JPS59216198A - Sound/soundless discrimination system for voice - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION The field of industrial application not yet invented relates to a voiced ≦ unvoiced determination method used in the voice analysis stand 5y.

(b) Prior art At present, the Nick-Call method is the mainstream voice synthesizer.
The approximate WP1 is shown in FIG. )
is a silent sound source generation circuit that generates a noise sound source signal by simulating the non-periodic nature of turbulent vibration in the human trachea. (3) is a selection switch for selecting the periodic pulse sound source signal obtained from the voiced sound source generation circuit (1) or the noise sound source signal obtained from the unvoiced sound source generation circuit (2). (4) is a multiplier that adds the original amplitude component to the sound source signal obtained from the switch (3); (5) is a digital filter that simulates the acoustic characteristics of the human vocal tract; (4
), a voice waveform signal of voiced or unvoiced voice is output by filtering the voiced or unvoiced sound source signal obtained from the source. (6) is a D converter that converts the digital audio waveform signal obtained from the digital filter (5) into an analog value.
@ is a converter, and (7) is a speaker that produces synthesized speech based on the audio waveform signal from the D-A converter (6). (8) is a parameter memory, which is the period of the periodic pulse sound source signal in the voiced sound source generation circuit (1)? The pitch parameter P that is set and also instructs the selection of the sound source signal with the selection switch (3), the amplitude component element A of the multiplier (4), and the angle parameter A that determines the amplitude component element of the multiplier (4); ) is the Percoll coefficient 'iF(= (Kn K21 '''+ Lo) that determines the filter characteristics of
is obtained every frame period. Therefore, when synthesizing unvoiced speech, the value of the pitch parameter P is 0", and the selection switch (3) detects this value "0". , the noise source signal from the unvoiced source generating circuit (2) is introduced into the multiplier (4), and by multiplying this source signal by the value of the amplifier parameter A, an unvoiced source signal is obtained. , when synthesizing voiced speech, the value of bitch parameter P indicates the pitch period by a non-zero value, and depending on this value P, the voiced sound source generation circuit (1) determines the period of the periodic pulse sound source signal. In addition to setting the pitch parameter P, the selection switch (3)
The voiced sound source generation circuit (1
) is introduced into the multiplier (4) and this sound source signal is multiplied by the value of the amplifier parameter, thereby obtaining a voiced sound source signal.

The voiced or unvoiced sound source signal thus obtained is filtered by a digital filter (51) whose filter characteristics are controlled by a Percoll coefficient Kl+"2s...+K10, and further subjected to D-A conversion (6). A voiced or unvoiced synthesized voice is uttered by a speaker (7).

In such a speech synthesizer, parameters P and A should be stored in its parameter memory (8).

Set 1° to K1- in advance based on the waveform of the original audio.
It is necessary to analyze and extract every frame period of about 0m sea.As a conventional device for this analysis, as shown in Figure 2, the autocorrelation values from the 0th to the 140th order of the audio waveform signal S are used. V. -■140k correlator fu to calculate
ll, and each 0th to 10th order autocorrelation function value ■ obtained from the correlator (11). A Percoll coefficient extractor a21 that derives Percoll coefficients from 1st order to 10th order to ~Lr+ based on ~V, o<C, and the above-mentioned correlator 11 separately from this.
Each autocorrelation function value Vll to V140 obtained from 11 and the above Percoll coefficient extractor (1 obtained from 12+
The modified correlation function W(τ
), and a maximum value detector that calculates the maximum value pm' of the correlation function W(τ) from the wise correlator date and outputs the delay time τ at this time as the pitch parameter P. and the Percoll coefficient Kl' of all the above Percoll coefficient extractor goods = residual power Ei in one frame period unit obtained along with the extraction of KIO; the pitch parameter P obtained from the maximum value detector a phrase Using an amplifier parameter extractor that calculates the amplifier parameter A distributed for each pitch based on the parameter P, A, ~K
I was getting IG. ζ et al., the voiced/unvoiced determination based on whether or not the value of the pitch parameter P is set to 60'' is performed by the voiced/unvoiced determining section, and this determination condition is detailed in Japanese Patent Publication No. 55-34956. As stated,
The maximum value ρm of the modified correlation function W(τ) obtained from the maximum value detection path 114+ and the Percoll coefficient extractor (+
The value obtained by adding the first-order Percoll coefficient obtained from 21 by 0.5 times 0-5x+, that is, 2m+0.5x
, and compare this value with a specific threshold t. When p m + 0.5 K, ≥ t, it is voiced speech, and conversely, when p m + 0.5 K1 < t, it is unvoiced speech. However, in such conventional Percoll-based voice analysis methods, the Percoll coefficient, which is the main parameter and has the largest amount of information, is In calculating +o', the autocorrelation coefficients from the 0th order to the 10th order are obtained using a correlator 1ull. 〜■1°, whereas it is necessary to calculate the pitch parameter P, which has a relatively small amount of information, and the voiced/unvoiced judgment signal Ui, the correlator [111 roughly 1°
Autocorrelation function values from 1st to 140th order Vll to V14
It is necessary to calculate 0, which requires a huge amount of calculation, and using the kneading 1st interval values v1□ to v14fl, f? :,
The even larger number of calculations in the modified correlator t131 increases the amount of calculation required for such an analysis process. was occupied.

