 21 interpolate the periphery of the plurality of local maximum points four times, and set the position having the maximum maximum value among the local maximum points as the position of the basic period candidate (step S504). ).

  The feature quantity calculation units 11 and 21 divide the value of the autocorrelation function at the position of the basic period candidate by the value of the zeroth-order autocorrelation function to obtain a value indicating the degree of periodicity (step S505). Then, the feature quantity calculation units 11 and 21 perform weighting, add a value indicating the degree of periodicity after weighting, and select an optimum basic period candidate as a basic period using the addition result as an index ( Step S506).

  Here, when the speech frame is a voiced sound section, the feature amount calculation units 11 and 21 may obtain the fundamental cycle of the speech frame and perform the following processing using only the fundamental cycle. Furthermore, when the voice frame is included in the unvoiced sound section or the silent section, the feature amount calculation units 11 and 21 obtain the basic period by interpolating the basic period of the voice frame included in the preceding and following voiced sound sections. The following processing may be performed by using them. Details will be described later.

The feature quantity calculation units 11 and 21 calculate the logarithmic fundamental frequency (LF0) by taking the natural logarithm of the inverse of the fundamental period as the fundamental frequency (F0) (step S507). The feature quantity calculation units 11 and 21 calculate the primary difference ΔLF0 as the primary regression coefficient and the secondary difference Δ ² LF0 as the secondary regression coefficient for the one-dimensional logarithmic fundamental frequency (LF0) (step S508, Step S509). The feature quantity calculators 11 and 21 set the logarithmic fundamental frequency (LF0), the primary difference ΔLF0, and the secondary difference Δ ² LF0 as pitch feature quantities (step S510).

Thus, for each voice frame, a pitch composed of one logarithmic fundamental frequency (LF0), one primary difference ΔLF0, and one secondary difference Δ ² LF0, D _F (= 3) coefficients. Features can be obtained.

The feature amount calculation units 11 and 21 use the EM algorithm to calculate the speech waveform of the speaker from the pitch feature amount (coefficient in all speech frames) composed of D _F (= 3) coefficients calculated for each speech frame. The GMM parameter λ relating to the entire data is calculated (step S511). Then, the feature quantity calculation units 11 and 21 obtain the GMM super vector M from the GMM parameter λ (step S512).

  Based on the GMM supervector M, the feature quantity calculation units 11 and 21 calculate the i vector: w, which is an acoustic feature quantity, using the method described in Non-Patent Document 1 (step S513).

  The processes of the similarity calculation units 22-1 to 22-M, the similarity addition units 23-1 to 23-M, and the selection unit 24 are the same as those in FIG.

  As described above, the feature amount calculation units 11 and 21 according to the second embodiment configure a mixed Gaussian distribution model (GMM) based on pitch feature amounts for the sound waveform data of the sound read from the program sound DB 10 or the complementary sound DB 20. The GMM super vector M is obtained by combining the average vectors μ by the number of mixtures C. Then, the feature amount calculation units 11 and 21 calculate an i vector that is an acoustic feature amount using the pitch feature amount based on the GMM super vector M.

  The selection unit 24 at the latter stage is based on the i vector calculated by the feature quantity calculation units 11 and 21, and the program audio, among the complementary audio DBs 20-1 to 20 -M (among M complementary audio speakers), A complementary speech DB 20-c (complementary speech speaker c) having an acoustic feature that is least similar is selected.

  Here, the complementary speech DB 20-c (complementary speech speaker c) is selected with the acoustic feature amount calculated from the pitch feature amount as an index, and the pitch feature amount is a numerical value representing the pitch of the sound.

  Therefore, even if the program sound and the complementary sound are simultaneously presented as a result of adding the complementary sound to the program sound, the person who listens to these sounds can easily distinguish the program sound and the complementary sound. Thus, it is possible to obtain complementary speech that allows the speaker's voice to be easily recognized.

