JPS58220195A - Pitch extraction system - 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 (Technical Field of the Invention) The present invention provides a pitch extraction method that can reduce the number of calculations when extracting the pitch, which is the fundamental period of speech, using an autocorrelation function of a speech waveform. It is related to.

(Prior Art and Problems) The pitch of speech (voiced sound) can be extracted using the autocorrelation function of the Otohara speech waveform or the residual waveform after analyzing the speech waveform. In this case, the autocorrelation amount is determined by integrating the product of the waveform x (t) and the waveform x (t+τ) that is shifted by a certain time τ from the waveform x (t).

In other words, the autocorrelation function R(τ) is It is expressed as

In addition, in the case of digital processing, the autocorrelation function R (kT) is It is expressed as

Here, T is the sampling period.

Conventionally, when trying to find the pitch of a voice with the accuracy of the sampling period using equation (2), a correlation calculation circuit was prepared for the time shift kT that corresponds to the entire statistical distribution range of the pitch. I was doing calculations. For example, assuming that the pitch distribution range is 2 to 16 mB, the sampling period T =
When 125 μs is set, the range of time lag 4 in which the autocorrelation function should be calculated is as follows.

16T≦kT≦128T (3) Therefore, in order to perform the calculation, a multiplier and an integrator are required.

113 dividers are required for each, and the same number of taps in the delay memory for setting the time lag of the signal are also required, resulting in a significantly larger hardware scale.

(Object of the Invention) The present invention aims to solve the problems of the prior art, and its purpose is to An object of the present invention is to provide a pitch extraction method that can reduce the number of autocorrelation functions required for extraction without reducing accuracy.

(Structure of the Invention) The pitch of speech is considered to be the period of an impulse train output from the vocal cords, and the application of this impulse and
Looking at the relationship with the voice waveform generated at that time, it is observed that the voice waveform shows a sudden rise due to the application of the impulse, and that the polarity of the voice waveform changes from negative to positive at the time of the rise. . In other words, it can be seen that the time point at which the impulse is estimated to have been applied is at the point where the voice waveform changes from negative to positive. In this way, when pitch is considered as the period of an impulse train, calculation of the autocorrelation function for pitch extraction only needs to be performed at the point where the polarity of the audio waveform changes from negative to positive.

However, even in this case, there are still many points where the polarity changes. That is, the polarity change point is the zero crossing point of the waveform, and this is because the influence of the high frequency component superimposed on the audio waveform is large at this portion.

FIG. 1 shows an example of the vicinity of a polarity change point in an audio waveform. In the figure, three zero crossing points x, y, and z are recognized, but two of them (X and 2) are the points of change from negative to positive. Therefore, in the case of the waveform shown in FIG. 1, the number of changing points at which autocorrelation should be calculated is doubled due to the influence of high frequency components. Therefore, in the present invention, in order to remove the influence of high frequency components near the zero crossing point, a threshold value is set at ±a with zero as the center, as shown by the broken line in FIG. By not determining that there has been a polarity change even if there is a zero crossing until the range is exceeded, the number of polarity change points that require autocorrelation calculation is reduced. It goes without saying that the threshold values in this case do not necessarily have to have equal positive and negative values.

(Embodiment of the Invention) FIG. 2 is a block diagram showing the configuration of an embodiment of the pitch extraction method of the present invention. In the figure, 10 is an input terminal, 20 is a change point detection circuit, □□□ is a threshold generation circuit,
40 is a counter for counting the change point interval, (capital) is a correlator, (
b) is a tap position setting circuit for the delay memory in the correlator blade i; 70 is a tap position control circuit that regulates the tap position set by the tap position setting circuit ω to fall within a predetermined range; A pitch determination circuit for determining pitch, 90 is an output terminal.

In FIG. 2, a waveform input from terminal 10 is applied to a change point detection circuit 20. In FIG. The change point detection circuit 20 is given a threshold value of, for example, ±a from the threshold generation circuit, and detects a change point from negative to positive in the input waveform that exceeds the range of this threshold value.

The point of change may be any one of the points at which the threshold value ±a is crossed for the first time, the points at which it is crossed for the second and subsequent times, or similar points at zero crossing points. Further, the threshold value generated by the threshold generation circuit (9) is selected from an appropriate value based on the average power, amplitude, added noise power, amplitude, and other factors of the input waveform, and is adaptive. Good too,
Alternatively, it may be fixed. The counter 40 counts the interval between change points detected by the change point detection circuit.

The tap position setting circuit ω sets the tap position of the delay memory between the correlators based on the count result of the counter 40. The delay memory provides a time lag for calculating the autocorrelation function in the correlator power, and is used by the counter 40.
Based on the counting result of p, taps corresponding to specific time differences are sequentially set from a certain reference point according to the detected change point. This is because a plurality of change points may occur within the l pitch range. The tap position control circuit 70 is configured to prevent the tap position set by the tag position setting device (a) when the tap position is not within a predetermined range determined statistically. Performs controlling action.

The correlator (7) calculates an autocorrelation function of the terminal 100 input signal using a time shift corresponding to the tap position set in the delay memory. The pitch determination circuit 80 determines, for example, the maximum correlation function value among the correlation function values determined by the correlator power, and outputs a time value according to the corresponding tap position of the delay memory to a terminal (A) as a pitch. do.

In the above example, the applied impulse was assumed to be positive, and the changing point at which the autocorrelation function should be calculated was the changing point from negative to positive in the waveform. However, similar results can also be obtained by treating the applied impulse as negative. In this case, the point of change from positive to negative in the waveform may be selected as the point of change at which the autocorrelation function should be calculated.

(Effects of the Invention) As explained above, according to the pitch extraction method of the present invention, the time lag for calculating the autocorrelation function is the time difference between points of change in a certain direction in a waveform, for example from negative to positive. Furthermore, when determining the change point, set two threshold values at an appropriate interval, and limit the change point to t-1 while passing through the range defined by these thresholds. As a result, the number of time lags for calculating the autocorrelation function is reduced, and the number of multipliers, integrators, and dividers for calculating the correlation function is greatly reduced, realizing an economical pitch extraction device. be able to. According to an example that was actually designed and manufactured, it was possible to extract the pitch by calculation for only a few tap positions, and the result was 1/20th of the 113 points of the conventional method.

[Brief explanation of the drawing]

FIG. 1 is an explanatory diagram showing the vicinity of a polarity change point in an audio waveform, and FIG. 2 is a block diagram showing the configuration of an embodiment of the pitch extraction method of the present invention.
・Changing point detection circuit, garden...threshold generation circuit, 40・
・・Counter for counting the change point interval, male ・correlator, ■・
...Tap position setting circuit, 70...Tap position control circuit, (capital)...Pitch determination circuit, 90...Output terminal. Patent Applicant: Fujitsu Limited Representative Patent Attorney: Hisashi Tamamushi Department 5 (3 others) 1st
Figure 2

Claims (1)

[Claims] In the pitch extraction method, which extracts the pitch of the voice based on the autocorrelation function value of the input waveform obtained from the correlation with the nine waveforms given a certain time lag between the input waveform and this, a threshold value having an arbitrary width above and below zero. means for detecting a changing point that crosses the range in one direction; means for counting the interval between the changing points; and determining an autocorrelation function value of the input waveform by setting a time lag corresponding to the interval between the changing points. 1. A pitch extraction method comprising: a correlator, and extracting a time shift corresponding to the maximum autocorrelation value determined by the correlator as a pitch of an input waveform.