JPS5912185B2 - Voiced/unvoiced determination device - Google Patents

Abstract

PURPOSE:To discriminate the voice or voiceless condition, by measuring the periodic nature or timing similarity of the series of correlativity in which the part around delayed zoro having greater effect of correlativity by the noise surrounded is removed. CONSTITUTION:The audio waveform signal including surrounding noise is inputted to the A/D converter 402 via 401. The 402 samples the signal and quantizes it, and outputs it to the temporal memory 403. 403 temporarily memorizes the sampled waveform signal continuously inputted and outputs it to the window processor 404. 404 executes the window operation to the sampled waveform signal, cuts out the sampled waveform signal, and outputs the signal cut out to the correlativity unit 405. 405 measures the correlativity from the delay time having almost no effect of surrounding noise to the integer times time of the maximum value of the pitch period and outputs it to the degree of voice measurement unit 406. 406 measures the degree of voice from the correlativity, discriminates the voice or voiceless condition, and outputs the result of discrimination via the output terminal 407 as the voice and voiceless discrimination signal.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a voiced/unvoiced determining device for determining voiced/unvoiced speech, and more particularly to a voiced/unvoiced determining device for performing a good voiced/unvoiced determination in a high ambient noise environment.

Speech analysis and synthesis of voiced/unvoiced characteristics in speech waveforms,
It is known to be an important parameter in recognition, etc.

For example, in a speech analysis and synthesis system, the voiced/unvoiced determination result determined by the analysis section has a large effect on the quality of synthesized speech synthesized by the synthesis section. Conventionally, various methods have been used to determine voicedness or unvoicedness of speech waveforms, such as energy of speech waveforms, zero crossing rate, autocorrelation coefficient at unit delay, first-order linear prediction coefficient, predicted residual power in linear prediction analysis, etc. Combinations of the various methods described above are known. However, in conventional methods, such as methods that use the energy of voice waveforms, in a high ambient noise environment, unvoiced sounds with low energy increase in energy due to the influence of noise.
Unvoiced sounds are often mistaken for voiced sounds. Similarly, the method using the zero crossing rate often mistakes voiced sounds with relatively low energy, such as so-called crossing parts between voiced sounds or the vicinity of the end of a voiced sound, as unvoiced sounds. In addition, the method using the autocorrelation coefficient due to unit delay, the method using the first-order linear prediction constant, and the method using the prediction residual power in linear prediction analysis are generally different from each other when the ambient noise is expressed as a power spectrum. Since it is a loud colored noise, it has a great influence. As described above, the conventional voiced/unvoiced determination method has the disadvantage that it is easy to make a determination error in a high ambient noise environment. SUMMARY OF THE INVENTION An object of the present invention is to provide a voiced/unvoiced determination device that reduces the influence of ambient noise and enables more reliable voiced/unvoiced determination.

The present invention is based on a speech waveform (hereinafter referred to as residual waveform n) that is generated directly from a speech waveform or passed through an inverse filter created using parameters expressing the speech spectrum (for example,
Fumitada Itakura, "Speech feature extraction using statistical methods," 8th Symposium sponsored by the Institute of Electrical Communication, Tohoku University.

Tsutomu Kobayashi, Kei Yamamoto, “Determination of voiced/voicedness using polar correlation method”, Proceedings of the Acoustical Society of Japan Research Conference,
Autocorrelation, cepstrum, polar correlation coefficient, Difference coefficient (for example, Hitoshi Fukuma, Kenichi Nakamura "A method of extracting voice pitch using 30 differences"
, 1971 Information Division National Conference of the Institute of Electronics and Communication Engineers), etc., and a means for measuring the periodicity or temporal similarity of the correlation coefficient.
The first feature of the present invention is that the correlation coefficient of a voiced sound has an approximate periodicity of 35, and the approximate period closely matches the pitch period.
, and the second property that the correlation coefficients of unvoiced sounds and ambient noise in many cases converge in the direction of time delay: the correlation coefficient of ambient noise is The purpose of this method is to measure the periodicity or temporal similarity of a correlation coefficient sequence from which the portion near zero delay, which is largely affected by the noise, is removed to determine whether or not there is a voice.

