JPH06250695A - Method and device for pitch control - Google Patents

Abstract

PURPOSE:To faithfully reproduce the change in a voice characteristic by making the formant power of a voice, which is synthesized while superimposing waveforms, closer to the one that corresponding to the pitch frequency of a target using the correlation between the pitch frequency of input voice signals and the formant power. CONSTITUTION:A formant extracting section 103 extracts formant of input voice from a spectrum envelope by cepstrum in few frames of a voice section. A local peak extracting section 104 extracts a peak position of the input voice, a spectrum envelope and phoneme display symbols or the like are outputted to a control coefficient setting section 106 and a control function is selected. Moreover, a pitch frequency that corresponds to each peak position, is computed from a local peak position. Pitch control coefficients and filter control coefficients are set from the pitch frequency of the input voice, the size of formant and the pitch control width inputted from a control value input section 105, outputted to a voice deformation section 107 and are controlled so that the size of formant is made closer to a target value employing a band pass filter.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and apparatus for pitch control of a speech waveform, and more particularly to controlling formant power using a bandpass filter to simulate the correlation between pitch frequency and formant power in natural speech. The present invention relates to a pitch control method and device capable of performing the above.

[0002]

2. Description of the Related Art As an example of conventional pitch control in speech synthesis technology, "Application of local peak extraction of speech waveform and pitch control" (October 1991, Proceedings of the Acoustical Society of Japan 2-6-1). There is a device described in. When pitch control of the voice waveform is performed as it is with this pitch control device, first, the local peak position is extracted from the input voice signal, and the window function for waveform cutout is determined in pitch units based on this peak position. A waveform is cut out by this window function, and the cut out waveforms are re-superposed so as to have a target pitch period to generate a synthetic voice.

[0003]

In such a conventional technique, since the pitch period of the voice cut out from the input voice is changed and the superposition is performed again, the conventional vocal tract even after the pitch period is changed. The properties are almost preserved. Therefore, the vocal tract characteristic corresponding to the pitch frequency cannot be reproduced, and there is a problem that the discomfort increases as the pitch control width increases.

The present invention has been made in view of the above problems, and pitch control for reproducing a vocal tract characteristic corresponding to a pitch frequency in accordance with an update of a pitch period of a voice waveform to generate a pitch control voice closer to a natural voice. It is to provide a method and an apparatus for implementing the method.

[0005]

According to the pitch control method of the present invention which achieves the above object, the pitch frequency of a synthesized voice is determined by cutting out the waveform of an input voice signal having a pitch structure and superimposing the cut out waveforms. In the method of controlling to the target value, the correlation between the pitch frequency of the input speech signal and the formant power is approximated to a polynomial function, and the formant power corresponding to the target value is determined from this polynomial function, and synthesized at the time of waveform superposition. The formant power of the voice to be reproduced is brought close to the determined formant power.

According to this method, a bandpass filter covering a plurality of formants, a pitch frequency calculating means for calculating a pitch frequency of an input voice signal and a control target value thereof, and a correlation between a pitch frequency and a formant power value for each phoneme. Function holding means for holding a polynomial function that approximates the relationship, function selecting means for selecting a specific polynomial function from the function holding means according to phoneme, selected polynomial function, the pitch frequency and its control target value Filter control means for controlling the gain of the bandpass filter based on the above. The bandpass filter is preferably composed of a plurality of bandpass filter groups whose gains are individually adjustable.

[0007]

Function: There is a correlation between the pitch frequency and the formant power value. Moreover, this correlation can be approximated by a quadratic or cubic polynomial function. The pitch control method and device of the present invention are focused on this point, and a polynomial function that approximates the correlation between the pitch frequency of a voice signal having a pitch structure and the formant power value, that is, the maximum value of the spectrum envelope. It is possible to reproduce the change in vocal tract characteristics due to pitch conversion by preparing in advance and associating this polynomial function with the pitch frequency of the input voice and controlling the gain of the bandpass filter. As a result, a feeling of strangeness in the synthesized voice when the pitch conversion width is increased is suppressed.

[0008]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block configuration diagram of a pitch control device according to an embodiment of the present invention. This pitch control device has a storage device (not shown) that can secure a work area necessary for processing the input voice signal and its processing.

In FIG. 1, reference numeral 101 denotes an input terminal, from which a voice signal and a phoneme notation symbol of the input voice signal are introduced from a preceding stage device. The pre-processing unit 102 converts the audio signal into digital data that matches the signal form of the processing unit of the present apparatus.
Also, it is determined from the phoneme notation symbols whether or not the voiced sound is pitch controllable. Here, the following pitch control processing is performed for voiced sound, and the pitch control processing is not performed for unvoiced sound. The pitch control process for voiced sounds will be described below.