Therefore, it is difficult to develop a speech analysis device using such conventional methods. There are limits to how much space can be reduced and how quickly the analysis time can be shortened.

Real-time processing of speech analysis was impossible using a microcomputer-class computing system.

On the other hand, a method has been proposed for extracting the pitch parameter P and the voiced/unvoiced determination signal U from the cancer-voice waveform signal in @contact time, and if these 16 are adopted,
This eliminates the need for the modified correlator t131 in the conventional speech analysis method, and further reduces the division iF of the correlator l to 1J, which is especially difficult to use in the conventional voiced/unvoiced determination method. Because of the low accuracy, there is a large error in identifying voiced and unvoiced regions of speech, which has the disadvantage of deteriorating the quality of synthesized speech obtained by the speech synthesizer as shown in FIG.

(C) Object of the Invention The present invention has been made in view of the above-mentioned points, and provides a voiced/unvoiced voice determination method that is highly accurate and capable of performing determination processing in real time.

2) Structure of the Invention The voiced/unvoiced voice determination method of the present invention calculates the maximum value of the low-frequency component signal of the voice waveform signal as the L-1 voice wave and the 0th order autocorrelation function value (rv., 1st order) of the shadow signal. The Percoll coefficient of 1.
When different constants are a, b, c, and d, (1) L plate x(a (ll) L max > a, and <b(I
ff) L rnax) a, and b(K, < c, and ■. < d) When it satisfies any of the following expressions, it is determined to be voiceless, and otherwise it is determined to be voiced. .

(E) A speech analysis device employing the speech voiced/unvoiced determination section formula of the present invention in the embodiment W diagram? show. In figure 1jj1, (21) is P
Autocorrelation function values V from the 0th order to the 10th order of the CM audio waveform signal S. ~■,. The correlator that outputs q is the autocorrelation function value obtained from the correlator 1). This is a Percoll coefficient extraction type that derives Noku-Call coefficients from the first order to the tenth order based on ~v1°. □□□ has a rotary digital filter configuration that blocks high frequency signals of the PCM audio waveform signal S, and completely passes the low-frequency component signal L below 103 Hz, where there is a pitch period that is the fundamental frequency of audio. The influence of formants that exist above 100 Hz) is completely removed. (24+ fl pitch parameter extractor, which converts the difference value of the low-frequency component signal from the low-pass filter (23) into three values of positive, negative, and zero. The autocorrelation function value W(τ) of this trivalent signal is calculated as , and the delay time τ when this i1n becomes the maximum value closest to 1"
The pitch I (0125 output as a parameter P) is an ambuno parameter extractor, and the above Percoll coefficient is added to the 1 obtained by extracting 1 degree to ~.
The residual power E per frame period is calculated as 1 based on the pitch parameter P extracted from the pitch parameter extractor.
Calculate the amplifier parameter A distributed for each pitch. (
28) D. A voiced/unvoiced determination unit maintained in the present invention, which is the correlator c! 10 from the 0th order of the audio waveform signal S calculated in step 11
The autocorrelation function up to [V. ~■,. Among them, the power value of the audio waveform signal represented by the 0th order ■. and above)(
- 10 from the first order extracted by the call coefficient extractor (22)
The correlation coefficient of the voice waveform represented by the first-order one of 1 to KIO to the next call coefficient @1 sample delay, and the voice waveform signal S obtained from the above-mentioned low-pass filter (23). Voiced/unvoiced judgment is performed using the low M component signal @L. In other words, the above-mentioned power (■) is indispensable for Percall-based voice analysis without using any special functions or parameters. A determination process is performed using the above-mentioned low frequency component signal with a correlation degree of 1 and a pitch parameter P of 8. The voiced/unvoiced determination unit @ includes a maximum value detector tube that detects the maximum value Lmax in one frame of the low frequency component signal, and a maximum value Lmax from the maximum value detector tube.
a first comparator that compares nax and a specific value a);
A second comparator (29) that compares the correlation degree of 1 from the Percoll coefficient extractor with a specific value C, and a sixth comparator that also compares the correlation degree of 1 and a specific value C (b<c). comparator (
2)) and the above correlator (power from 21+) and % constant value d
and a fourth comparator (3I) for comparing L max≦a
When the low frequency component of the audio waveform signal S is recognized to be small by detecting < 1) i is detected and the randomness of the audio waveform signal S is recognized to be high. or when the sixth comparator (30) detects (c and the fourth comparator 11 detects V.<d, and the audio waveform signal is 8 degrees lower due to randomness) However, when the power is small, it is determined that the audio waveform signal S of this frame is due to the unvoiced five tones.Specifically, wo is within the condition range of the optimal constants determined experimentally. Yes, for example (1) Lmax≦ヲ7 (lll L mx >7t, and <0(if
f) L max >2-, and 0<K, <0.8
751 and V.

