JPH0816183A - Sound quality adaption device - Google Patents

Abstract

PURPOSE:To attain high quality editing synthesis by converting to a spectrum parameter of a reference sound in an optimum section decided based on a decision result of a distance value between an adaptive input sound and the reference sound. CONSTITUTION:A distance value calculation means 14 calculates and outputs a spectrum parameter distance value 15 such as spectrum strain by correspondence between frames by e.g. DP matching between a partial reference sound spectrum parameter 12 and a partial adaptive input sound spectrum parameter 13. An optimum section decision means 16 inputs the spectrum parameter distance value 15, and detects an optimum reference sound section to output it as an optimum reference sound partial section 17. Then, a conversion means 18 converts the partial adaptive input sound spectrum parameter 13 into the partial reference sound spectrum parameter 12 of the optimum reference sound partial section 17 with frame correspondence by the DP matching to output it as a conversion sound spectrum parameter 19.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice quality adaptation device for associating an input voice with a reference voice by learning, and in particular, connects voice data accumulated between voices having different utterances from different speakers. The present invention relates to a voice quality adaptation device applicable to edit synthesis.

[0002]

2. Description of the Related Art Edit synthesis is a technique for outputting a fixed message by voice by connecting a large amount of prerecorded and accumulated words, phrases, or voice data in units of sentences according to the purpose. If there is a request to output an unregistered word due to updating of the contents of, etc., it is necessary to add new voice data. However, the additionally recorded voice data may have different voice quality from the existing data due to the difference in the speaker, or even in the case of the same speaker, due to the difference in the recording environment and the state of the vocalization. The synthetic sound becomes unnatural at the connection point in synthesis. To avoid this, if you try to match the quality of the existing data with the additional data,
It requires a great deal of effort to re-record all existing data. Therefore, a technology is required to make the quality of the added voice data (input voice data) equal to that of the existing data (reference voice data). Conventionally, the method is based on codebook mapping applied to voice quality conversion. There is a spectrum adaptation method, and the literature "Voice quality conversion by vector quantization" (Abe, Nakamura, Kano, Kuwahara, Proceedings of the Acoustical Society of Japan, October 1987, pp.215.
-216).

FIG. 5 is a block diagram showing an example of the configuration of a conventional spectrum adaptation system by codebook mapping based on this technique. In the figure, 3 is a spectrum adapting means for performing spectrum adaptation by codebook mapping, 2 is an input speech spectrum parameter which is an input of the spectrum adapting means 3, and 4 is an output of the spectrum adapting means 3, which is a reference speech. The adapted input speech spectrum parameter subjected to spectrum adaptation, 32 is a reference speech spectrum parameter for learning which is a small number of reference speech data accumulated in the spectrum adaptation means 3 and used for learning between the input speech and the reference speech, and 33 is Similar to the learning reference speech spectrum parameter 32, the learning input speech spectrum parameter is a small number of input speech data having the same utterance content as the learning reference speech data used for learning.
Further, 24 is a vector quantizing means for searching a code vector having the highest degree of similarity in the code book with respect to the input spectral parameter and outputting the code number, 22
Is a reference voice vector quantization codebook used for vector quantization of the reference voice, 34 is an input voice vector quantization codebook used for vector quantization of the input voice, and 25 is a reference voice vector quantization of the spectrum parameter of the reference voice. Reference voice vector quantization code number sequence obtained by vector quantization in the codebook 22, 35 is an input voice vector quantization code number sequence obtained by vector quantization of the input voice spectrum parameter in the input voice vector quantization codebook 34. , 36 is a learning means for making a correspondence between the input voice data for learning and the reference voice data to create a codebook for converting the input voice data into the reference voice data, and 23 is an adaptive input which is an output of the learning means 36. Speech vector quantization codebook, 26 is the spectrum of the input speech The adaptive input speech vector quantization code number sequence obtained by quantizing the parameter with the adaptive input vector quantization codebook 23, 37 inputs the vector quantization code number sequence, and the code vector at that code number in the codebook Is a decoding means for outputting.

