JP3302075B2 - Synthetic parameter conversion method and apparatus - Google Patents

Abstract

PURPOSE:To use parameters as they are by conducting a simple transformation of cepstrum.parameters even though the smapling frequency is different during a voice synthesis. CONSTITUTION:F1 [Hz] synthesis cepstrum.parameters, which are transformation object to F2 [Hz] synthesis cepstrum.parameters, are transformed into a logarithmic power.spectrum by conducting a discrete Fourier transform for every frame by a Fourier transformation processing section 1. The logarithmic power.spectrum is inputted to a spectrum data number changing section 2. If F1>F2, the data corresponding to a high frequency side that is higher than F2/2 [Hz] of the spectrum are discarded. If F1<F2, the data with small constant values are added to the high frequency side which is higher than F1/2 [Hz] of the spectrum. The power.spectrum, to which a data number is added or reduced, is discrete Fourier inverse transformed in a Fourier inverse transformation processing section 3 and F2 [Hz] synthesis cepstrum.parameters are obtained.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech synthesizing method using a cepstrum as a phoneme parameter, and more particularly to a method for converting cepstrum parameters suitable for synthesizing speech at a sampling frequency different from that at the time of sampling and analyzing a speech signal. And equipment.

[0002]

2. Description of the Related Art Voice cepstrum, as is well known,
The voice uttered by a person is converted from analog to digital (A / D) and digitized, and a short-time window such as a hamming window is applied to the digital voice data while being shifted at a fixed period, and the voice data in each window is applied. Is obtained by performing logarithmization on a spectrum obtained by performing Fourier transform of, and then performing inverse Fourier transform.

[0003] Among the obtained cepstrum, the higher-order cepstrum stores the pitch component of the voice power spectrum, and the lower-order cepstrum stores the envelope component of the voice power spectrum. Cepstrum is widely used for analyzing and synthesizing speech and synthesizing rules of speech described below by utilizing such properties.

[0004] First, in the speech analysis and synthesis, the speech uttered by an announcer or the like is cepstrum-analyzed by the above-described method, and the low-order components of each frame of the obtained cepstrum parameters are determined together with the voiced / unvoiced information and pitch frequency of each frame. save. Then, at the time of synthesis, the information is input to a synthesizer configured with an LMA (logarithmic amplitude approximation) filter or the like, and the synthesized voice is D / A converted and output as a voice.

Next, in the rule synthesis of speech, cepstrum analysis is performed in advance on a voice uttered by an announcer or the like by the above-described method, and a cepstrum analysis frame corresponding to a basic unit of rule synthesis, for example, a CV (consonant + vowel chain) unit. Then, a low-order cepstrum holding the envelope information of the voice is extracted and stored in the storage device as a voice unit.
Then, when synthesizing speech, the speech units in CV units are interpolated and connected according to the phoneme sequence to be synthesized, and this is used as a phoneme parameter, and is configured by an LMA filter or the like together with a prosody parameter consisting of pitch information generated on the other side. It is input to the synthesizer, and the synthesized voice is D / A converted and output as voice.

[0006]

As described above, techniques for analyzing and synthesizing speech using cepstrum and rules for synthesizing speech have been known.

However, conventionally, for example, when the processing speed of the synthesizer processing system is slow and the sampling frequency of the synthesized speech signal output from the synthesizer is lowered to reduce the number of operations per unit time, cepstrum parameter A at the time of creation
If the sampling frequency of the D / A conversion differs from the sampling frequency of the D / A conversion during speech synthesis,
The cepstrum parameter analyzed by the / D conversion cannot be used as it is. For example, in the case of analysis and synthesis, one of the following two methods has to be performed to regenerate audio data composed of cepstrum parameters. (1) The voice is A / D converted again at a desired sampling frequency, and the cepstrum analysis is performed again.

(2) After appropriate sampling of the A / D-converted discrete signal data, the signal is passed through a low-pass filter so that aliasing does not occur, and cepstrum analysis is performed again.

