JPS6151320B2 - - Google Patents

Abstract

PURPOSE:To determine a retrieval range accurately by retrieving a pitch period where the retrieval range of the pitch period in each frame of the sound section except the first word head is limited near the pitch period detected in the past frame. CONSTITUTION:Sound/non-sound discriminator 405 discriminates whether each frame has sound or not, and continuous sound sections and continuous non-sound sections are detected by detectors 407 and 408 respectively on a basis of discrimination signals. Then, busy discriminator 410 discriminates a busy state or non-busy state on a basis of continuous sound section information and continuous non-sound section information, and a fixed time interval is given to each frame by the first word head detector 411 on a basis of sound/non-sound information and busy/non- busy information. Then, integral-multiple pitch period error correction is performed in the first work head section by 413, and the pitch period which has a formant error corrected is extracted by 414, and the pitch period where the retrieval range of the pitch period in each frame of sound sections except the first word head is limited near the pitch period detected in the past frame is retrieved by 409. As a result, the retrieval range can be determined accurately.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a pitch extraction device that performs pitch extraction based on an autocorrelation coefficient obtained by analyzing a speech waveform with a frame period approximately equal to the pitch period. This invention relates to a pitch extraction device that can significantly reduce pitch extraction errors in continuous sections.

It is known that the voiced part of a speech waveform has a periodic repeating waveform, and the change characteristic of the period (pitch period) is an important parameter in speech analysis, synthesis, recognition, etc. For example, in a speech analysis and synthesis system, the pitch extraction result extracted by the analysis section has a large effect on the quality of synthesized speech synthesized by the synthesis section.

As a method for extracting the pitch period of an audio waveform, methods using various analysis parameters are conventionally known, such as a method of calculating and extracting an autocorrelation coefficient for each frame having a time length approximately equal to the pitch period.

Pitch extraction methods based on autocorrelation coefficients are widely used because the autocorrelation coefficients can be obtained by processing in the time domain, and the influence of the phase between the analyzed waveform and the frame is relatively small. ing. However, the pitch extraction method based on the autocorrelation system uses integral multiples of the pitch period or the pitch period.
This method has the disadvantage that a period of N1/N2 times is often mistakenly detected as a pitch period. (however
(N1 and N2 are integers, and N1 < N2) The analyzed waveform in which the above defects occur can be classified into the so-called voiced sound stationary part and the so-called voiced sound transient part such as the beginning of a word. Broadly classified.

One of the reasons why the above-mentioned defects occur in voiced sound stationary parts is that the waveform to be analyzed has extremely strong stationarity.
This is because if the so-called stationary part of a voiced sound is observed over a relatively long period of time, for example, several hundred mSEC, the waveform element whose pitch period is one unit will have the following pitch period and waveform element shape: It is recognized that there is a gradual change. However, for various segments of the stationary voiced part, the time length of the waveform cut out for each frame period (for example, 30 m
When observed over a relatively short period of time (SEC), the waveform often exhibits almost perfect stationarity, that is, periodicity. For example, the autocorrelation coefficient waveform of a sine wave becomes a cosine wave with the same period as the sine wave, and as is well known, the autocorrelation coefficient waveform of a waveform with stationarity, that is, periodicity, has a period. have sex. Therefore, when a waveform cut out at a time length of, for example, about 30 mSEC for each frame period exhibits almost perfect stationarity, that is, periodicity, the autocorrelation coefficient waveform exhibits almost perfect periodicity. Therefore, for example, as shown in Figure 2, the maximum value of the autocorrelation coefficient in the pitch period, 201, is almost equal to the maximum value, 202, in the double pitch period, and due to the effects of calculation accuracy and slight disturbances, the maximum value of the autocorrelation coefficient in the pitch period, 201, This is because the maximum value 202 in the double pitch period frequently occurs larger than the maximum value 201.