The operation of the conventional spectrum adaptation method by codebook mapping will be described below with reference to FIGS. 5 and 6. In the spectrum adaptation means 3,
The vector quantization means 24 vector-quantizes the learning reference speech spectrum parameter 32, which is learning data, by the reference speech vector quantization codebook 22, and outputs a reference speech vector quantization code number sequence 25. Similarly, the learning input voice spectrum parameter 33, which is learning data, is also vector-quantized by the input voice vector quantization codebook 34 by the vector quantization means 24 and output as the input voice vector quantization code number sequence 35. In the learning means 36, the reference speech vector quantization code number sequence 25 and the input speech vector quantization code number sequence 35 are respectively referred to as a reference speech vector quantization codebook 22 and an input speech vector quantization codebook 34.
Then, the quantized reference voice data and the input voice data are associated with each other in the time direction. As a general method for efficiently performing this association, DP (Dynami)
c Programming) There is a method of nonlinear expansion / contraction of the time axis by matching, and as shown in FIG. 6, by selecting a path that minimizes the sum of the spectral distance values between both data, the i-th frame of the reference audio data is selected. Is associated with the j-th frame of the input voice data. Then, the vector quantization code number of the input voice that most corresponds to the vector quantization code number of the reference voice is obtained, and the code vector of the reference voice vector quantization codebook 22 is calculated based on the vector quantization code number of the input voice vector quantization codebook 34. An adaptive input speech vector quantization codebook 23 is created by converting into a code vector and is output. The learning operation between the codebooks up to this point is performed in advance. The adaptive input speech vector quantization codebook 23 vector-quantizes the inputted input speech spectrum parameter 2 by the vector quantizing means 24 for each frame, and outputs an adapted input speech vector quantization code number sequence 26. The decoding means 37 inputs the adapted input speech vector quantization code number sequence 26, refers to the reference speech vector quantization codebook 22, and derives a code vector corresponding to the vector quantization code number from the reference speech vector quantization codebook 22. , Output as the adapted input speech spectrum parameter 4.

[0005]

Since the conventional voice quality adaptation apparatus based on codebook mapping is configured as described above, the values of the code vectors in each codebook are discrete, and the adaptation input speech is particularly important. Since the vector quantization codebook uses the input speech codevector which has the largest correlation with the reference speech codevector by the DP matching operation, the discreteness between the codes becomes larger. As a result, the converted spectrum parameter becomes discontinuous between frames, and there is a problem in that the continuous feature held by natural speech in the time direction cannot be reproduced. Also, there was a problem in using the codebook itself. That is, since the codebook uses the centroid of the set of corresponding spectral parameters as each code vector, the phonological features such as formants are flattened when the variance of the spectral parameters for each code is large. As a result, there is a problem in that the phonological property is blurred and causes deterioration of the sound quality. Further, as a method for solving the problem of such a codebook mapping method, a voice quality conversion method by associating segment unit voices as disclosed in Japanese Patent Laid-Open No. 4-158397 has been proposed. However, since this method uses pre-segmented data as a standard pattern, the result is that the parameter has characteristics in the time direction, which tends to increase quantization distortion compared to the codebook mapping method. To avoid this, there was a problem that a huge amount of standard patterns of various time lengths had to be prepared.
Furthermore, as the number of standard patterns increases as compared with the codebook mapping method, the number of learning samples corresponding to the standard patterns decreases, and as a result, the influence of habit due to the uneven distribution of the samples becomes stronger.

The present invention has been made to solve the above problems, and an object of the present invention is to provide a voice quality adaptation device which can be applied to high-quality edit synthesis.

[0007]

A speech quality adaptation device according to the first aspect of the present invention comprises a spectrum adaptation means for adapting a spectrum feature of each frame of an input speech using learning data, and a phoneme model. Using the labeling means for outputting the phoneme label including the phoneme boundary information for each of the adapted input speech and the reference speech which are the outputs of the spectrum adaptation means, and the input speech from the phoneme label string of the adapted input speech given by the labeling means. A partial section determining unit that determines a partial section and detects a section that matches the phoneme label string of the input speech partial section from the phoneme label string of the reference speech given by the labeling unit, and a section that is an output of the partial section determining unit. A cutting-out means for cutting out each of the adapted input voice and the reference voice based on the section information, and the cut-out means. A distance value calculation means for calculating a distance value between the cut-out adapted input speech and the reference sound by means,
An optimum section determining means for selecting the optimum section of the reference voice by judging the distance value obtained by the distance value calculating means, and a reference voice in the optimum section determined by the optimum section determining means for the adapted input voice cut out by the cutout means. And a conversion means for converting into.