In any of the above methods (1) and (2), since the cepstrum parameters that have been created first cannot be used, A / D conversion and decimation processing of speech, re-execution of cepstrum parameter analysis, and the like are performed. Must be done, inefficient.

Also, taking the case of rule synthesis as an example, in addition to the labor of (1) or (2) in the above-mentioned analysis synthesis,
It is necessary to redo a speech unit (extraction of a low-order cepstrum of a cepstrum frame corresponding to a necessary speech portion), which is even more troublesome.

It is an object of the present invention to provide a method and apparatus for synthesizing parameters which can be used as they are even by simple conversion of cepstrum parameters, even if the sampling frequency during speech synthesis is different.

[0012]

According to the present invention, a first cepstrum parameter obtained by analyzing voice signal data sampled at a first frequency is discrete-Fourier-by-frame for each frame. The power spectrum is converted into a power spectrum by the conversion, and the power spectrum whose number of data is increased or decreased by discarding data from the highest frequency side of the power spectrum or adding data to the highest frequency side is inversely transformed by a discrete Fourier transform. And obtaining a second cepstrum parameter for speech synthesis at a second frequency different from the second frequency.

[0013]

In the above configuration, A / D conversion is performed at the first frequency.
When it is desired to obtain second cepstrum parameters for speech synthesis at a second frequency from first cepstrum parameters of speech obtained by conversion and cepstrum analysis, first the first cepstrum parameters must be After converting into a power spectrum by the discrete Fourier transform for each frame, data may be discarded from the highest frequency side of the power spectrum (in the case of first frequency> second frequency) or from the lower frequency part to the highest frequency side. By adding data of a small value (when the first frequency <the second frequency) and increasing or decreasing the number of data of the power spectrum, a new power spectrum in which the frequency domain is reduced or expanded can be obtained. , By performing an inverse discrete Fourier transform of this new power spectrum, the second frequency (sampler It is possible to obtain the second cepstral parameters comparable to those analyzed in grayed frequency).

As described above, the cepstrum parameters obtained by analyzing the cepstrum once A / D converted once are converted into cepstrum parameters equivalent to the cepstrum parameters analyzed by A / D conversion at different sampling frequencies.
Since the parameters can be obtained, there is no need to perform resampling, thinning, low-pass filtering, reanalysis, and the like, and the speech can be converted from the first speech unit, so that there is no waste and the efficiency is extremely high.

[0015]

An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing a configuration of a synthesizing parameter (cepstrum parameter) converter according to the embodiment.

The apparatus shown in FIG. 1 converts a cepstrum parameter (F1 [Hz] synthesis cepstrum parameter) obtained by analyzing audio signal data A / D converted at a first sampling frequency F1 [Hz]. , For converting to a cepstrum parameter for speech synthesis (F2 [Hz] cepstrum parameter for synthesis) at a second sampling frequency F2 [Hz] (for converting a synthesized sound sampling frequency),
Fourier transform processing unit 1, spectrum data number changing unit 2,
And an inverse Fourier transform processing unit 3. The Fourier transform processing unit 1 transforms the cepstrum parameter for F1 [Hz] synthesis into a power spectrum by discrete Fourier transform for each frame.

The number-of-spectral-data changing unit 2 determines whether to cut off data from the highest frequency side (F1> F2) for the power spectrum converted and output from the Fourier transform processing unit 1 according to the magnitude relationship between F1 and F2. ) Or add data to the highest frequency side (if F1 <F2) and perform a data number change process.

The inverse Fourier transform processing unit 3 performs an inverse discrete Fourier transform of the power spectrum whose number of data has been increased / decreased by the spectrum data number changing unit 2 to generate a cepstrum parameter for F 2 [Hz] synthesis.

Next, the operation of the synthesizing parameter conversion device configured as shown in FIG. 1 will be described. The cepstrum parameters for F1 [Hz] synthesis are the m-th order (where m <256) cepstrum parameters C ₀ -C. An example in which the parameter is converted to a cepstrum parameter for F2 [Hz] synthesis assuming _m is described with reference to the flowchart in FIG.