  As illustrated in FIG. 5, the feature amount calculation units 11 and 21 obtain a fundamental period for a speech frame, calculate a logarithmic fundamental frequency (LF0) and the like using the fundamental period, and an i vector that is an acoustic feature amount: Calculate w. In this case, the feature quantity calculation units 11 and 21 may calculate the i vector: w, which is an acoustic feature quantity, using only the basic period of the voice frame included in the voiced sound section. In addition, the feature amount calculation units 11 and 21 obtain the basic period of the unvoiced sound period and the silent period by interpolating the basic period of the voice frame included in the preceding and following voiced sound periods. Then, the feature quantity calculation units 11 and 21 calculate the i vector: w, which is an acoustic feature quantity, using the basic period of the voice frame included in the voiced sound section and the basic period of the unvoiced sound section and the silent section. May be.

  FIG. 6 is a flowchart illustrating a process example of voice frame section determination as pre-processing of the process illustrated in FIG. 5. The feature amount calculation units 11 and 21 determine, as pre-processing of the processing illustrated in FIG. 5, voiced sound segments, unvoiced sound segments, and silent segments as sections including a voice frame.

  The feature quantity calculators 11 and 21 read the speaker's voice waveform data from the program voice DB 10 or the complementary voice DB 20 (step S601), and perform high frequency emphasis (pre-emphasis) on the voice waveform data (step S602). ). Then, the feature amount calculation units 11 and 21 cut out audio data (audio frames) of a frame having a predetermined window width from the audio waveform data after high-frequency emphasis (step S603). The processes in steps S604 to S612 described below are performed for each audio frame.

  The feature quantity calculation units 11 and 21 calculate the power of the audio frame (step S604), and determine whether or not the power of the audio frame is greater than a preset threshold value (step S605). If it is determined in step S605 that the power of the audio frame is greater than the threshold value (step S605: Y), the feature amount calculation units 11 and 21 assume that the audio frame is included in the sound section, and the process proceeds to step S607.

  On the other hand, if it is determined in step S605 that the power of the audio frame is not greater than the threshold value (step S605: N), the feature amount calculation units 11 and 21 assume that the audio frame is included in the silent section, and the section is silent. A section is set (step S606).

  The feature quantity calculation units 11 and 21 shift from step S605 to calculate the zero crossing number of the audio frame when the power of the audio frame is larger than the threshold (step S607). Then, the feature amount calculation units 11 and 21 determine whether or not the zero crossing number of the audio frame is smaller than a preset threshold value (step S608). If it is determined in step S608 that the number of zero crossings of the audio frame is smaller than the threshold value (step S608: Y), the feature amount calculation units 11 and 21 determine that the audio frame is included in the non-frictional section, and go to step S610. Transition.

  On the other hand, if the feature amount calculation units 11 and 21 determine in step S608 that the number of zero crossings of the audio frame is not smaller than the threshold (step S608: N), the audio frame is included in the frictional section, The section is set as an unvoiced sound section (step S609).

  The feature quantity calculation units 11 and 21 shift from step S608 to calculate the autocorrelation function of the audio frame when the number of zero crossings of the audio frame is smaller than the threshold (step S610). Then, the feature amount calculation units 11 and 21 determine whether or not the autocorrelation function of the audio frame is larger than a preset threshold value (step S611). When the feature quantity calculation units 11 and 21 determine in step S611 that the autocorrelation function of the audio frame is greater than the threshold (step S611: Y), the audio frame is included in the voiced sound interval, and the relevant interval is included. It is set to the voice interval (step S612).

  On the other hand, if the feature amount calculation units 11 and 21 determine in step S611 that the autocorrelation function of the audio frame is not greater than the threshold (step S611: N), the audio frame is included in the unvoiced sound interval, Is set as an unvoiced sound section (step S609).