Therefore, it is possible to accurately extract the feature of voiced sound that the correlation coefficient has approximately periodicity in a high ambient noise environment. Therefore, there is an effect that more accurate voiced/unvoiced determination can be performed under the above environment. In general, the voiced part of a speech waveform is approximately periodic, so the correlation waveform of the voiced part directly determined from the speech waveform is approximately periodic, such as the autocorrelation waveform shown in FIG. 1, for example. On the other hand, ambient noise is usually
Correlation waveforms that have weak periodicity and can be directly determined from ambient noise waveforms, such as the autocorrelation waveform shown in Figure 2, show large changes near zero delay, and converge to a constant value as the delay increases. I'll go. The convergence value is, for example, zero for an autocorrelation coefficient, cepstrum, polar correlation coefficient, etc., and is a positive value for a difference coefficient, etc. Furthermore, the cross-correlation between ambient noise and voiced sound is generally almost zero, and if we ignore the dispersion of cross-correlation due to the influence of the analysis window, the correlation waveform of voiced sound mixed with ambient noise will be less than the autocorrelation waveform. For example,
For example, as shown in FIG. 3, it is approximately expressed by linear addition of the autocorrelation waveform of the voiced sound shown in FIG. 1 and the autocorrelation waveform of the ambient noise shown in FIG. 2. Regarding other correlation coefficients, such as the kebstrum, the autocorrelation coefficient is the inverse Fourier transform of the Pacou spectrum of the waveform, whereas the kebstrum is the inverse Fourier transform of the logarithm of the absolute value of the Pacou spectrum of the waveform. As is clear, the cepstrum of a voiced sound mixed with ambient noise is approximately expressed by linear addition of the cepstrum of the voiced sound and the cepstrum of the ambient noise. The same applies to polarity correlation coefficients, difference coefficients, etc.
Therefore, the correlation waveform or the change in the correlation waveform at a delay time of about the pitch period or more than the pitch period almost represents the correlation waveform of the voiced sound. Furthermore, the correlation waveform of unvoiced sounds generally does not have periodicity. Therefore,
By measuring the similarity in the pitch period and the delay time of an integral multiple of the pitch period of the correlation waveform of the direct voice mixed with ambient noise, or by measuring the periodicity of the correlation waveform, it is possible to determine whether the voice is voiced or unvoiced. Judgments can be made reliably. Furthermore, even when using the correlation coefficient obtained from the residual waveform, the periodicity of voiced parts is preserved (for example, as described above in "Determination of voiced/unvoiced using polar correlation method"), and It can be handled in the same way as the correlation coefficient obtained from speech. Next, the present invention will be explained in detail with reference to the drawings.

FIG. 4 is a block diagram showing a first embodiment of the present invention. First, an audio waveform signal containing ambient noise is input to the A/D converter 402 via the waveform input terminal 401.

The A/D converter 402 samples the waveform signal, further quantizes it, and outputs it to the temporary memory 403. temporary memory 4
03 temporarily stores the continuously input sampled waveform signal and outputs it to the window processor 404. The window processor 404 performs a window function operation such as rectangular window, . . . ming window, etc. on the sampled waveform signal,
The sampled waveform signal is cut out, and the cut out waveform signal is output to the correlator 405. The correlator 405 uses a delay time that is almost unaffected by ambient noise (for example, 2 mSEC) to an integer multiple of the maximum value of the pitch period (for example, 45 times three times 15 mSEC).
The voicing degree measuring device 406 measures the correlation coefficient up to
Output to.

The voicing degree measuring device 406 measures the voicing degree from the correlation coefficient, determines whether the signal is voiced or unvoiced, and outputs the determination result as a voiced/unvoiced determination signal via a voiced/unvoiced determination signal output terminal 407 .
Next, a first configuration example of the voicing degree measuring device 406 will be described in detail with reference to the drawings.

