JPH04147300A - Speaker's voice quality conversion and processing system - Google Patents

Abstract

PURPOSE:To enable high quality voice conversion by executing vectorization of plural rhyme in a frequency axis, obtaining a corresonding path by means of matching according to a dynamic planning method and utilizing this correspondence for frequency conversion. CONSTITUTION:Plural rhymes vectorization means 1 and 2 obtains a spectral envelope at maximum power for each of plural number of rhyme included in voice produced by each speaker. A corresponding path extracting means 3 obtains the corresponding path according to the dynamic planning method by matching the spectral envelope for the two rhyme corresponding to each other by means of the dynamic planning method. A frequency conversion means 4 executes the frequency conversion for each rhyme included in the sound of a speaker A by utilizing the corresponding path. Thus, since nonlinear conversion is executed for the spectral envelope for plural number of rhyme, the voice conversion in higher quality can be obtained.

Description

[Detailed Description of the Invention] [Summary] The voices of two speakers with different voice qualities are matched on the frequency axis using the dynamic programming method (hereinafter referred to as DP matching), and the corresponding dynamic programming extraction bus is (hereinafter referred to as DP Babasu).

This invention relates to a speaker voice quality conversion processing method that converts the voice quality of one speaker to the voice quality of another speaker.

The purpose of this invention is to enable higher-quality voice quality conversion when converting the voice quality of a speaker.

DP matching is performed on a vector sequence of spectral patterns of multiple phonemes on the frequency axis, the DP path is defined as a correspondence between speakers on the frequency axis, and voice quality conversion is performed using this correspondence. The spectral pattern is used as a vector sequence in the frequency domain.

[Industrial application field]

D on the frequency axis for the voices of two speakers with different voice qualities
The present invention relates to a speaker voice quality conversion processing method that performs P matching to convert the voice quality of one speaker to the voice quality of another speaker based on the corresponding DP Babas.

Such voice quality conversion processing is used for speaker adaptation in speech recognition and voice quality conversion by voice synthesis.

[Conventional technology]

FIG. 2 shows the configuration of a conventional case. Reference numeral 11 in FIG. 2 is a control unit that performs control to execute the process of obtaining a spectrum envelope and the process of performing DP matching; 12.
14 are maximum power point spectrum envelope extraction units; 13
．． 15 is an envelope storage unit that stores the obtained spectrum envelope; 16 is a frequency axis DP (dynamic programming) unit that performs DP matching on two spectrum envelopes to obtain a DP bus; 17 is a frequency conversion table that holds information for frequency conversion based on the obtained DP Babas 51
Reference numeral 8 denotes a frequency conversion unit that nonlinearly converts the spectrum envelope of speaker A's voice.

First, the control unit 11 selects speaker A. A single vowel (for example, /a/) of speaker A is read, and the maximum power point spectrum envelope extraction unit AI2 extracts the utterance of the single vowel (for example, /a/). Extract the spectral envelope at the maximum power point. This spectral envelope is stored in the envelope storage section A13.

Next, the control unit 11 reads the monophthong of speaker B (e.g. /a/) for selecting speaker B, and extracts the same monophthong as speaker A (e.g. /a/) at the maximum power point spectral envelope extraction unit B14. The spectrum envelope at the time when the power of the utterance a/) is maximum is extracted. This spectral envelope is stored in the envelope storage section B15. Maximum power point spectrum envelope extraction unit B14
The operations of the envelope storage section B15 and the maximum power point spectrum envelope extraction section A12 and the envelope storage section A13 are the same.

Next, the control unit 11 selects DP. As a result of this instruction, the one frequency axis DP section 16 has an envelope storage section A13 and an envelope storage section B.
15, the spectral envelope of the same single vowel of speaker A and speaker B is read and DP matching is performed in 9 frequency domains. Through this DP pass, the correspondence relationship between speaker A and speaker B in the frequency domain Seek. This correspondence relationship in the frequency domain is written into the frequency conversion table 17.