  Thereby, it is determined whether the voice frame is included in any of a voiced sound section, an unvoiced sound section, or a silent section. The feature quantity calculation units 11 and 21 calculate the i vector: w, which is an acoustic feature quantity, using only the basic period of the voice frame included in the voiced sound section. In addition, the feature quantity calculation units 11 and 21 obtain an unvoiced sound section or a basic period of a silent section based on a basic period of a voice frame included in the preceding and following voiced sound sections, and use this basic period as an acoustic feature quantity. A certain i vector: w may be calculated.

  FIG. 7 is a diagram for explaining an example in which the basic period of the silent section and the unvoiced sound section is obtained from the basic period of the preceding and following voiced sound sections. As shown in FIG. 7, it is assumed that voice frame sections are determined in time series. The feature quantity calculation units 11 and 21 obtain the basic period of the voice frame included in the voiced sound section for the voiced sound section. In addition, for the silent section (see the portion α in FIG. 7), the feature amount calculation units 11 and 21 calculate the basic period of the voice frame included in the silent section sandwiched between the voiced sound sections in the preceding voiced sound section. Calculation is performed by interpolation processing using the basic period near the end and the basic period near the start of the subsequent voiced sound section. The same applies to the unvoiced sound section (see β in FIG. 7).

Example 3
Next, Example 3 will be described. The third embodiment is an example in which the first and second embodiments are combined, and a complementary sound that is easy to distinguish from the program sound is selected from the viewpoint of voice quality and voice pitch. Specifically, in the third embodiment, one supplement from a plurality of complementary sounds is performed based on the acoustic feature using the spectral feature of the first embodiment and the acoustic feature using the pitch feature of the second embodiment. Select audio.

  As in the first embodiment, the feature quantity calculation units 11 and 21 combine the average vectors constituting the mixed Gaussian distribution model (GMM) based on the spectral feature quantity by the number of mixtures as the acoustic feature quantity to obtain the GMM super vector. , I vector is calculated. Similarly to the second embodiment, the feature quantity calculation units 11 and 21 combine the average vectors constituting the mixed Gaussian distribution model (GMM) based on the pitch feature quantity as the acoustic feature quantity by the number of the mixture, thereby combining the GMM super vectors. And the i vector is calculated.

Specifically, the feature quantity calculation units 11 and 21 calculate the i vector: w _s based on the spectrum feature quantity by performing the process shown in FIG. 2, and perform the process shown in FIG. Then, i vector: w _p based on the pitch feature amount is calculated.

The similarity calculation unit 22-m receives, from the feature amount calculation units 11-1 to 11-N, an i vector based on the first to Nth spectral feature amounts: w _s and an i vector based on the pitch feature amount: w. Enter _p . Also, the similarity calculation unit 22-m receives the i vector: w _s based on the m-th spectrum feature amount and the i vector: w _p based on the pitch feature amount from the corresponding feature amount calculation unit 21-m. input.

The similarity calculation unit 22-m, for each of the i vector: w _s based on the spectral feature amount and the i vector: w _p based on the pitch feature amount, each of the first to Nth i vectors: w, The similarity between the m-th i-vector and w is calculated. Then, the similarity calculation unit 22-m outputs the similarities between the first to Nth program sounds and the mth complementary sound to the corresponding similarity addition unit 23-m.

The similarity adding unit 23-m receives the first to first i vectors: w _s based on the spectrum feature quantity and i vectors: w _p based on the pitch feature quantity from the corresponding similarity calculation section 22-m. Each similarity between the Nth program sound and the mth complementary sound is input. Then, the similarity adding unit 23-m adds the similarities to each of the i vector: w _s based on the spectrum feature quantity and the i vector: w _p based on the pitch feature quantity, thereby calculating the sum of the similarity degrees. calculate. Thereby, the sum total of the similarity in the acoustic feature amount using the spectral feature amount and the sum of the similarity in the acoustic feature amount using the pitch feature amount are obtained. The similarity addition unit 23-m weights and adds the two calculation results with a preset weighting coefficient, obtains the sum of the similarities, and outputs it to the selection unit 24.