Also, the voice quality adaptation device according to the second aspect of the present invention uses the input voice instead of the adaptive input voice as the input of the labeling means and the clipping means in the first aspect of the invention.

The speech quality adaptation apparatus according to the third aspect of the present invention further comprises a reference speech vector quantization codebook and an adaptive input speech vector quantization codebook for vector quantization of the reference speech and the input speech for each frame. And a vector quantizing means for determining the partial interval for the input speech vector quantization code number sequence output from the vector quantizing means and outputting the input speech partial vector quantization code number sequence and It is an output of the code number sequence cutting-out means for detecting and cutting out a section that matches or is similar to the input voice partial vector quantization code number sequence for the output reference voice vector quantization code number sequence, and the output of the code number sequence cutting out means. Find distance value between input voice and reference voice for partial code number sequence The separation value calculating means, the optimum interval determining means for judging the distance value obtained by the distance value calculating means and selecting the optimum interval of the reference voice, and the input voice partial vector quantization code number output by the code number string cutting means. Conversion means for converting the sequence into a reference voice in the optimum section determined by the optimum section determination means.

A voice quality adaptation device according to a fourth aspect of the present invention is a spectrum adaptation means for adapting the spectrum feature of each frame of voice, and an adaptation input which is an output of the spectrum adaptation means. A distance value is calculated between the input voice cutout unit that cuts out a voice segment, the adaptive input voice cut out by the input voice cutout unit, and all the partial segments of the reference voice, and the optimum section of the reference voice is determined by the distance value determination. The reference section determining means for obtaining and the converting means for converting the adapted input speech cut out by the input speech cutting out means into the reference speech in the optimum section determined by the reference section determining means are provided.

[0011]

In the speech quality adaptation apparatus according to the present invention, the labeling means is used to give a phoneme label containing phoneme boundary information to the adapted input speech using the phoneme model, and the sub-interval determining means is supplied with the adapted input. The input phoneme label section is determined from the phoneme label string of the voice and output, and the section that matches the phoneme label string of the input phoneme section section is searched from the phoneme label string of the input reference speech, and the section is set as the reference voice section section. Output. Further, the cutout unit cuts out the adapted input speech in the input speech partial section, outputs the partially adapted input speech, similarly cuts out the reference speech in the reference speech partial section, and outputs the partial reference speech. Further, the distance value calculating means calculates a spectral distance value between the partially adapted input speech and the partial reference speech, and the optimum section selecting means determines the distance value and detects the optimum reference speech section. It is output as the optimum reference voice partial section. Then, the conversion means converts the partially adapted input speech into a partial reference speech in the optimum reference speech partial section.

Further, in the speech quality adaptation apparatus according to the present invention, the labeling means gives a phoneme label containing phoneme boundary information to the input adaptation input speech using the phoneme model, and the partial interval determining means is inputted. The input speech part interval is determined from the phoneme label string of the adapted input speech and output, and the reference speech part is searched from the phoneme label string of the input reference speech for a section that matches the phoneme label string of the input speech part interval. Output as a section. The cutout unit cuts out the input voice in the input voice partial section, outputs the partial input voice, similarly cuts out the reference voice in the reference voice partial section, and outputs the partial reference voice. Further, the distance value calculating means calculates a spectral distance value between the partial input speech and the partial reference speech, and the optimum section selecting means determines the distance value, detects the optimum reference speech section, and detects the optimum reference speech. Output as a partial section. Then, the conversion means converts the partial input voice into a partial reference voice in the optimum reference voice partial section.

