JP3030869B2 - Sound source data generation method for speech synthesizer - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE to the present invention is the rule speech synthesis system
That relates to a speech synthesis apparatus, and more particularly to a method of generating a sound source data.

[0002]

2. Description of the Related Art A speech synthesizer using a regular speech synthesis system divides an input character string into words and phrases by syntactic analysis, and determines intonation and accent for each.
Words and phrases are decomposed into syllables and even phonemes, and sound source waves and articulation filter parameters are obtained for each syllable or phoneme, and a synthesized speech is obtained as a response output of the articulation filter to the sound source wave.

In such a speech synthesizer, there is a method using impulse and noise as sound source information, or a method using residual information. In the method of using the residual as sound source information, a speech waveform is subjected to linear prediction analysis to determine articulation parameters, a speech waveform is input to an articulation filter based on these parameters, and a residual waveform is obtained at the output thereof. The waveform is used as sound source information by sampling and encoding. In addition, the extraction of the audio waveform is performed by multiplying the original waveform by a window function (Humming window, Hanning window, etc.) so that no abrupt change occurs at both ends of the extraction section.

[0004]

As a conventional means for obtaining sound source information using a residual waveform, a reference point (pitch synchronization point) for extracting each pitch waveform is extracted from several zero-crossing point candidates on the basis of a voice waveform. ing. For example, in the case of “U” utterance by the female voice shown in FIG. 4A, the number of candidate zero-crossing points is as small as two to three, and the zero-crossing point can be accurately obtained in a waveform having a small amplitude of the pitch waveform. .

However, in the case of "U" utterance by the male voice shown in FIG. 4B, when there are many zero-crossing points and the attenuation of the waveform amplitude for each pitch is large, a zero-crossing point near the appearance of the residual peak can be obtained accurately. Instead, a point shifted in phase is often selected as a reference. If stored in a sound source file with the phase shifted, waveform distortion occurs during synthesis .
4A and 4B, a to e and a 'to c' are pitch synchronization points.

When analyzing a female voice using the above-described method, since the zero-crossing point is accurately obtained, it is possible to accurately divide the waveform into individual pitch unit waveforms by the pitch synchronization means. When analyzing male voice,
Due to the difference in pitch waveforms, the error at the zero-crossing point is large and a division reference cannot be obtained at an accurate position.

SUMMARY OF THE INVENTION The present invention has been made in view of the above circumstances, and has been made in view of the above circumstances. It is intended to provide a generation method.

[0008]

SUMMARY OF THE INVENTION In order to achieve the above object, the present invention provides original waveform data obtained by sampling audio.
And voiced sound section over the entire section where voice exists
Pitch corresponding to the repetition period of the audio waveform seen in
Calculate all the pitch sections corresponding to the
Perform waveform analysis to obtain residuals, selected for each frame
After calculating the autocorrelation of the residual, the autocorrelation is performed for each frame.
The point where the correlation value near the pitch period becomes maximum on the Seki pattern
And use this value to calculate the pitch interval at the center of each frame.
The relative peak position of the residual waveform amplitude within the
Between the frames with the highest energy
This frame is used as the reference frame when calculating the
Of the pitch waveform in the center of the frame
Calculate the peak position of the resulting residual, and then set the time axis
Find the correlation between adjacent frames in the forward direction, and
First, find the correlation between frames that are
Use the remaining peak position for the obtained voiced sound section.
The pitch period is calculated for each frame, and the total
Performs averaging of the pitch periodic sequence pattern for the
Using this averaged pitch period sequence,
After determining the difference peak position, the peak position and the actual residual
Compare to the peak position of the difference waveform and
If there is a residual waveform peak at
And reanalyze the residual peak position obtained by fine adjustment.
To the residual amplitude maximum point near the previous residual peak position.
As a residual peak, and the phase of that peak is
The residual waveform for the pitch period is
It is characterized by cutting out and generating sound source data.

[0009]

After the division of the pitch section from the original waveform is detected, the residual of all frames is calculated. After that, the residual peak position is extracted. After this extraction, the maximum residual point is extracted as the peak position of the section. On the other hand, a cross-correlation calculation or the like between a section and a section adjacent in the front-back direction of the time axis is performed. The peak positions are extracted one after another by the calculation, the pitch between adjacent pitches is set as a pitch, and a row of the pitch is averaged. Thereafter, the peak position is extracted again, the peak position is finely adjusted, and the final peak position is extracted.

[0010]

An embodiment of the present invention will be described below with reference to the drawings.