The operation of the apparatus shown in FIG. 1 (of which the number of spectrum data changing unit 2 is different) differs depending on the magnitude relationship between F1 and F2. Therefore, the operation when F1> F2 will be described first with reference to the operation explanatory diagram of FIG.

First, the Fourier transform processing unit 1 analyzes m-th order cepstrum parameters C ₀ -C obtained by analyzing the audio signal data A / D converted at the sampling frequency F 1 [Hz].
C _m ,

[0022]

(Equation 1)

Thus, the array X is set (step S1). Then, the Fourier transform processing unit 1 calculates the array X
Is subjected to a discrete Fourier transform of 256 points to obtain a logarithmic power spectrum (step S2). At this time, a well-known fast Fourier transform (FET) algorithm is used.

When the Fourier transform processing unit 1 performs a Fourier transform on the array X, for example, from the m-th order cepstrum parameter shown in FIG. 3 (a), as shown in FIG. F1 from [Hz] to Nyquist frequency
A logarithmic power spectrum in the range of / 2 [Hz] is obtained in a folded form on the high frequency side. When the logarithmic power spectrum is obtained by the Fourier transform processing unit 1, the spectrum data number changing unit 2 is activated.

The number-of-spectral-data changing unit 2 firstly selects F1
And F2 is determined (step S3). If F1> F2 as in this example, the logarithmic power spectrum, that is, the logarithmic power spectrum folded back as shown in FIG. The following processing is performed on the spectrum.

FIG.
In the folded logarithmic power spectrum shown in (b), the data corresponding to the higher frequency side than F2 / 2 [Hz] (the data in the hatched portion in the figure) is discarded, and FIG. The number of data points is 256 × F2 / F1 as shown in FIG.
It is reduced to points (step S4). The spectrum data shown in FIG. 3C with the reduced number of data points is 0 [H
The logarithmic power spectrum in the frequency range from z] to F2 / 2 [Hz] is folded on the high frequency side. this is,
A / D conversion at the sampling frequency F2, 256 × F2
This is almost the same as the logarithmic power spectrum obtained by performing a discrete Fourier transform by windowing the / F1 point.

As described above, when F1> F2, data corresponding to a higher frequency side than F2 / 2 [Hz] of the logarithmic power spectrum obtained by the Fourier transform processing unit 1 shown in FIG. The data points are thinned out by the spectrum data number changing unit 2 and the number of data points is reduced to 25 as shown in FIG.
It is reduced to 6 x F2 / F1 points. When a new logarithmic power spectrum whose number of data points is reduced by the Fourier transform processing unit 1 is obtained, the Fourier inverse transform processing unit 3 is activated.

The inverse Fourier transform processing unit 3 determines that the number of data points is reduced to 256 × F2 / F1 points in FIG.
The inverse Fourier transform of the logarithmic power spectrum shown in (c) is performed (step S6). As a result, a new cepstrum parameter as shown in FIG. 3D equivalent to that obtained by sampling at the sampling frequency F2 [Hz], that is, a cepstrum parameter for F2 [Hz] synthesis is obtained. Next, the operation when F1 <F2 will be described with reference to the operation explanatory diagram of FIG.

The operation in the case of F1 <F2 is different from the operation in the case of F1> F2 in that the number of data points is increased by the spectrum data number changing unit 2.
Operation of Other Fourier Transform Processing Unit 1 (Steps S1, S1
2) and the operation of the inverse Fourier transform processing unit 3 (step S6)
Is the same as in the case of F1> F2. For this reason, if the cepstrum parameter for F1 synthesis is the same as the cepstrum parameter in FIG. 3A as shown in FIG. 4A, the logarithmic power spectrum obtained by the Fourier transform processing unit 1 is also obtained. As shown in FIG.
This is the same as the power spectrum of (b).