Another reason why the above-mentioned defects occur in the voiced stationary part is that the utterer of the analyzed waveform has a narrow band width for one formant, and furthermore, the center frequency of the first formant is twice the pitch frequency (the reciprocal of the pitch period). For example, if the pitch frequency is an integer multiple of
The harmonics resonate with the formant, and the pitch frequency is 2
The double frequency component is extremely emphasized, and the fundamental frequency of the analyzed waveform becomes twice the pitch frequency. When the apparent period of the analyzed waveform in which the frequency component twice the pitch frequency is extremely emphasized, that is, the reciprocal of the apparent fundamental frequency becomes 1/2 of the original pitch period, the autocorrelation coefficient waveform of the analyzed waveform shows periodicity with a period that is 1/2 of the original pitch period. Therefore, for example, as shown in Figure 1, the maximum value of the autocorrelation coefficient, 101, which appears due to the resonance between the formant and the pitch, and the maximum value, 102, in the original pitch period are almost equal, which prevents false detection of the pitch period. Cause.

The reason why the above-mentioned defects occur in the voiced sound transition part is that the pitch period and the shape of the voice waveform in the voiced sound transition part change greatly and are relatively irregular. The autocorrelation coefficient waveform of the analyzed waveform, which has a large change in the pitch period and the shape of the audio waveform and is relatively irregular, has rough periodicity due to the pitch period in many cases. The maximum values of the autocorrelation coefficients at periods that are integral multiples of the pitch period are relatively uneven, and the maximum value of the autocorrelation coefficient at the pitch period is often smaller than the maximum value at periods that are an integral multiple of the pitch period, so-called integral multiple pitch periods. Many errors occur.

Note that voiced transient parts generally have large changes in the pitch frequency, and the effect on the speech waveform due to resonance between the harmonics of the pitch frequency and the formant frequency is small compared to the effect in the steady voiced part; The frequency of occurrence of so-called formant pitch errors is smaller than that in voiced stationary parts.

The impact of pitch detection errors on the quality of synthesized speech in a speech analysis and synthesis system is auditory:
The error in the constant sound part is large, and the effect of the error in the transient part of the sound sound is relatively small.

In the past, various methods have been attempted to reduce or eliminate pitch detection errors, particularly in voiced sound stationary parts, which have a large effect on the quality of synthesized speech. However, because the conventional method treats both the steady part of voiced sounds and the transient part such as the beginning of a word, if a pitch detection error happens to occur at the beginning of a word, for example, the pitch detection error will affect future pitch detection characteristics. It has the disadvantage of having negative effects.

Conventional methods, for example, take advantage of the fact that the pitch period of audio changes relatively slowly and limit the difference in pitch period between adjacent frames to a predetermined range to extract the pitch period. There is a known method for preventing pitch period detection errors. However, as shown in FIG. 3, the above-mentioned method of limiting the search range to the basic pitch period curve 301, for example, is difficult to detect due to a detection error, etc. Once the double pitch period is detected, it is difficult to detect the correct basic pitch period again. In particular, when there is a transition from a silent part such as the beginning of a word to a voiced part, or from an unvoiced part to a voiced part, there is a high risk of erroneously detecting the double pitch period.

In order to alleviate the above-mentioned drawback, when determining the pitch search range from the pitch cycles detected in the past few frames, there is a drawback that it is difficult to determine the pitch search range at the beginning of a word.

SUMMARY OF THE INVENTION An object of the present invention is to provide a pitch extraction device that performs pitch extraction based on an autocorrelation coefficient, which prevents pitch cycle detection errors and enables more reliable pitch detection. .

The present invention includes means for determining whether each frame is voiced or unvoiced, means for detecting a continuous voiced section and a continuous unvoiced section from the determined voiced/unvoiced information, and a means for detecting a continuous voiced section and a continuous unvoiced section from the continuous voiced section information and continuous unvoiced section information. means for determining busy/disconnected speech; and means for detecting a first word beginning having a predetermined time interval from the voiced/unvoiced information and the busy/discontinued information; Integer multiple pitch period errors are corrected, and pitch periods with formant errors corrected are extracted, and the search range for pitch periods in each frame of the voiced section excluding the first word beginning is set in the vicinity of pitch periods detected in past frames. A pitch extraction device is obtained which is characterized in that it searches for a pitch period with a limit.