In FIG. 1, step S ₁ is a process for dividing and detecting a pitch section from the original waveform shown in FIG. 2. When detecting a reference point in this step S ₁ , once the entire frame is detected without using the waveform as a reference. Calculate the residual of. This is the process of step S _2. The process of step S ₂ performs the process of step S ₃ if finished. Step S ₃ performs a correlation calculation of the residual sequence to each other, obtained for each pitch period. The results of the correlation calculation to extract a residual peak position at Step S _4.

[0012] Turning to the process in step S ₅ If the residual peak position is extracted. Step S ₅ is a maximum energy of the frame of the vowel section (typically a strong peak resistance), it extracts the maximum point of the residual as a peak position of the section. Further, in the longitudinal direction of the time axis from the section, calculating the cross-correlation between the next section in step S _6. The result of this calculation, than the width of the maximum correlation, decided As shown in FIGS. 2 to 3 one after another peak position at step S _7.

[0013] For example, as shown in the sound source waveform diagram of FIG. 3, Peak (M) → Peak (M-1) → Peak (M-2) ...... , and Peak (M) → Peak (M + 1) → Peak
If (M + 2) ...... and the peak position has been determined, performing interval of the next to each other pitch P (1), with P (2) ... points and step S _8. The pitch column averaging process in step S ₉ at step S _8, so as to smoothly transition. Next, the averaged pitches P (1) ', P
(2) '... to determine again peaks at Step S ₁₀ the.
After determining the peak, to fine tune the peak position in the vicinity in step S _11. For example, if there is a larger peak nearby, it is shifted. Then, to determine the final peak position in step S _12. The final peak position from the step S ₁₃ based on established a phase in performing the processing of S ₁₅ each pitch interval stored in the music file. Incidentally, Step S ₁₃ is a process of cutting out on the basis of the final peak position, step S _14, turn the peak position based on
And out reanalysis a process for production of the sound source data, the step S ₁₅ is a process of storing the music file established a phase at each pitch interval.

[0014]

As described above, according to the present invention,
An accurate peak position can always be obtained even for a waveform having a large number of zero-crossings and a large waveform attenuation within a pitch.

[Brief description of the drawings]

FIG. 1 is a flowchart showing one embodiment of the present invention.

FIG. 2 is an original waveform diagram.

FIG. 3 is a waveform diagram of a sound source.

FIG. 4A is a female voice waveform diagram, and B is a male voice waveform diagram .

Continuation of the front page (58) Field surveyed (Int. Cl. ⁷ , DB name) G10L 11/00-21/06 JICST file (JOIS)

Claims (1)

(57) [Claims] 1. Extraction from voice as one analysis target
The audio waveform in the section that was
The pitch period corresponding to the repetition period of the speech waveform
The corresponding section is called the pitch section, and
One frame is defined as a section between a certain time and a certain time.
Cut out and cut out as speech waveforms for three pitch sections
Apply a Hamming window to the extracted waveform, and sound the cutout section
On the section, while shifting slightly,
Find the parameters and set the shift width
And a pair of voice parameters for each pitch section.
In the method for calculating the data, the original waveform data obtained by sampling the sound is used to calculate the sound.
Sounds that are seen in voiced sections over all sections where
Corresponds to the pitch period corresponding to the repetition period of the voice waveform
Calculate all pitch sections and analyze the waveform of each frame
To get the residuals, and the self-phase of the residuals chosen for each frame
After calculating the correlation, the autocorrelation pattern is calculated for each frame.
Finds the point where the correlation value near the pitch period is maximal.
Value, the residual wave in the central pitch section in each frame
The relative peak position of the shape amplitude is detected, and
Calculate residual peak position for frames with large energy
The frame at which the maximum energy is
Obtained by analyzing the pitch waveform at the center of the
Calculate the peak position of the residual, and then calculate the correlation between frames that are adjacent in the forward direction on the time axis.
Frame that is next to the time axis
Between the remaining voiced speech sections
The pitch period is calculated for each frame using the peak position, and the pitch period sequence pattern for all the obtained frames is obtained.
Averaging processing is performed, and this averaged pitch period sequence is used.
After calculating a new residual peak position,
And the actual peak position of the residual waveform.
If there is an actual residual waveform peak near the position,
The residual peak position obtained by shifting and
Is analyzed again, and the residual vibration near the previous residual peak position is analyzed.
The maximum width point is extracted as the final residual peak, and the peak
The pitch period of each frame
A sound source data generating method for a speech synthesizer , wherein a residual waveform of a minute is cut out to generate sound source data.