FIG. 5 is a block diagram showing a first configuration example of the voicing level measuring device 406. This configuration example is effective for correlation coefficients such as autocorrelation coefficients, cepstral correlation coefficients, and polar correlation coefficients. Correlator 40
The correlation coefficient measured in step 5 is input to the similarity measuring device 502 via the correlation coefficient input terminal 501.

The similarity measuring device 502 measures the similarity, R(K), for example.

However, ρP is the correlation coefficient at delay P, and K is the temporal distance between the delay times, which is the minimum value of the pitch search range (for example, 2.5 m).
SEC) to the maximum value of the pitch search range (for example, 15 mS
EC). N is a delay time that is hardly affected by ambient noise, and M is an integral multiple of the maximum value of the pitch period (
For example, the delay time is shorter than 45 mSEC, which is three times 15 mSEC. The similarity measurer 502 further outputs the measured R(K) to the maximum value searcher 503. Maximum value searcher 503 searches for the maximum value RMAX of R(K) and outputs it to presence/absence determination unit 504 . The presence/absence determination unit 504 compares RMAX with a preset determination reference value, and determines whether RMAX
If it is larger than the determination reference value, it is determined to be voiced, and if it is smaller, it is determined to be unvoiced, and the determination result is outputted as a voiced/unvoiced determination signal via the voiced/unvoiced determination signal output terminal 505. Next, a second configuration example of the voicing degree measuring device 406 will be described in detail with reference to the drawings.

FIG. 6 is a block diagram showing a second configuration example of the voicing degree measuring device 406. This configuration example is effective for correlation coefficients such as autocorrelation coefficients, cepstral correlation coefficients, and polar correlation coefficients. Correlator 40
The correlation coefficient measured in step 5 is input to a similarity measuring device 602 and a maximum value searching device 605 via a correlation coefficient input terminal 601.

The similarity measuring device 602 measures the similarity S(K), for example.

However, ρ2 is the correlation coefficient at delay P, and K is the temporal distance of the delay time, which is the minimum value of the pitch search range (for example, 2.5 mS
EC) to the maximum value of the pitch search range (e.g. 15mSE
C). N is a delay time that is hardly affected by ambient noise, and M is a delay time that is shorter than an integral multiple (for example, 45 mSEC) of the maximum value of the pitch period. The similarity measurer 602 further outputs the measured S(K) to the minimum value searcher 603. The minimum value searcher 603 searches for the minimum value SMIN of S(K) and outputs it to the presence/absence determination unit 604.
The maximum value searcher 605 searches for the maximum value ρMAX of the correlation coefficients supplied via the correlation coefficient input terminal 601 and outputs it to the determination signal generator 606. The determination signal generator 606 generates a determination reference value signal that monotonically increases as silk AX increases, and outputs it to the presence/absence determiner 604. Presence/absence determiner 60
4 compares SMIN supplied from the minimum value searcher 603 with the determination reference value signal supplied from the determination signal generator 606, and if SMIN is smaller than the determination reference value, it is voiced, and if it is larger, it is voiced. The determination result is output as a voiced/unvoiced determination signal via the voiced/unvoiced determination signal output terminal 607. Next, a third configuration example of the voicing degree measuring device 406 will be described in detail with reference to the drawings. Figure 7 shows the voicing level measuring device 40.
FIG. 6 is a block diagram showing a third configuration example of No. 6; This configuration example is effective for correlation coefficients such as autocorrelation coefficients, cepstral correlation coefficients, and polar correlation coefficients. The correlation coefficient measured by the correlator 405 is inputted to the pitch extractor 702 and the integer multiple pitch searcher 703 via the correlation coefficient input terminal 701.