Next, embodiments of the present invention will be described with reference to the drawings.

FIG. 4 is a block diagram for explaining an embodiment of the present invention, and the portion surrounded by a dashed line 401 in FIG. 4 shows the configuration of the present invention. Waveform input terminal 402
The audio waveform input signal from which high frequencies have been removed is input to the A/D converter 403 via the A/D converter 403 . The A/D converter 403 samples the audio waveform input signal at a frequency that is twice or more the highest frequency component included in the audio waveform input signal, further quantizes the audio waveform input signal, generates a sampled audio waveform signal, and converts the audio waveform input signal into a sampled audio waveform signal. The audio waveform signal is output to temporary storage 404.

A temporary storage device 404 temporarily stores the sampled audio waveform signal, and outputs the sampled audio waveform signal to the presence/absence determination device 405 and the autocorrelation coefficient measuring device 406 for each frame period. The presence/absence determiner 405 classifies the sampled audio waveform signal cut out for each frame into voiced sounds and unvoiced sounds using various generally well-known voiced/unvoiced sound discrimination means. In addition,
The unvoiced sound includes silence.

Further, the presence/absence determining unit 405 outputs the classification result as a presence/absence signal to a continuous voicing detector 407, a continuous unvoiced detector 408, a pitch search range measuring device 409, a first word beginning detector 411, and a second word beginning detector 412. . Continuous voicing detector 407 detects continuous voicing when the busy determination signal supplied from busy determiner 410 (to be described later) indicates a disconnection and the presence/absence signal supplied from presence/absence determiner 405 is continuously voiced (for example, 100m SEC continuous voiced). ), a continuous voiced signal is supplied to a busy determiner 410, which will be described later.

Further, the function of the continuous voicing detector 407 is suppressed when a busy determination signal supplied from a busy determining device 410 (to be described later) indicates that the telephone is busy. Continuous silence detector 4
08 is a presence/absence determination device when a busy determination signal supplied from a busy determination device 410 (to be described later) indicates a busy state;
When the presence/absence signal supplied from 05 is continuously silent (for example, when 300 m SEC is continuously silent), the continuous silent signal is supplied to the busy determiner 410, which will be described later. Further, the function of the continuous silence detector 408 is suppressed when a busy determination signal supplied from a busy determining device 410 (described later) indicates a disconnection.

The busy determiner 410 is a bistable multivibrator, and is set to be busy by the continuous voiced signal and to be disconnected by the continuous unvoiced signal. The busy determiner 410 converts the busy determination signal into a continuous voicing detector 407.
In addition to the continuous silence detector 408, the signal is also supplied to a first word-beginning detector 411 and a second word-beginning detector 412.

The first word beginning detector 411 is an AND gate,
When the busy determination signal supplied from the busy determiner 410 indicates a disconnection and the presence/absence signal supplied from the presence/absence determiner 405 indicates voiced, a first word beginning signal is generated and the pitch search range measuring device 409 The first word beginning signal is supplied to an integer pitch period error corrector 413 and a formant pitch error correction indicator 414. Therefore, the maximum time width of the first word-initial signal is the time set by the continuous voicing detector 407, for example, 100 mSEC.
matches.

The second word beginning detector 412 is an AND gate with a timer, and when the busy determination signal supplied from the busy determiner 410 indicates a speech interruption and the presence/absence signal supplied from the presence/absence determiner 405 indicates voiced. A second word beginning signal is generated and at the same time a timer starts operating. The timer is set to about 30 mSEC, for example, and advances when the busy determination signal indicates a disconnection and the presence/absence signal indicates voice presence, and returns to the initial state when at least one of the two conditions is not met. Return to Further, when the timer reaches a preset time, for example, 30 mSEC, the second word beginning signal is suppressed. The second word beginning detector 412
Formant pitch error correction indicator 41
Output to 4. Formant error correction indicator 414
is supplied with the first word-initial signal, and is supplied with the second word-initial signal.
A correction instruction signal is output to the formant error corrector 415 when the word beginning signal is not supplied.