When the Fourier transform processing unit 1 obtains the logarithmic power spectrum shown in FIG. 4B, the spectrum data number changing unit 2 changes the logarithmic power in the range from 0 [Hz] to F1 / 2 [Hz]. When a logarithmic power spectrum with a spectrum folded on the high frequency side is obtained, F1 and F2
Is determined (step S3).

In the case of F1 <F2 as in this example, the spectrum data number changing section 2 sets the F1 / 2 [H of the folded logarithmic power spectrum as shown in FIG. 4B.
z], on the higher frequency side, for example, add data of a constant value smaller than that of the low frequency part as log power, and increase the number of data points to 256 × F2 / F1 points as shown in FIG. ). The spectrum data shown in FIG. 4C in which the number of data points is increased is 0 [Hz].
The logarithmic power spectrum in the frequency range from F2 to F2 [Hz] is folded on the high frequency side. The A / D conversion is performed on the sound that has passed through a steep low-pass filter having a frequency of F1 [Hz] at a sampling frequency of F2, and 256 × F2
This is equivalent to a logarithmic power spectrum obtained by performing a discrete Fourier transform with a window of / F1 point.

As described above, when F 1 <F 2, the low-frequency portion is shifted to a higher frequency side than F 1/2 [Hz] of the logarithmic power spectrum obtained by the Fourier transform processing section 1 and shown in FIG. Is added by the spectrum data number changing unit 2, and the number of data points is increased to 256 × F2 / F1 points as shown in FIG.

Therefore, FIG. 4 in which the number of data points is increased
The logarithmic power spectrum shown in (c) is subjected to discrete Fourier inverse transform by the inverse Fourier transform processing unit 3, thereby obtaining a sample at the sampling frequency F2 [Hz] and analyzing it as shown in FIG. F2 [Hz] Cepstrum parameters for synthesis can be obtained. A case where the above-described synthesis parameter (cepstrum parameter) conversion is applied to analysis and synthesis will be described with reference to FIGS. 5 (a) and 5 (b). FIGS. 5A and 5B show a block configuration of the speech synthesizer.

First, the speech synthesizer shown in FIG. 5A includes a storage unit 11, a sound source generation unit 12, a speech synthesis operation unit 13,
2 kHz D / A converter 14 and low-pass filter 15
And a well-known configuration composed of

In the storage unit 11 of the speech synthesizer, a speech read out from a sentence or word is sampled at 12 kHz,
For example, a 20th-order (m = 20) cepstrum parameter (12 kHz synthesis cepstrum parameter) obtained by analysis is stored in the cepstrum parameter storage unit 111.
The voiced / unvoiced information to be synthesized is a voiced / unvoiced information storage unit 11
2 is stored in the pitch information storage unit 113. Then, based on the contents of the voiced / unvoiced information storage unit 112 and the contents of the pitch information storage unit 113, the sound source generation unit 1
2, sound source data is generated, and from the sound source data and the contents of the cepstrum / parameter storage unit 111, voice signal data is generated by a voice synthesis calculation unit (LMA filter calculation unit) 13 configuring an LMA filter. The D / A converter 14 converts this audio signal data into a 12 kHz D / A signal.
A conversion is performed and the signal is passed through a low-pass filter 15. In this way, a 12-kHz sampling voice is synthesized.

On the other hand, the speech synthesizer shown in FIG.
It has a configuration similar to that of the voice synthesizer of FIG. 1A, and includes a storage unit 21, a sound source generation unit 22, a voice synthesis operation unit 23,
It comprises a D / A converter 24 and a low-pass filter 25. However, the speech synthesizer of FIG.
Since the operation speed of the speech synthesis operation unit (LMA filter operation unit) 23 is slow, the operation time is extremely long when attempting to synthesize a 12 kHz audio signal.
A 6 KHz audio signal that requires only half the number of calculations per unit time is synthesized. For this reason, D in FIG.
Unlike the / A converter 14, a D / A converter 24 that performs 6 kHz D / A conversion is used.