Further, in the voice quality adaptation device according to the present invention, the input voice partial vector quantized code number sequence is determined and output from the vector quantized code number sequence of the input voice to which the code number sequence cutout means is input. , Search a section that matches or is similar to the input speech partial vector quantization code number sequence from the vector quantization code number sequence of the input reference speech, cut out the reference speech vector quantization code number sequence in that section, and refer to it. The audio partial vector quantization code number sequence is output. Further, the distance value calculation means refers to the reference speech vector quantization codebook and the adaptive input speech vector quantization codebook, and determines between the input speech partial vector quantization code number sequence and the reference speech partial vector quantization code number sequence. Calculate the spectral distance value according to the codebook. The optimum section determination means determines the distance value, detects the optimum reference speech section and outputs it as the optimum reference speech partial section, and the conversion section refers to the optimum reference speech partial section from the input speech partial vector quantization code number sequence. Convert to voice.

Further, in the voice quality adaptation device according to the present invention, the input voice cutout means cuts out a certain section of the input adaptive input voice, and the reference section determination means appropriately cuts out the reference voice. , The spectral distance value is calculated between all the cut out reference speeches and the partially adapted input speech, the distance value is determined, the optimal reference speech section is detected, and the optimal reference speech partial section is output.
Further, the conversion means converts the partially adapted input speech into a partial reference speech in the optimum reference speech partial section.

[0015]

【Example】

Example 1. A first embodiment of the present invention will be described with reference to FIG. FIG. 1 is a configuration diagram showing a configuration of a voice quality adaptation device, where 1 is a reference voice spectrum parameter input to this device for voice quality adaptation of an input voice spectrum parameter 2 and 5 is an input spectrum parameter. Labeling means for giving a phoneme label to the label, 6 is a labeling means 5
Is a reference speech phoneme label string given to the reference speech spectrum parameter 1, 7 is an adapted input speech phoneme label string given to the adapted input speech spectrum parameter 4 by the labeling means 5, and 8 is a phoneme label string to be inputted. Based on the voice data, a sub-segment determining means for deciding a sub-segment having a length of about a syllable or a voice unit, 9 is a reference voice partial segment that is a sub-segment for a reference voice, 10
Is an input voice partial section which is a partial section for the input voice. Further, 11 is a cutout unit for extracting the spectrum parameter in the partial section, and 12 is a cutout unit 11
The partial reference speech spectrum parameter extracted from the reference speech spectrum parameter 1 by 13 is the partially adapted input speech spectrum parameter 13 extracted from the adapted input speech spectrum parameter 4 by the clipping means 11, and 14 is the two speech data in the partial section. Distance value calculating means for obtaining a spectral distance value between, 15 is a spectrum parameter distance value, 16 is an optimum interval selecting means for selecting a partial reference speech spectrum parameter 12 considered to be optimal based on the determination of the spectrum parameter distance value 15, 1
Reference numeral 7 is the optimum reference speech partial section, 18 is the speech spectrum parameter of the input partial section, and is the optimum reference speech partial section 1
The conversion means for converting the partial reference speech spectrum parameter 12 in 7 and the conversion spectrum parameter 19 output from the conversion means 18. In addition, 2, 3 and 4 in the figure
Is equivalent to the conventional example.