The pitch search range measuring device 409 is the presence/absence determination device 40
When the presence/absence signal supplied from 5 indicates silence, it instructs the pitch search unit 416 to a preset pitch search range or to prohibit pitch search. Further, the pitch search range measuring device 409 instructs the pitch search device 416 to a preset pitch search range when the first word beginning detector 411 is outputting the first word beginning signal. Further, the pitch search range measuring device 409 receives a signal from the formant error corrector 415 when the presence/absence signal supplied from the presence/absence determiner 405 indicates voicing and the first word-initial detector 411 is not outputting the first word-initial signal. The pitch search range is determined from past pitch output data, and the pitch search range is instructed to the pitch search unit 416. Note that the pitch search range is determined as, for example, L _MAX =C ₁ L _AV L _MIN =C ₂ L _AV . Also, L _AV is replaced by, for example, aL _AV +bPITCHP for each frame. However, L _MAX is the maximum value of the pitch cycle search range, L _MIN is the minimum value of the pitch cycle search range, and L _AV is the past average pitch cycle PITCH _P
is the pitch period in the immediately previous frame, a is a constant (0<a<1), and b is a constant (0<b<1).
Also, generally a>b. Further, C ₁ and C ₂ are constants, and C ₁ >1>C ₂ >0. Note that it is clear that various methods other than the above method can be used to determine the pitch search range domain.

The autocorrelation coefficient measuring device 406 is a temporary memory device 404
A sequence of autocorrelation coefficients is measured from a sampled audio waveform signal supplied every frame period from , and the sequence of autocorrelation coefficients is output to a pitch searcher 416 . The pitch searcher 416 searches for the maximum value and local maximum value of the autocorrelation coefficient sequence in the pitch search range specified by the pitch search range measuring device 409, and further searches for each of the autocorrelation coefficients corresponding to the maximum value and local maximum value. Find the delay time.

Further, the pitch searcher 416 outputs each delay time information and each autocorrelation value corresponding to each delay time to the integer pitch period error corrector 413. The integer multiple pitch period error corrector 413 performs integer multiple pitch period error correction only when the first word beginning signal is supplied from the first word beginning detector 411. Integer multiple pitch period error correction is performed using, for example, a delay time that is an integer fraction of the delay time corresponding to the maximum value (for example, 1/2,
1/3, 1/4, ......), the delay time corresponding to the maximum value is determined to be the pitch period, and the maximum value is determined to exist. If not, the delay time corresponding to the maximum value is determined as the pitch period.

The integer multiple pitch period error corrector 413 further outputs the pitch period, each delay time, and the autocorrelation coefficient corresponding to each delay time to the formant error corrector 415. Formant error corrector 41
5 performs formant error correction only when a correction instruction signal is supplied from the formant error correction indicator 414. Formant error correction is performed when there is no pitch period supplied from the integer multiple pitch period error corrector 413 in the vicinity of the predicted value based on the pitch period in the past several frames (hereinafter referred to as predicted pitch period), and a local maximum value corresponding to the pitch period is further corrected. This is carried out by correcting and determining the delay time in the vicinity as the pitch period when there is a delay time of a local maximum value other than the predicted pitch period.

Note that if all the delay times of all maximum values are not in the vicinity of the predicted pitch period, the pitch period supplied from the integer multiple pitch period error corrector 413 is directly determined as the pitch period. The formant error corrector 415 supplies the corrected pitch period data to the pitch search range measuring device 409, and
The pitch period data is outputted via the output terminal 417.