The audio signal converted into digital data by the preprocessing unit 102 is guided to the formant extraction unit 103. The formant extraction unit 103 extracts the formant of the input voice from the spectrum envelope by the cepstrum in several frames of the voice section, and stores the spectrum envelope shape and the formant position in the storage device. Next, the local peak extraction unit 104
At, the peak position of the input voice is extracted by the peak search method and stored in the storage device. Spectral envelope calculated so far, formant position, local peak position,
The voice signal and the phoneme notation symbol are output to the control coefficient setting unit 106. The control coefficient setting unit 106 selects a control function for determining the control target value of the formant from the phoneme notation symbols. The selection criteria will be described later. Further, the pitch frequency corresponding to each peak position is calculated from the local peak position. Then, the pitch control coefficient and the filter control coefficient are set from the pitch frequency of the input voice, the size of the formant, and the pitch control width input from the control value input unit 105, and these are output to the voice transformation unit 107.

The voice transformation unit 107 transforms the pitch period of the voice waveform by the pitch synchronization type waveform superposition method. This modified waveform is controlled so that the size of the formant is close to the target value using a bandpass filter whose gain is determined based on the filter control coefficient set by the control coefficient setting unit 106. Then, the post-processing unit 108 performs conversion into analog data and outputs the result to the output terminal 109.

Next, a detailed description of the control coefficient setting unit 106 will be given with reference to FIG. Control coefficient setting unit 106
Is composed of a control function holding unit 201, a control function selection unit 202, a pitch calculation unit 203, a pitch conversion coefficient calculation unit 204, and a filter coefficient control value calculation unit 205.

The control function holding unit 201 holds a polynomial describing the correlation between the pitch frequency and the formant power for each phoneme from the first to the fourth formants. A set of correlations for each formant power for each phoneme is called a control function. The control function selection unit 202 selects a control function from the control function holding unit 201 based on the input phoneme symbol notation. Next, the pitch calculator 203 calculates the pitch frequency corresponding to each peak position from the information of the local peak position. The pitch conversion coefficient calculation unit 204 determines the peak interval ratio after pitch control based on the control width given from the control value input unit 105 and the local peak position of the input voice. This peak interval ratio is called a pitch conversion coefficient. The filter coefficient control value calculation unit 205 calculates the gain of the bandpass filter from the size of the formant of the input voice and the ideal value of the formant at the control target pitch frequency. This gain calculation processing is performed for each of the first to fourth formants, and this is used as the filter coefficient control value. The pitch conversion coefficient and the filter coefficient control value obtained here are guided to the audio transformation unit 107 (see FIG. 1) together with the audio signal and the like.

FIG. 3 is a detailed block diagram of the voice transformation unit 107. Hereinafter, the details of the voice transformation unit 107 will be described with reference to FIG. The waveform superposition position determination unit 301 calculates a target local peak position after pitch control from the local peak position of the input voice and the pitch conversion coefficient described above. At this time, the time length of the input voice is saved by repeating the peaks and the thinning process. The waveform superposition processing unit 302 performs superposition processing for each pitch based on the correspondence between the input voice and the local peak position of the target voice, and updates the pitch period (pitch period type waveform superposition method). Next, the filter coefficient setting unit 303 sets the filter gain of the band corresponding to each formant based on the previously determined filter coefficient. The filter processing unit 304 uses the bandpass filter whose gain is set by the filter coefficient setting unit 303, and processes the audio signal with the updated pitch period. The audio signal subjected to the above processing is output to the post-processing unit 108.

Next, details of the control function using the correlation between the pitch frequency and the formant power will be described with reference to FIGS. 4, 5 and 8.
Will be described with reference to. A correlation as shown in FIG. 8 is recognized between the pitch frequency and the formant power. This correlation can be approximated by a quadratic or cubic polynomial function. For example, the approximate expression of the plot illustrated in FIG. 8 is y
Where x is the power and x is the pitch frequency. y = -0.005x ² + 1.312x-7.040

FIG. 4 is an example in which this correlation is approximated by a quadratic polynomial function curve 401. Since this correlation differs depending on the speaker, phoneme, and formant (Nth formant), when applied to the pitch control device of this embodiment, the required correlation is analyzed in advance and approximated respectively. Find the polynomial function.

FIG. 5 is a conceptual diagram showing the correlation between the pitch and the formants for the first to fourth formants (F1 to F4) of the five vowels / a / to / o /. In the present embodiment, the approximation is performed by a quadratic or cubic polynomial function, but an appropriate approximation method or pattern table can be used depending on the behavior of the vocal tract characteristic to be analyzed.

A method of controlling the formant after the pitch control by using the above approximate expression function will be described below. In FIG. 4, the first plot point 402 is a plot of the corresponding position between the arbitrary pitch frequency fo of the input voice and the formant power Po at that time. On the other hand, the second plot point 403 is the control target point, and the target pitch frequency fa is calculated from the pitch frequency fo and the pitch control width described above. After calculating this target pitch frequency fa, this value fa is substituted into the variable x of the polynomial function expressing the correlation, and the corresponding formant power Pa at the target pitch is calculated. Since the formant power axis of the correlation is logarithmic, the gain Fn-gain of the filter controlling the formant is calculated by Pa-Po. Fn corresponds to the first to fourth formants,
Actually, it is expressed as F1-gain to F4-gain.