When it is desired to synthesize the same content as the voice stored in the storage unit 11 shown in FIG. 5A by such a voice synthesizer as shown in FIG. A synthesis parameter conversion device may be used. That is, F1 and F2 in the flow chart (algorithm) of FIG. 2 are calculated as follows: F1 = 12 [kHz] F2 = 6 [kHz]

As described above, by the synthesis parameter conversion device,
The cepstrum parameter analyzed from the voice data of the sampling of 12 kHz is converted to the sampling frequency of 6 kHz.
Cepstrum equivalent to that sampled and analyzed in
The parameters may be converted and stored in the cepstrum / parameter storage unit 211 of the speech synthesizer in FIG. As mentioned above, although one Example of this invention was described, this invention is not limited to said Example. For example, regarding the Fourier transform, the Fourier transform of 256 points is used in the above embodiment, but the number of data points and the algorithm of the Fourier transform are not limited at all.

In the above embodiment, when F1 <F2, as shown in FIG. 4C, the spectrum data number changing unit 2 changes the logarithmic power to the highest frequency side (the center of the spectrum data). A case has been described where the number of data points is increased by adding data of a fixed value that is smaller than that of the low frequency portion. However, as shown in FIG. 6, for example, as shown in FIG. (In the case of FIG. 6A) or a peak may be added (in the case of FIG. 6B).

Although the application example shown in FIG. 5 is an example of speech analysis and synthesis, it is a speech synthesis method using cepstrum parameters as speech segments, and can be used for speech segment conversion. It is. In short, the present invention can be variously modified and implemented without departing from the gist thereof.

[0041]

As described above, according to the present invention, the cepstrum parameter obtained by analyzing the audio signal data is converted into a power spectrum by a discrete Fourier transform for each frame. Since the power spectrum whose number of data has been increased or decreased by truncating the data from the highest frequency side or adding data to the highest frequency side is inversely transformed by discrete Fourier transform, cepstrum parameters sampled and analyzed at different frequencies The cepstrum parameter equivalent to can be obtained.

Therefore, according to the present invention, it is necessary to reduce the number of operations per unit time due to, for example, the inherent processing speed of the synthesizer processing system. From the cepstrum parameters obtained by D-conversion and cepstrum analysis, necessary frames are cut out, cepstrum parameters equivalent to cepstrum parameters sampled and analyzed at different frequencies (target frequencies) can be easily obtained. Efforts such as re-sampling and thinning, low-pass filtering, analysis / slicing-out, and the like are eliminated, and there is no waste since conversion can be performed from the initially created cepstrum parameters.

[Brief description of the drawings]

FIG. 1 is a block diagram showing a configuration of a synthesis parameter (cepstrum parameter) conversion device according to an embodiment of the present invention.

FIG. 2 is a flowchart showing a procedure of a synthesis parameter (cepstrum parameter) conversion process for converting a synthesized sound sampling frequency by the apparatus of FIG. 1;

FIG. 3 is a diagram for explaining an operation when obtaining an F2 [Hz] synthesis cepstrum parameter having a lower frequency than the F1 [Hz] synthesis cepstrum parameter;

FIG. 4 is a diagram for explaining an operation in a case where an F2 [Hz] synthesis cepstrum parameter having a higher frequency than the F1 [Hz] synthesis cepstrum parameter is obtained;

5A and 5B are diagrams showing an example of application to analysis and synthesis. FIG. 5A is a block diagram of a speech synthesizer that synthesizes speech having a sampling frequency of 12 kHz, and FIG. FIG. 2 is a block diagram of a speech synthesizer for synthesizing.

FIG. 6 is a diagram showing a modification of the data addition method shown in FIG. 4 (c).

[Explanation of symbols]

DESCRIPTION OF SYMBOLS 1 ... Fourier-transform processing part, 2 ... Spectrum data number change part, 3 ... Fourier inverse transformation processing part, 13,23 ... Speech synthesis operation part, 14 ... 12kHz D / A conversion part, 24 ... 6kHz
zD / A converter, 111 ... 12 kHz synthesis cepstrum parameter storage, 211 ... 6 kHz synthesis cepstrum parameter storage.