Here, let s _Sm be the sum of the similarities obtained by the equations (1) and (2) for the i vector: w _s based on the spectral feature quantity. Also, let s _Pm be the sum of the similarities obtained by the above equations (1) and (2) for the i vector: w _p based on the pitch feature quantity. When the weighting coefficient is g, the similarity sum s _SPm , which is the result of weighted summation of similarities s _Sm and s _Pm , is expressed by the following equation.

ｇ＝１．０の場合は実施例１を示し、ｇ＝０．０の場合は実施例２を示す。 The weighting coefficient g is a real number that takes a value in the following range.

In the case of g = 1.0, Example 1 is shown, and in the case of g = 0.0, Example 2 is shown.

  The selection unit 24 inputs the sum of similarity additions from the similarity addition units 23-1 to 23 -M, and specifies the sum of the minimum similarity among the addition sums of these similarities. Then, the selection unit 24 selects the complementary speech DB 20 (complementary speech speaker) corresponding to the sum total of the minimum similarity among the supplementary speech DBs 20-1 to 20-M (among M supplementary speech speakers). Select and output selection information.

Here, the supplementary speech DB 20 (complementary speech speaker c) corresponding to the sum total s _SPm of the minimum similarity is set as the supplementary speech DB 20-c (complementary speech speaker c), and the selection information is c (1 to N). If any value), the selection information c is selected by the following equation.

  As described above, the feature amount calculation units 11 and 21 according to the third embodiment calculate the i vector based on the spectral feature amount and the i vector based on the pitch feature amount as the acoustic feature amount.

  The similarity calculation unit 22-m, for each of the i vector based on the spectral feature amount and the i vector based on the pitch feature amount, is between the first to Nth program sounds and the mth complementary sound. The similarity is calculated. Then, the similarity addition unit 23-m calculates the sum of the similarities by adding the similarities for each of the i vector based on the spectral feature quantity and the i vector based on the pitch feature quantity, thereby calculating two calculations. The results are weighted and added to determine the sum of the similarities.

  Based on the sum total of the similarities, the selection unit 24 complements the acoustic features that are most similar to the program audio among the complementary audio DBs 20-1 to 20 -M (among M complementary audio speakers). The voice DB 20-c (complementary voice speaker c) is selected.

  Here, the complementary speech DB 20-c (complementary speech speaker c) is selected using as an index the acoustic feature amount calculated from the spectral feature amount and the acoustic feature amount calculated from the pitch feature amount. Further, as described above, the spectral feature amount reflects the frequency component of the voice, and the voice quality is determined by the frequency component of the voice. The pitch of the sound is determined by the pitch feature amount.

  Therefore, even if the program sound and the complementary sound are simultaneously presented as a result of adding the complementary sound to the program sound, the person who listens to these sounds can easily distinguish the program sound and the complementary sound. In addition, it is possible to obtain complementary speech in which the voice quality and pitch of the speaker can be easily distinguished.

  In particular, the addition sum of similarities, which is an index for selecting the complementary speech DB 20-c (complementary speech speaker c), is weighted for each of the i vector based on the spectral feature amount and the i vector based on the pitch feature amount. Is reflected. That is, when importance is attached to voice quality, the sum of similarities in which the voice quality is reflected is calculated by bringing the weighting coefficient of the i vector based on the spectrum feature amount close to 1.0. In addition, when importance is attached to the voice pitch, the sum of the similarities reflecting the voice pitch is calculated by bringing the weighting coefficient of the i vector based on the pitch feature amount close to 1.0. . Therefore, by setting a weighting coefficient corresponding to the program sound in advance, it is possible to obtain complementary sound that is easier to distinguish from the program sound.