The pitch extractor 702 uses a well-known pitch extraction method using a correlation coefficient to obtain the pitch period and the correlation coefficient ρMAX in the pitch period. Furthermore, the pitch extractor 702 uses the pitch searcher 7 to multiply the pitch period by an integer.
03, ρMAX is output to the presence/absence determination unit 704, respectively. The integer multiple pitch searcher 703 uses the pitch period information supplied from the pitch extractor 702 to calculate a delay time, for example, twice the pitch period, of the correlation coefficient supplied via the correlation coefficient input terminal 701. Search for the maximum correlation value, ρ2MAX, etc. near an integral multiple of the delay time. Furthermore, the integer multiple pitch searcher 703 determines whether ρ2MAX etc.
Output to 4. The presence/absence determiner 704 measures the ratio of the ρ2MAX to the ρMAX, and if the ratio is larger than a preset determination reference value, it is determined to be voiced, and if it is smaller, it is determined to be unvoiced, and the determination result is determined as voiced/unvoiced. It is output as a signal via the voiced/unvoiced determination signal output terminal 705. Next, a fourth configuration example of the voicing degree measuring device 406 will be described in detail with reference to the drawings. FIG. 8 is a block diagram showing a fourth example of the configuration of the voicing level measuring device 406. This embodiment is effective for correlation coefficients such as difference coefficients, and correlation coefficients such as autocorrelation coefficients, cepstrum, and polar correlation coefficients. The correlation coefficient measured by the correlator 405 is transmitted to the correlation coefficient manual terminal 80.
1 to a minimum value extractor 802, a pitch-delayed minimum value extractor 803, a maximum value extractor 804, a pitch-delayed maximum value extractor 805, and a pitch extractor 806.

The minimum value extractor 802 searches for the minimum value, ρMIN, and the delay time at ρMIN, TMIN, from the delay time correlation coefficient sequence that is hardly affected by ambient noise, sends ρMIN to the presence/absence determiner 807, and sends TMIN to the delay time by a pitch. Minimum value extractor 80
Output each to 3. The pitch extractor 806 obtains the pitch period Tp using a well-known pitch extraction method using a correlation coefficient, and outputs it to the pitch delay minimum value extractor 803 and the pitch delay maximum value extractor 805. The pitch delay minimum value extractor 803 searches for the minimum value ρMIND from the correlation coefficient sequence near TMlN+TP, and the presence/absence determination unit 807 searches for the minimum value ρMIND.
Output to. The maximum value extractor 804 searches the maximum value, ρMAX, and the delay time at ρMAX, TMAX, from the delay time correlation coefficient sequence that is almost unaffected by ambient noise, and calculates ρMAX.
X is output to the presence/absence determiner 807, and TMAX is output to the pitch delay local maximum value extractor 805, respectively. The pitch delay local maximum value extractor 805 calculates TMAX+TP using the TMAX and Tp supplied from the pitch extractor 806, searches for the local maximum value ρMAXD from the correlation coefficients around the calculated value, and determines whether the value is present or not. Output to. The presence/absence determination unit 807 measures the presence/absence degree V, for example.

Further, the presence/absence determiner 807 compares v with a determination reference value set in advance, for example, and if v is larger than the determination reference value, it is determined to be voiced, and if it is smaller, it is determined to be voiceless, and the determination result is determined as voiced/unvoiced. It is output as a signal via the voiced/unvoiced determination signal output terminal 808. Next, a second embodiment of the present invention will be described in detail with reference to the drawings.

FIG. 9 is a block diagram showing a second embodiment of the present invention. First, an audio waveform signal containing ambient noise is input to the A/D converter 902 via the waveform input terminal 901.