When the spectral envelope of the 9 speakers A is input, the frequency conversion unit 18 nonlinearly expands and contracts the spectral envelope of the 9 speakers by referring to the correspondence relationship in the frequency domain written in the 1-frequency conversion table 17. , outputs the spectral envelope after voice quality conversion.

[Problem to be solved by the invention]

Since the phoneme used in the past was a single phoneme (for example, /a/), there was a high possibility that the correspondence relationships in the frequency domain for other phonemes would be significantly different. Therefore.

There was a need for a voice quality conversion method that would provide appropriate correspondence in the frequency domain for other tones.

An object of the present invention is to perform voice quality conversion of five speakers with higher quality.

[Means to solve the problem]

Figure 1 shows the principle configuration diagram of the present invention with reference numeral 1.2 in Figure 0.
3 represents a vectorization means for a plurality of phoneme spectra, 3 represents a corresponding bus extraction means for obtaining a corresponding bus by DP matching on the frequency axis, and 4 represents a frequency conversion means using the corresponding bus.

The plurality of phoneme spectrum vectorization means 1.2 obtains the spectral envelope at the maximum power point for each of the plurality of sounds #I in the speech uttered by each speaker.

The corresponding bus extraction means performs DP matching on the spectrum envelopes of two corresponding phonemes.

Find the corresponding DP bus. Then, the frequency conversion means 4 using the corresponding bus performs frequency conversion on each phoneme in the speech of the two speakers A using the corresponding path.

[For production]

Means 1 for vectorizing the speech of speaker A into a plurality of phoneme spectra
The speech of speaker B is input to the multi-phonetic spectrum vectorization means 2 to generate a vector sequence in the frequency domain according to the spectral pattern. The means 3 for determining a corresponding bus by performing DP matching on the vector series on one frequency axis determines the correspondence relationship in the frequency II range of both voices. Using this correspondence relationship in the frequency domain, the frequency conversion means 4 according to the corresponding path reads the frequency characteristics of speaker A.

Outputs frequency features after voice quality conversion.

That is, the respective phonemes AI and Bi of speaker A and speaker B correspond to each other on the frequency axis as shown in the figure 101, 102, 103, . . .
If the sound is as follows, then the sound fi of speaker A is
The frequency of point P on Ai is expressed as point P' on speaker B's sound 11Bi.
is converted to the frequency of

C Embodiment] FIG. 3 shows the configuration of an embodiment of the present invention. The reference numeral 21 in FIG. 3 is a control section corresponding to the control section 11 shown in FIG. Envelope extraction sections 23 and 29 are storage section selection sections for selecting locations where the extracted spectral envelopes corresponding to each phoneme are to be stored.

Reference numerals 24 to 26 and 30 to 32 are respectively envelope storage units corresponding to the envelope storage units 13 and 15 shown in FIG. 2, and 27 and 33 are vector generation units, respectively. Alternatively, the spectrum envelope stored in 30 to 32 is extracted and the frequency axis DP section 34
34 is a frequency axis DP unit that performs DP matching on the frequency axis for a given spectrum envelope, and 35 is a frequency conversion table that holds conversion tables corresponding to a plurality of phonemes. , 36 represents a frequency conversion unit that performs frequency conversion for each of a plurality of phonemes.

First, one of the monophthongs of speaker A (for example, /a/) for selecting speaker A is read by the control unit 21, and the spectrum envelope extraction unit A22 selects it at the maximum power point.

The spectrum envelope at the time when the power of the utterance of this monophthong (for example, /a/) is maximum is extracted. This spectral envelope is selected by the storage section selection section A23 for the input vowel (for example, /a
The next single vowel (for example, /i/) of speaker A stored in the envelope storage unit A124 corresponding to /) is read, and the maximum power point spectrum envelope extraction unit A22 extracts this single vowel. The svector envelope at the time of maximum vocalization power of a vowel (for example, /l/) is extracted. This spectral envelope is determined by the memory selection unit A23 for the input vowel (for example, /l/
) is stored in the envelope storage unit A225. In the same way, up to n other single vowels are input, and the envelope memory unit A
Stored by n 26. Note that n here is the basic number of vowels of the language used.