  The present invention has been described with reference to the first to third embodiments. However, the present invention is not limited to the first to third embodiments, and various modifications can be made without departing from the technical idea thereof. For example, in the first embodiment, the acoustic feature amount based on the spectral feature amount is calculated, and in the second embodiment, the acoustic feature amount based on the pitch feature amount is calculated. In the third embodiment, the acoustic feature quantity based on the spectral feature quantity and the acoustic feature quantity based on the pitch feature quantity are calculated. In the present invention, the calculation method of the acoustic feature amount is not limited to the method based on the spectrum feature amount or the method based on the pitch feature amount, and other methods may be used.

  For example, it is assumed that three different types of acoustic feature quantities are calculated using three different types of methods. The feature amount calculation units 11 and 21 calculate the first to third i vectors using the first to third methods, respectively. The similarity calculation unit 22-m calculates the similarity between each of the first to Nth program sounds and the mth complementary sound for each of the first to third i vectors. Then, the similarity adder 23-m calculates the sum of the similarities by adding the similarities for each of the first to third i vectors, adds the three calculation results by weighting, and adds the similarities. Find the total sum of. Based on the sum total of the similarities, the selection unit 24 complements the acoustic features that are most similar to the program audio among the complementary audio DBs 20-1 to 20 -M (among M complementary audio speakers). The voice DB 20-c (complementary voice speaker c) is selected.

  Note that a normal computer can be used as the hardware configuration of the voice selection device 1 according to the embodiment of the present invention. The voice selection device 1 is configured by a computer including a volatile storage medium such as a CPU and a RAM, a non-volatile storage medium such as a ROM, an interface, and the like. Feature amount calculation units 11-1 to 11-N, feature amount calculation units 21-1 to 21-M, similarity calculation units 22-1 to 22-M, and similarity addition unit 23-1 included in the voice selection device 1 Each of the functions of ˜23-M and the selection unit 24 is realized by causing the CPU to execute a program describing these functions. These programs (voice selection programs) may be stored and distributed in a storage medium such as a magnetic disk (floppy (registered trademark) disk, hard disk, etc.), an optical disk (CD-ROM, DVD, etc.), or a semiconductor memory. It can also be transmitted / received via a network.

1 Voice Selector 10-1 to 10-N Program Voice DB
11-1 to 11-N, 21-1 to 21-M feature quantity calculation units 20-1 to 20-M Complementary speech DB
22-1 to 22-M Similarity Calculation Units 23-1 to 23-M Similarity Addition Unit 24 Selection Unit

Claims (5)