Claims (8)

(57) [Claims] A first cepstrum parameter obtained by analyzing voice signal data sampled at a first frequency is converted into a power spectrum by a discrete Fourier transform for each frame, and the highest frequency of the power spectrum is obtained. A discrete cosine inverse transform of the power spectrum reduced in data number by truncating the data from the side to obtain a second cepstrum parameter for speech synthesis at a second frequency lower than the first frequency. Characteristic synthesis parameter conversion method. 2. A first cepstrum parameter obtained by analyzing audio signal data sampled at a first frequency is converted into a power spectrum by a discrete Fourier transform for each frame, and the highest frequency of the power spectrum is obtained. A discrete cosine inverse transform of the power spectrum whose data number has been increased by adding data to the second side to obtain a second cepstrum parameter for speech synthesis at a second frequency higher than the first frequency. Characteristic synthesis parameter conversion method. 3. Converting a first cepstrum parameter obtained by analyzing voice signal data sampled at a first frequency into a cepstrum parameter for voice synthesis at a second frequency different from the first frequency. Converting the first cepstrum parameter into a power spectrum by discrete Fourier transform for each frame, and if the first frequency is higher than the second frequency, By truncating the data from the highest frequency side of the power spectrum, the first frequency is
If the frequency is lower than the frequency, the data is added to the highest frequency side of the power spectrum to obtain a new power spectrum with the number of data increased / decreased, and the discrete power Fourier transform of the new power spectrum is performed. Obtaining the second cepstrum parameter according to: 4. The power spectrum according to claim 2, wherein a power value of data added to a highest frequency side of the power spectrum is smaller than a power value of a low frequency part of the power spectrum. The described synthesis parameter conversion method. 5. Fourier transform processing means for transforming a first cepstrum parameter obtained by analyzing audio signal data sampled at a first frequency into a power spectrum by discrete Fourier transform for each frame; Spectrum data number changing means for reducing the data number of the power spectrum by truncating the data from the highest frequency side of the power spectrum converted by the conversion processing means, and reducing the data number by the spectrum data number changing means. Inverse Fourier transform processing means for performing a discrete Fourier inverse transform of the power spectrum to obtain a second cepstrum parameter for voice synthesis at a second frequency lower than the first frequency. Synthesis parameter converter. 6. Fourier transform processing means for transforming a first cepstrum parameter obtained by analyzing audio signal data sampled at a first frequency into a power spectrum by discrete Fourier transform for each frame; A spectrum data number changing means for increasing the number of data of the power spectrum by adding data to the highest frequency side of the power spectrum converted by the conversion processing means, and the number of data is increased by the spectrum data number changing means. Inverse Fourier transform processing means for performing a discrete Fourier inverse transform of the power spectrum to obtain a second cepstrum parameter for voice synthesis at a second frequency higher than the first frequency. Synthesis parameter converter. 7. Converting a first cepstrum parameter obtained by analyzing voice signal data sampled at a first frequency into a cepstrum parameter for voice synthesis at a second frequency different from the first frequency. A Fourier transform processing means for transforming the first cepstrum parameter into a power spectrum by a discrete Fourier transform for each frame, wherein the first frequency is higher than the second frequency. If expensive
By adding data to the highest frequency side of the power spectrum converted by the Fourier transform processing means, if the first frequency is lower than the second frequency, data is added to the highest frequency side of the power spectrum. A spectrum data number changing means for increasing or decreasing the number of data of the power spectrum, and performing a discrete Fourier inverse transform on the power spectrum having an increased or decreased number of data by the spectrum data number changing means, An inverse Fourier transform processing means for obtaining cepstrum parameters. 8. A power value of data added to a highest frequency side of the power spectrum by the spectrum data number changing means is smaller than a power value of a low frequency portion of the power spectrum. The synthesis parameter conversion device according to claim 6 or 7.