Next, the operation will be described. The spectrum adapting means 3 inputs the input speech spectrum parameter 2, converts the spectrum from the input speech to the reference speech in each frame by, for example, a codebook mapping method using learning data in the conventional example, and adapts the input. The voice spectrum parameter 4 is output. The labeling means 5 inputs the reference speech spectrum parameter 1 and the adapted input speech spectrum parameter 4, respectively, and uses the phoneme model to provide a reference speech phoneme label sequence 6 including phoneme boundary information and an adapted input speech phoneme label. Row 7
Is output. The partial section determining means 8 inputs the reference speech phoneme label sequence 6 and the adapted input speech phoneme label sequence 7, and determines a section delimited by a phoneme boundary by observing the adapted input speech phoneme label sequence 7. The input speech subsection 10 is output, and on the other hand, the reference speech phoneme label sequence 6 is searched for a section that matches the adapted input phoneme label sequence 7 in the input speech subsection 10 and is output as the reference speech subsection 9. The clipping means 11 clips the reference speech spectrum parameter 1 based on the reference speech partial section 9 and outputs a partial reference speech spectrum parameter 12. Similarly, the adaptive input speech spectrum parameter 4 is cut out based on the input speech partial section 10 and the partially adapted input speech spectrum parameter 13 is output. The distance value calculation means 14 causes a spectrum distortion between the partial reference speech spectrum parameter 12 and the partially adapted input speech spectrum parameter 13 by, for example, associating frames by DP matching as shown in FIG. The spectrum parameter distance value 15 is calculated and output. The optimum section determination means 16 receives the spectrum parameter distance value 15 and detects the optimum reference speech section by a distance value determination method that detects a distance value that is the smallest and is equal to or less than a certain threshold. The optimal reference voice partial section 17 is output. The conversion means 18 converts the partially adapted input speech spectrum parameter 13 into the partial reference speech spectrum parameter 12 of the optimum reference speech partial section 17 in frame correspondence by DP matching shown in FIG.
Output as 9.

According to the present embodiment, it is possible to solve the problem that the spectrum parameter is discontinuous between frames and the phonology is blurred without using a standard pattern for each segment, and the quality is high. Voice quality adaptation can be achieved.

Example 2. A second embodiment of the present invention will be described with reference to FIG. FIG. 2 is a configuration diagram showing a configuration of the voice quality adaptation device, 2 is equivalent to the conventional example, and 1, 5, 6, 8, 9, 10, 11, 11, 12, 14, 1
5, 16, 17, 18, and 19 are the same as those described in the first embodiment. 20 is an input speech phoneme label string given to the input speech spectrum parameter 2 by the labeling means 5, and 21 is a partial input speech spectrum parameter extracted from the input speech spectrum parameter 2 by the clipping means 11.

Next, the operation will be described. The labeling means 5 inputs the reference speech spectrum parameter 1 and the input speech spectrum parameter 2, respectively, and outputs the reference speech phoneme label sequence 6 and the input speech phoneme label sequence 20 including the phoneme boundary information for each using the phoneme model. .
The sub-segment determining means 8 inputs the reference speech phoneme label sequence 6 and the input speech phoneme label sequence 20, determines a segment demarcated by a phoneme boundary by observing the input speech phoneme label sequence 20, and the input speech sub-segment 10 On the other hand, the reference speech phoneme label sequence 6 is searched for a section in the input speech partial section 10 that matches the input phoneme label sequence 20, and is output as the reference speech partial section 9. Cutting means 11
Cuts out the reference speech spectrum parameter 1 based on the reference speech partial section 9 and outputs the partial reference speech spectrum parameter 12. Similarly, the input voice spectrum parameter 2 is cut out based on the input voice partial section 10 and the partial input voice spectrum parameter 21 is output. The distance value calculation means 14 associates the partial reference speech spectrum parameter 12 and the partial input speech spectrum parameter 21 with each other by, for example, DP matching as shown in FIG. The parameter distance value 15 is calculated and output. The optimum section determination means 16 receives the spectrum parameter distance value 15 and detects the optimum reference speech section by a distance value determination method that detects a distance value that is the smallest and is equal to or less than a certain threshold. The optimal reference voice partial section 17 is output. The conversion means 18 converts the partial input speech spectrum parameter 21 into
The partial reference speech spectrum parameter 12 of the optimum reference speech partial section 17 is converted corresponding to the frame by the DP matching shown in FIG. 6, and is output as the converted speech spectrum parameter 19.

According to the present embodiment, it is possible to solve the problem that the spectrum parameter is discontinuous between frames and the phonology is blurred without using a standard pattern for each segment, and the quality is high. Voice quality adaptation can be achieved.

Example 3. A third embodiment of the present invention will be described with reference to FIG. FIG. 3 is a configuration diagram showing the configuration of the voice quality adaptation device, in which 2, 22, 23, 24, 25 and 26 are equivalent to those of the conventional example, and 1, 14, 15, 1
6, 17, 18 and 19 are the same as those described in the first embodiment. Reference numeral 27 is a code number sequence cutout means for extracting a code number sequence of a partial section of a certain length from the input code number sequence, and 28 is extracted from the reference voice vector quantization code number sequence 25 by the code number sequence cutout means 27. The reference voice partial vector quantized code number sequence, 29 is the input voice partial vector quantized code number sequence extracted from the adaptive input voice vector quantized code number sequence 26 by the code number sequence slicing means 27.