When any one of the six equations for stopping is satisfied, it is silent,
If it is voiced at other times, g is added with an accuracy close to 1oos and this judgment @No.U is output.

For reference, the specific configuration of the pitch parameter extractor g:(24+) of the above-mentioned ronnox filter is shown in Figure 4.The low-pass filter c!31 in the same figure is composed of two adders... One multiplier (41) and two delay elements (4
2 digital filter configuration? Therefore, if fc is set to about 100 songs, then the PCM
The formant component of the converted speech waveform @s[2 is considerably reduced, and the error in extracting the pitch parameter P can be reduced.

An example of pitch parameter extraction is the above low-pass filter (
c) Audio waveform signal e 1 (low frequency component signal S of zl, <Z) obtained from c) and signal S obtained by delaying this by one sample. (A comparison circuit (43I) is provided to compare 2-9 with When i!i!9 time τ is the maximum, it is output as the pitch parameter P.In other words, the amount of calculation for autocorrelation processing of the ternary signal is, for example, the audio waveform signal sl of 12 bits per sample. Although the amount of calculation is greatly reduced compared to the amount of calculation required for direct autocorrelation processing, there is almost no extraction of the pitch parameter P.

(f) Effects of the Invention As is clear from the above explanation, the voiced/unvoiced voice determination method of the present invention determines the maximum value ff L of the low frequency component signal of the voice waveform signal.
maX s The zero-order autocorrelation function value of the audio waveform signal is ■
Let the first-order Percoll coefficient be 1, and let the different constants be a,
When b, c, and d, medium Lmax≦a (11), Lmax > a s and x, <b(1i
l) L max > a and b <KI<:
IC, eyes V. When any of the expressions < d is satisfied, it is determined that it is unvoiced, and otherwise it is determined that it is voiced. Without performing two high-order autocorrelation processes,
Judgment processing can be performed in real time with a small amount of calculation.

Moreover, since 1tVok is used for the valid data L-x+ of the above-mentioned number as a test case, high-quality determination processing of h embryos is possible.

In addition, since it is possible to determine whether voice is voiced or unvoiced by reusing the data used to derive voice characteristic parameters in the voice analysis process, the amount of calculation for this determination process is extremely small.
Such voice analysis processing? It becomes possible to execute in real time on a small microcomputer-class calculation processing system.

[Brief explanation of the drawing]

Figure 1 is a block diagram showing the configuration of a general speech synthesis device, Figure 2 is a block diagram showing the configuration of a conventional speech analysis device, and Figure 6 adopts the voiced and unvoiced method of the present invention. FIG. 4 is a block diagram showing the configuration of a speech analysis device according to the present invention, and FIG.
(21) is a correlator, t is a Percoll coefficient extractor, ha-hello bass filter, 4) is a pitch parameter extractor, t
25) is an amplifier parameter extractor, (in) is a voiced/unvoiced judgment unit, (27) is a maximum value detection circuit, (to) (29) @is
u indicates the entire comparison circuit 0

Claims (1)

[Claims] (1) The maximum value Lmax of the low-frequency component signal obtained from the low-pass filter that completely blocks the high-frequency components of the audio waveform signal,
Autocorrelation function of audio waveform signal? Zero-order autocorrelation flash value V obtained from the correlator to be calculated. and the 0th to nth autocorrelation obtained by the upper phase 5lil device [...function 1 obtained from the Percoll coefficient extractor that derives 1 to Kne from the 1st to 1 for the next Percoll coefficient and constants a, b, c. The following formula expressed by d and % formula % ( (When any of the formulas in formula 06 is satisfied, the above audio waveform signal is determined to be an unvoiced audio signal, and conversely, when neither formula is satisfied. , a patented voiced/unvoiced voice determination method that determines that the audio waveform signal is a voiced audio signal.