A/D converter 902 samples the waveform signal, further quantizes it, and outputs it to temporary memory 903. Temporary memory 9
03 temporarily stores the continuously input sampled waveform signal and outputs it to the window processor 904. The window processor 904 performs a window function operation such as rectangular window, . . . ming window, etc. on the sampled waveform signal,
A sampled waveform signal is cut out, and the cut out waveform signal is output to a spectrum parameter extractor 905 and an inverse filter 906. A spectrum parameter extractor 905 extracts spectrum parameters using a linear prediction method or the like and outputs them to an inverse filter 906. The inverse filter 906 constitutes an inverse filter expressed by the spectrum parameters, filters the waveform supplied from the window processor 904, and outputs it to the correlator 907. The correlator 907 measures the correlation coefficient from a delay time that is hardly affected by ambient noise (for example, 2 mSEC) to an integral multiple of the maximum value of the pitch period (for example, 45 mSEC, which is three times 15 mSEC).
Output to 08. The voicing degree measuring device 908 measures the voicing degree from the correlation coefficient, determines whether it is voiced or unvoiced, and outputs the determination result as a voiced/unvoiced determination signal via a voiced/unvoiced determination signal output terminal 909 . Note that the voicing degree measuring device 908 is equivalent to the voicing degree measuring device 406 shown in FIG. 4 in the first embodiment.

The spectrum parameter extractor 905 and the inverse filter 906 are arranged between the A/D converter 902 and the temporary memory 903, or between the temporary memory 903 and the window processor 90.
It is clear that the present invention can also be practiced by moving between 4 and 4.

[Brief explanation of the drawing]

Figure 1 is a waveform diagram for explaining the characteristics of the autocorrelation waveform of a voiced part, Figure 2 is a waveform diagram for explaining the characteristics of the autocorrelation waveform of ambient noise, and Figure 3 is a waveform diagram for explaining the characteristics of the autocorrelation waveform of a voiced part. FIG. 4 is a waveform diagram for explaining the characteristics of the autocorrelation waveform of a voiced part. FIG. 4 is a block diagram for explaining the first embodiment of the present invention. In Fig. 4, 401 is a waveform input terminal, 402 is an A/
D converter, 403 temporary memory, 404 window processor, 405 correlator, 406 voicing degree measuring device, 407
is the voiced/unvoiced judgment signal output terminal. FIG. 5 is a block diagram for explaining a first configuration example of the voicing degree measuring device 406. In FIG. 5, 501 is a correlation coefficient input terminal, 502 is a similarity measuring device, 503 is a maximum value search device, 504 is a presence/absence determination device, and 505 is a voiced/unvoiced determination signal output terminal. FIG. 6 is a block diagram for explaining a second configuration example of the voicing degree measuring device 406. In FIG. 6, 601 is a correlation coefficient input terminal, 602 is a similarity measurer, 603 is a minimum value searcher, 604 is an existence/absence judger, 605 is a maximum value searcher, 606 is a judgment signal generator, and 607 is a voiced Silence judgment signal output terminal. FIG. 7 is a block diagram for explaining a third configuration example of the voicing degree measuring device 406. In FIG. 7, 701 is a correlation coefficient input terminal, 702 is a pitch extractor, 703 is an integer multiple pitch searcher, 704 is a presence/absence determination device, and 705 is a voiced/unvoiced determination signal output terminal. FIG. 8 is a block diagram for explaining a fourth configuration example of the voicing degree measuring device 406. In FIG. 8, 801 is a correlation coefficient input terminal, 802 is a minimum value extractor, 803 is a pitch-delayed minimum value extractor, and 80
4 is a local maximum value extractor, 805 is a pitch-lag local maximum value extractor,
806 is a pitch extractor, 807 is a presence/absence determination device, and 808 is a voiced/unvoiced determination signal output terminal. FIG. 9 is a block diagram for explaining a second embodiment of the present invention. In FIG. 9, 901 is a waveform input terminal, 902 is an A/
D converter, 903 temporary memory, 904 window processor, 905 spectrum parameter extractor, 906
907 is an inverse filter, 907 is a correlator, 908 is a voicing degree measuring device, and 909 is a voiced/unvoiced determination signal output terminal.

Claims (1)

[Claims] 1. A voiced/unvoiced determination device that determines voiced/unvoiced speech, comprising means for measuring a correlation coefficient from a speech waveform, and means for measuring periodicity or temporal similarity of the correlation coefficient after a certain delay time. A voiced/unvoiced determination device.