Next, the control unit 21 selects speaker B. The monophthong that is read is the voice uttered by speaker B.

The subsequent operations are the same as those when speaker A is selected. Each of the components 28 to 32 corresponds to the components 22 to 26 described above.

Next, the control unit 21 selects to perform DP pine seventing. In response to this instruction, the 1-frequency axis DP section 34 converts and reads n spectral patterns from the envelope storage section A124 to the envelope storage section An 26 through the vector generation section A27 into an n-dimensional vector series in the frequency domain, Through this DP path, n spectral patterns are converted into an n-dimensional vector sequence in the frequency domain from the envelope storage unit B130 to the envelope storage unit Bn32, and DP matching is performed in each of the two frequency domains. The correspondence relationship with B in the frequency domain is determined. This correspondence relationship in the frequency domain is written into the frequency conversion table 35.

When the spectral envelopes of 9 speakers A are input, the frequency converter 36 non-linearly expands and contracts the spectral envelopes of 1 speaker A with reference to the correspondence relationships in the frequency domain written in the 9 frequency conversion table 35. , outputs the spectral envelope after voice quality conversion.

In the above explanation, it is assumed that the spectral envelopes are extracted for a plurality of phonemes, but in the case of the present invention, (i) the spectral envelopes are found for a plurality of stationary bases, or (it) the spectral envelopes are found for a plurality of basic bases. It is more effective to However, it is of course not limited to this.

〔Effect of the invention〕

As explained above, according to the present invention, since the spectral envelopes of multiple phonemes are nonlinearly transformed, it is possible to perform voice quality conversion with higher quality than in the conventional case. .

[Brief explanation of drawings]

FIG. 1 shows the principle configuration of the present invention, FIG. 2 shows the conventional configuration, and FIG. 3 shows the configuration of an embodiment of the present invention. In the figure, l and 2 are means for vectorizing multiple phoneme spectra;
Is 3 the DP on the frequency axis? 4 represents means for determining a corresponding bus by searching, and 4 represents a frequency conversion means using the corresponding bus. Patent expert Chikujuhi Co., Ltd. agent patent attorney Mori 1) Hiroshi (and 2 others) 3L 118r ton example history

Claims (4)

[Claims] (1) The spectral envelope of the phoneme is extracted in response to the pronunciation pronounced by speaker A, and the spectral envelope of the phoneme is extracted in response to the same pronunciation pronounced by speaker B. Then, the spectral envelopes of the same phoneme for both speakers are matched using dynamic programming in the frequency domain, and based on the dynamic programming extraction path obtained in the matching, the spectral envelopes of both speakers in the frequency domain are A speaker voice quality conversion processing method that calculates the correspondence, non-linearly expands and contracts the spectral envelope of the voice uttered by speaker A, and converts the voice quality into a form that corresponds to the spectral envelope of the voice uttered by speaker B. a plurality of phoneme spectrum vectorization means (1, 2) for extracting respective spectral envelopes for a plurality of phonemes corresponding to the speaker A and the speaker B; and for each of the extracted phonemes. , a corresponding path extraction means (3) that performs dynamic programming matching in the frequency domain and obtains the dynamic programming extraction path; and a spectral envelope corresponding to each of the phonemes in the speech uttered by speaker A. Frequency conversion means (4
), and converts the voice quality of the voice uttered by speaker A to the voice quality of speaker B. (2) The dynamic programming extraction path for each phoneme obtained by the corresponding path extraction means (3) is calculated using the frequency conversion table (35
), and the frequency conversion means (4) has the frequency conversion table (35
2. The speaker's voice quality conversion processing method according to claim 1, wherein the contents of the speaker's voice are read out and used. (3) The frequency conversion means (4) non-linearly expands and contracts the spectral envelope corresponding to each phoneme in the speech uttered by speaker A, based on the contents of the frequency conversion table (35). The speaker's voice quality conversion processing method according to claim (2). (4) The speaker's voice quality conversion processing method according to claim (1), wherein the plurality of phonemes correspond to a plurality of vowels.