Next, the operation will be described. The vector quantization means 24 receives the reference speech spectrum parameter 1 and performs vector quantization by the reference speech vector quantization codebook 22 and outputs a reference speech vector quantization code number sequence 25. On the other hand, the input speech spectrum parameter 2 is input and vector quantization is performed using the adaptive input speech vector quantization codebook 23 that is associated with the reference speech vector quantization codebook 22 as in the conventional example. , And outputs an adapted input speech vector quantization code number sequence 26. The code number sequence cutting-out means 27 uses the reference voice vector quantization code number sequence 25.
And the adaptive input speech vector quantization code number sequence 26
Is input, the input voice partial vector quantization code number sequence 29 is determined and output by observing the adaptive input voice vector quantization code number sequence 26, and the input is performed from the reference voice vector quantization code number sequence 25. A section that matches or is similar to the voice partial vector quantization code number sequence 29 is searched, and the reference voice vector quantization code number sequence 25 in that section is cut out and output as the reference voice partial vector quantization code number sequence 28. Distance value calculation means 1
4 is the reference voice partial vector quantization code number sequence 2
8 and the input speech partial vector quantization code number sequence 29
By inputting the reference speech vector quantization codebook 2
2 and the adaptive input speech vector quantization codebook 2
3, each vector quantization code number sequence is decoded, and a frame by DP matching as shown in FIG. 6, for example, between the spectrum parameter of the decoded partial reference voice and the partial input voice spectrum parameter is decoded. A spectral parameter distance value 15 such as spectral distortion is calculated and output by associating between the values. The optimum interval determination means 16 receives the spectrum parameter distance value 15 and detects the optimum reference speech interval by a distance value determination method that detects a distance value that is the minimum and is less than a certain threshold value. , As the optimum reference voice partial section 17. The converting means 18 converts the input speech partial vector quantization code number sequence 29 into the optimum reference speech partial section 1
The reference voice spectrum parameter 1 of No. 7 is converted corresponding to the frame by the DP matching shown in FIG. 6, and is output as the converted voice spectrum parameter 19.

In the present embodiment, by adopting such a configuration, it is possible to solve the problem that the spectrum parameter becomes discontinuous between frames and the phonology is blurred without using a standard pattern for each segment. Therefore, it is possible to realize high-quality voice quality adaptation.

Example 4. A fourth embodiment of the present invention will be described with reference to FIG. FIG. 4 is a configuration diagram showing the configuration of the voice quality adaptation device, 2, 3, 4 are equivalent to the conventional example, and 1, 13, 17, 18, 19 are equivalent to those described in the first embodiment. belongs to. Reference numeral 30 designates an input speech segmentation means for deciding a subsection having a length of about a syllable or a speech segment in the input speech data, and outputting an adapted input speech spectrum parameter in the section, and 31 a part of the partially adapted input speech spectrum parameter 13 It is a reference section determining means for determining a sub section of the reference voice that is considered to be optimal based on the distance value determination.

The operation will be described below. The spectrum adapting means 3 inputs the input speech spectrum parameter 2 and converts the spectrum from the input speech to the reference speech in each frame by, for example, the method of codebook mapping using learning data in the conventional example, and adapts the adapted input speech. The spectrum parameter 4 is output. The input speech cut-out means 30 receives the adapted input speech spectrum parameter 4 and selects, for example, a spectrum change position as a boundary, and the partially adapted input speech spectrum parameter 1
Output as 3. The reference section determining means 31 selects and cuts the reference speech spectrum parameter 1 by using, for example, a change position of the spectrum as a boundary, and cuts out all the extracted reference speech spectrum parameter 1 and the partially adapted input speech spectrum parameter 13. In between,
For example, by associating frames by DP matching as shown in FIG. 6, a spectrum parameter distance value such as spectrum distortion is calculated, and a distance value that is minimum and is equal to or less than a certain threshold is detected. The optimum reference voice section is detected by the distance value determination method, and is output as the optimum reference voice partial section 17. The conversion means 18 is
The partially adapted input speech spectrum parameter 13 is converted into the reference speech spectrum parameter 1 of the reference speech determination section 17 corresponding to the frame by the DP matching shown in FIG. 6, and output as the converted speech spectrum parameter 19.