In the audio selection device for selecting the complementary audio when presenting the program audio by adding the complementary audio, from a plurality of complementary audios,
A program audio DB (database) in which a predetermined number of program audio data of 1 or more are stored;
A complementary speech DB storing a predetermined number of supplemental speech data of 2 or more;
An acoustic feature is calculated for each of the predetermined number of program audio data stored in the program audio DB, and an acoustic feature is calculated for each of the predetermined number of complementary audio data stored in the complementary audio DB. A feature amount calculation unit to be calculated;
An acoustic feature amount for each of the predetermined number of program audio data calculated by the feature amount calculation unit and an acoustic feature amount for each of the predetermined number of complementary audio data calculated by the feature amount calculation unit A similarity calculation unit that calculates the similarity between,
For each of the supplementary audio data, the similarity between the acoustic feature amount for each of the predetermined number of program audio data calculated by the similarity calculation unit and the acoustic feature amount of the complementary audio data is added. , A similarity addition unit for calculating the sum,
A selection unit that specifies a minimum sum among the sums for each of the complementary speech data obtained by the similarity addition unit, and selects the complementary speech data corresponding to the minimum sum from the predetermined number of supplemental speech data When,
A voice selection device comprising: The voice selection device according to claim 1,
The feature amount calculation unit includes:
For each of the program audio data and the complementary audio data, audio data is cut out in units of frames of a predetermined length, a frequency characteristic is obtained for each audio data in units of frames, and a mel frequency cepstrum is obtained based on the frequency characteristics. Spectral features including a static coefficient composed of a coefficient and logarithmic energy, and a primary regression coefficient and a quadratic regression coefficient of the static coefficient are obtained. Calculating a GMM parameter comprising weights and Gaussian distributions for the number of mixtures, extracting an average vector of the Gaussian distributions from the GMM parameters, obtaining a GMM supervector obtained by combining the average vectors by the number of mixtures; Based on the super vector, the i vector which is the acoustic feature amount is calculated. To the voice selection device, characterized in that. The voice selection device according to claim 1,
The feature amount calculation unit includes:
For each of the program audio data and the complementary audio data, audio data is cut out in units of frames of a predetermined length, a basic period candidate is set for each audio data in units of frames, and the periodicity of the basic period candidates is set. A fundamental period is extracted from the fundamental period candidates by obtaining a degree, and a pitch feature amount including a logarithmic fundamental frequency and a primary regression coefficient and a secondary regression coefficient of the logarithmic fundamental frequency is obtained based on the fundamental period, Based on the pitch feature amount, an EM algorithm is used to calculate a GMM parameter composed of a mixture weight for the number of mixtures and a Gaussian distribution for the number of mixtures, and an average vector of the Gaussian distribution is extracted from the GMM parameter, and the average vector GMM supervectors obtained by combining the same number as the number of the mixture, and based on the GMM supervectors, Calculating the i vectors are sounding feature amount, the audio selection device, characterized in that. The voice selection device according to claim 1,
The feature amount calculation unit includes:
For each of the program audio data and the complementary audio data, audio data is cut out in units of frames of a predetermined length, a frequency characteristic is obtained for each audio data in units of frames, and a mel frequency cepstrum is obtained based on the frequency characteristics. Spectral features including a static coefficient composed of a coefficient and logarithmic energy, and a primary regression coefficient and a quadratic regression coefficient of the static coefficient are obtained. Calculating a GMM parameter comprising weights and Gaussian distributions for the number of mixtures, extracting an average vector of the Gaussian distributions from the GMM parameters, obtaining a GMM supervector obtained by combining the average vectors by the number of mixtures; Based on the super vector, the first i-vector that is the acoustic feature amount Is calculated,
For each frame of audio data, set a basic period candidate, obtain a degree of periodicity of the basic period candidate, extract the basic period from the basic period candidate, and based on the basic period, A pitch feature amount including a primary regression coefficient and a secondary regression coefficient of the logarithmic fundamental frequency is obtained, and an EM algorithm is used on the basis of the pitch feature amount from a mixture weight for the number of mixtures and a Gaussian distribution for the number of mixtures. GMM parameters are calculated, an average vector of the Gaussian distribution is extracted from the GMM parameters, a GMM super vector obtained by combining the average vectors by the number of the mixture is obtained, and the acoustic feature amount is calculated based on the GMM super vector. Calculate a second i-vector,
The similarity calculation unit includes:
A first i vector for each of the predetermined number of program audio data calculated by the feature amount calculation unit and a first for each of the predetermined number of complementary audio data calculated by the feature amount calculation unit. Calculate similarity between i vector and
A second i vector for each of the predetermined number of program audio data calculated by the feature amount calculation unit and a second i vector for each of the predetermined number of complementary audio data calculated by the feature amount calculation unit. calculate the similarity between the i vector and
The similarity adding unit includes:
For each of the complementary audio data, the similarity between the first i vector for each of the predetermined number of program audio data and the first i vector of the complementary audio data calculated by the similarity calculation unit Add the degree, find the first addition result,
For each of the complementary audio data, the similarity between the second i vector for each of the predetermined number of program audio data and the second i vector of the complementary audio data calculated by the similarity calculation unit Add the degree, find the second addition result,
A voice selection device characterized by weighting and adding the first addition result and the second addition result to obtain the sum.   A voice selection program for causing a computer to function as the voice selection device according to any one of claims 1 to 4.

Images