As explained above, the feature of the present invention is to detect continuous voiced sections and continuous unvoiced sections from voiced/unvoiced information,
Furthermore, by determining whether the person is talking or not speaking from the two sections, the first word beginning that lasts for a relatively long time (for example, about 100 mSEC) and the second word beginning that lasts for a relatively short time (for example, about 30 mSEC) can be determined. and sets the initial value of the pitch search range using the pitch period data detected at the beginning of the first word, and detects the search range of the pitch period in each frame of the voiced sound section excluding the first word beginning in past frames. The pitch period is detected by restricting it to the vicinity of the detected pitch period, and furthermore, the pitch period is corrected by an integer multiple at the beginning of the first word, and the pitch period due to the influence of the formant is detected at the beginning of the first word except for the second beginning. The purpose is to correct period errors by referring to predicted values from pitch period data in each frame in the relatively near past.

These means can solve the aforementioned drawback that it is difficult to determine the pitch search range at the beginning of a word. In addition, for voiced sections excluding so-called word beginnings, the pitch period search range is determined by the pitch period detected in past frames, and the delay time (approximately pitch) at the maximum value of the autocorrelation coefficient due to the influence of double pitch period and formant. periodic
The delay time is N1/N2. However, N1 and N2 are integers, and N1<N2. ), it is possible to eliminate most of the causes of pitch cycle errors.

Furthermore, by providing a first word beginning and a second word beginning, it is possible to more accurately correct an integer multiple pitch period error and a pitch period error due to the influence of formants. That is, for integer multiple pitch cycle errors that tend to occur particularly frequently in voiced parts near the boundary between unvoiced parts and voiced parts, correction is performed for the entire word-initial section, that is, the first word-initial section. In addition, for pitch errors due to the influence of formants, where there is a high possibility that the correct pitch period will be incorrectly corrected to, for example, a double pitch period, if not corrected carefully, correction should be performed while referring to the predicted values of the pitch period of several past frames. is desirable.

In the present invention, by providing a second word beginning, a predicted value of the pitch period is created from the pitch period at the second word beginning, and pitch errors due to the influence of the formant at the first word beginning, which is not included in the second word beginning, can be appropriately corrected. can be corrected. By using the pitch period data at the first word beginning accurately determined as described above, the search range for the pitch period in each frame of the voiced sound section excluding the first word beginning section can be determined accurately. It's obvious what you get.

As described above, according to the present invention, it is possible to appropriately and easily determine the pitch period search range, and it is also possible to extract accurate pitch periods even for so-called word beginnings, which provide data for determining the pitch period search range. effective.

[Brief explanation of the drawing]

Figure 1 is an autocorrelation coefficient waveform diagram to explain the influence of formant, Figure 2 is an autocorrelation coefficient waveform diagram to explain the influence of double pitch period, and Figure 3 is pitch detection that occurs in the conventional method. FIG. 4, which is a diagram for explaining an example of an error, is a block diagram for explaining an embodiment of the present invention. 402...Waveform input terminal, 403...A/D converter, 404...Temporary memory, 405...Presence/absence determination device, 406...Autocorrelation coefficient measuring device, 407...
...Continuous voiced detector, 408...Continuous unvoiced detector,
409... Pitch search range measuring device, 410... Busy determiner, 411... First word beginning detector, 412...
...Second word beginning detector, 413...Integer multiple pitch cycle error corrector, 414...Formant pitch error correction indicator, 415...Formant error corrector, 4
16... Pitch search device, 417... Pitch cycle data output terminal.

Claims (1)

[Claims] 1. In a pitch extraction device that performs pitch extraction based on an autocorrelation coefficient obtained by analyzing a speech waveform at a frame period approximately equal to the pitch period, there is provided a means for determining whether each frame is voiced or unvoiced, and means for determining whether each frame is voiced or unvoiced; means for detecting a continuous voiced section from unvoiced information; means for detecting a continuous unvoiced section from the determined voiced/unvoiced information; and determining busy/interrupted speech from the continuous voiced section information and the continuous unvoiced section information. and means for detecting a first word beginning having a certain time interval from the voiced/unvoiced information and the busy/disconnected information, and detecting an integer multiple pitch cycle error in the first word beginning section. At the same time, the pitch period whose formant error has been corrected is extracted, and the search range for the pitch period in each frame of the voiced section excluding the first word initial is limited to the vicinity of the pitch period detected in the past frame. A pitch extraction device characterized by searching for a period.