In the present embodiment, by adopting such a configuration, it is possible to solve the problem that the spectrum parameter becomes discontinuous between frames and the phonology is blurred without using a standard pattern for each segment. Therefore, it is possible to realize high-quality voice quality adaptation.

Example 5. In the distance value calculating means of the third embodiment, instead of decoding the vector quantization code number sequence of the input voice and the reference voice with the vector quantization codebook to calculate the spectrum parameter distance value, the reference voice codebook and adaptation are applied in advance. Equipped with a table that describes the spectral parameter distance values between all code vectors with the input speech codebook, and refers to the table according to the correspondence between the reference speech and the input speech frames, and the selected interframe distance values The distance value may be calculated based on the total sum of.

Example 6. In the partial section determining means of the first and second embodiments and the code number string cutting means of the third embodiment, the section of the reference voice that matches the section of the input voice is searched for from the longer matching section by the longest matching method. You may make it select and detect.

Example 7. In the distance value calculation means of the first, second, third, fifth embodiment, the reference section determination means of the fourth embodiment, and the conversion means of the first, second, third, and fourth embodiment, between the frames of the input voice and the reference voice. The correspondence may be performed by linear expansion / contraction instead of nonlinear expansion / contraction by DP matching.

[0030]

As described above, according to the present invention, a partial section is determined by comparing phoneme strings of the adaptive input speech, and the distance value between the adaptive input speech and the reference speech in the section is determined. Since it is designed to convert to the spectrum parameter of the reference speech of the optimum section decided based on the result, it is possible to reduce the influence of the discontinuity in the time direction and the blur of the phoneme in the conventional codebook mapping method. Therefore, a voice quality adaptation device capable of high quality editing and synthesis can be realized.

Further, according to the present invention, the partial segment is determined by comparing the phoneme strings of the input speech, the distance value between the input speech and the reference speech in the segment is determined, and the optimum segment reference based on the determination is performed. Since the conversion to speech spectral parameters is performed, the effects of discontinuity in the time direction and phonological blurring in the conventional codebook mapping method can be reduced, and high-quality edit synthesis is possible. A voice quality adaptation device can be realized.

Further, according to the present invention, the partial section is determined by comparing the vector quantization code number sequences of the adapted input speech, and the distance value between the adapted input speech and the reference speech in the section is determined, Since the spectrum parameters of the reference speech in the optimum interval determined based on it are converted, it is possible to reduce the effects of discontinuity in the time direction and phonological blur in the conventional codebook mapping method. Thus, it is possible to realize a voice quality adaptation device that enables high-quality editing / synthesis.

In addition, according to the present invention, the distance value of the spectrum is calculated between the adapted input speech and all the cut out reference speeches, and the distance value judgment is performed to determine the optimal reference speech of the partial section. However, since the conversion is performed to the spectrum parameter of the reference speech in that section, it is possible to reduce the influence of the discontinuity in the time direction and the phonological blur in the conventional codebook mapping method, and it is possible to obtain high quality. A voice quality adaptation device capable of editing and synthesizing can be realized.

[Brief description of drawings]

FIG. 1 is a configuration diagram showing a first embodiment of the present invention.

FIG. 2 is a configuration diagram showing a second embodiment of the present invention.

FIG. 3 is a configuration diagram showing a third embodiment of the present invention.

FIG. 4 is a configuration diagram showing a fourth embodiment of the present invention.

FIG. 5 is a configuration diagram showing an example of a conventional spectrum adaptation method.

FIG. 6 is an explanatory diagram of an operation of DP matching.