Details regarding the characteristics of the bandpass filter for controlling the formant size will be described with reference to FIGS. As shown in FIG. 6, the bandpass filter for controlling the formant is configured by combining a plurality of bandpass filters 602 each having a narrow passband and capable of adjusting individual gains. By doing so, the individual gains can be adjusted by an equalizer. In addition, when the whole is simply combined, ideally the passband has a flat characteristic 601.
Is equivalent to one bandpass filter with. When controlling the size of the formant, the gain of each bandpass filter is adjusted based on the previously calculated information on the formant position and the gain Fn-gain.

FIG. 7 is an explanatory diagram showing an example of ideal characteristics of a bandpass filter in which the gain is manipulated to control the formant. In the example of FIG. 7, an example of a filter setting 701 that emphasizes the low-frequency formant, a filter setting 702 that reduces the mid-range formant power, and an overall ideal characteristic 703 are shown.

As described above, according to the pitch control apparatus of the present embodiment, the pitch control is performed by the pitch synchronization type waveform superposition method, and when the formant power of the synthesized voice changes, the control value input unit 105 and the control coefficient setting are performed. Since it is corrected by the unit 106 and the voice transforming unit 107 at any time, it is possible to reproduce the change in the vocal tract characteristics due to the pitch conversion and generate a pitch control voice closer to a natural voice. The present invention is not necessarily limited to this embodiment. The procedure or configuration can be changed without departing from the spirit of the invention. For example, by making the model of the approximated polynomial function more precise, it becomes possible to simultaneously realize the amplitude control mechanism of the synthetic voice.

[0022]

As described in detail above, the pitch control method of the present invention utilizes the correlation between the pitch frequency of the input voice signal and the formant power, and the formant power of the voice synthesized at the time of waveform superposition. Is brought close to that corresponding to the target pitch frequency, so that the change in vocal tract characteristics due to the change in pitch frequency is faithfully reproduced.
Further, according to the pitch control device of the present invention, the gain of the bandpass filter that determines the combined formant is controlled by the output value of the filter control means based on the correlation, and the pitch control method is easily and reliably realized. Is effective. By configuring the bandpass filter with a plurality of bandpass filter groups whose gains can be adjusted individually, the vocal tract characteristics are reproduced more faithfully, and the naturalness of synthesized speech is improved.

[Brief description of drawings]

FIG. 1 is a block diagram showing the overall configuration of an embodiment of the present invention.

FIG. 2 is a detailed block diagram of a control coefficient setting unit of this embodiment.

FIG. 3 is a detailed block diagram of a voice transformation unit of the present embodiment.

FIG. 4 is an explanatory diagram showing an example in which a correlation between a pitch frequency and formant power is approximated.

FIG. 5 is an explanatory diagram showing an example of registration of correlations of first to fourth formants of five vowels.

FIG. 6 is an explanatory diagram showing ideal characteristics of a filter group used for formant control.

FIG. 7 is an explanatory diagram showing ideal characteristics of a filter whose gain is manipulated to control a formant.

FIG. 8 is an explanatory diagram showing a correlation between a pitch frequency analyzed from an actual voice signal and formant power.

[Explanation of symbols]

 102 ... Pre-processing unit 103 ... Formant extraction unit 104 ... Local peak extraction unit 105 ... Control value input unit 106 ... Control coefficient setting unit 107 ... Speech transformation unit 108 ... Rear Processing unit 201 ... Control function holding unit 202 ... Control function selection unit 203 ... Pitch calculation unit 204 ... Pitch conversion coefficient calculation unit 205 ... Filter coefficient control value calculation unit 301 ... Waveform superposition Position determination unit 302 ... Waveform superposition processing unit 303 ... Filter coefficient setting unit 304 ... Filter processing unit

Claims (3)

[Claims] 1. A pitch control method for controlling a pitch frequency of a synthesized voice to a target value by cutting out a waveform of an input voice signal having a pitch structure and superimposing the cut out waveforms. And formant power are approximated to a polynomial function, and the formant power corresponding to the target value is determined from this polynomial function, and the formant power of the voice synthesized at the time of the waveform superposition is the determined formant power. Pitch control method characterized in that 2. A device for realizing the pitch control method according to claim 1, wherein a band pass filter covering a plurality of formants, and a pitch frequency calculation means for calculating a pitch frequency of an input voice signal and a control target value thereof. A function holding means for holding a polynomial function that approximates the correlation between the pitch frequency and the formant power for each phoneme, and a function selecting means for selecting a specific polynomial function from the function holding means according to the phoneme. And a filter control means for controlling the gain of the bandpass filter based on the polynomial function and the pitch frequency and a control target value thereof. 3. The pitch control device according to claim 2, wherein the bandpass filter is composed of a plurality of bandpass filter groups each having an adjustable gain.