[Explanation of symbols]

1 reference speech spectrum parameter, 2 input speech spectrum parameter, 3 spectrum adaptation means, 4 adaptation input speech spectrum parameter, 5 labeling means, 6 reference speech phoneme label sequence, 7 adaptation input speech phoneme label sequence, 8 partial section determination Means, 9 reference speech partial section, 10 input speech partial section, 11 clipping means, 12
Partial reference speech spectrum parameter, 13 Partial adaptation input speech spectrum parameter, 14 Distance value calculating means, 15 Spectral parameter distance value, 16 Optimal interval determining means, 17 Optimal reference speech partial interval, 18 Transforming means, 19 Transforming speech spectrum parameter, 20 input speech phoneme label sequence, 21 partial input speech spectrum parameter, 22 reference speech vector quantization codebook, 23 adaptive input speech vector quantization codebook, 24 vector quantization means, 25 reference speech vector quantization code number sequence, 26 adaptive input speech vector quantization code number sequence, 27 code number sequence cutting-out means, 28 reference speech partial vector quantization code number sequence,
29 input speech partial vector quantization code number sequence, 30
Input voice cutout means, 31 Reference section determination means.

Claims (4)

[Claims] 1. A voice quality adaptation device for adapting a reference voice and an input voice having different utterances or different utterance times to bring the input voice closer to the voice quality of the reference voice, using learning data for each frame of the input voice. Spectrum adapting means for adapting the spectrum feature; and first and second phoneme models for outputting phoneme labels including phoneme boundary information for the adapted input speech and the reference speech output from the spectrum adapting means. 2 labeling means, determining an input speech subsection from the phoneme label string of the adapted input speech given by the first labeling means, and from the phoneme label string of the reference speech given by the second labeling means. Partial segment determining means for detecting a segment that matches the phoneme label string of the input speech partial segment, and the partial segment determining means First and second clipping means for clipping each of the adaptive input speech and the reference speech based on the output partial section information, and the adaptive input speech clipped by the first and second clipping means Distance value calculating means for obtaining the distance value between the reference voices, optimum section determining means for selecting the optimum section of the reference voice based on the determination result of the distance value obtained by the distance value calculating means, and cutout by the first cutout means. And a conversion means for converting the adapted input speech to a reference speech in the optimum section determined by the optimum section determination means. 2. The voice quality adaptation apparatus according to claim 1, wherein an input voice is used instead of the adapted input voice as an input to the first labeling means and the clipping means. 3. A voice quality adaptation device for adapting a reference voice and an input voice having different utterances or different utterance times to bring the input voice close to the voice quality of the reference voice, wherein a reference voice vector quantization codebook and an adaptive input are provided. First and second vector quantization means for performing vector quantization of the reference speech and the input speech for each frame by the speech vector quantization codebook, and the input speech vector quantum output from the first vector quantization means. Of the input voice partial vector quantized code sequence, the input voice partial vector quantization code number sequence is output, and the reference voice vector quantized code number sequence output from the second vector quantization means is input voice partial vector quantum. Code number string cutout that detects the section that matches or is similar to the code number string A step, a distance value calculation means for obtaining a distance value between an input voice and a reference voice for a partial code number sequence output from the code number sequence cutout means, and a determination of the distance value obtained by the distance value calculation means. Optimal interval determining means for selecting an optimal interval of the reference speech, and conversion of the input voice partial vector quantized code number sequence output by the code number sequence extracting means into reference speech in the optimal interval determined by the optimal interval determining means. A voice quality adaptation device comprising: a conversion unit. 4. A voice quality adaptation device for adapting a reference voice and an input voice having different utterances and different utterance times to bring the input voice closer to the voice quality of the reference voice, by adapting the spectrum feature of each frame of the voice. A spectrum adapting means for performing, an input speech clipping means for segmenting the adaptive input speech output from the spectrum adapting means, and all of the adaptive input speech and the reference speech clipped by the input speech clipping means A reference section determining means for determining a distance section and a distance section for determining an optimum section of the reference speech by the distance value determination, and the adapted input speech cut out by the input speech cutting section are determined by the reference section determining section. A voice quality adaptation device, comprising: a conversion means for converting into a reference voice in an optimum section.