JPS63142400A - Detection of max point of self- correlation function given by formant - 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] [Summary] In a pitch extraction method that uses an autocorrelation function, in order to reduce the error of determining that the maximum point of the autocorrelation function given by the formant representing the peak of the frequency characteristics of the vocal tract is a pinch,
The longest interval from the period or zero crossing of the function given by the integral of the autocorrelation function is taken out, and this is removed by using this as a criterion to detect the maximum point given by the formant from the maximum point of the autocorrelation function. This is used to reduce errors in pinch extraction.

[Industrial application field]

The present invention relates to a pinch extraction method using an autocorrelation function, and in particular to a method for detecting the maximum point of an autocorrelation function given by a formant, which can be used as a means to prevent the maximum point of an autocorrelation function given by a formant from being mistaken as a pitch. It is.

Speech processing systems that use telephones as input devices are a field that is highly expected to be used in the future.

A bandpass filter is inserted into the telephone line, and the fundamental frequency component corresponding to Pisochi is significantly attenuated. Although it is expected that pinch extraction of such band-limited speech will be very difficult, pitch is an important parameter in speech processing, and it is important to establish a pitch extraction method that is stable and has few extraction errors. This is something that is strongly desired.

If a pitch extraction method can be established, there will be many benefits, such as the ability to accurately calculate a spectrum by so-called pitch synchronization analysis, in which one pitch is extracted and Fourier transformed to obtain a frequency spectrum. For this reason, various pitch extraction methods have been proposed in the past, and the present invention relates to a method for reducing extraction errors in methods that utilize autocorrelation functions.

Extraction errors in pitch extraction methods that use autocorrelation functions include the error of assuming that the maximum point of the autocorrelation function given by the formant is the pitch, and the error of assuming that the maximum point of the autocorrelation function given by an integral multiple of the pitch is P-7. There are two mistakes. The present invention provides a method for detecting the maximum point of an autocorrelation function given by a formant, which can be used as a means to reduce the former of these two errors.

[Conventional technology]

FIG. 4 is a block diagram showing the circuit configuration of a conventional example.

A method often used as a means to reduce the error caused by pinching the maximum point of the autocorrelation function given by the formant described above is to configure the autocorrelation function to be calculated so that the maximum point given by the formant does not occur. . For example, there is a method of placing a low-pass filter in front of the autocorrelation function calculation circuit, and this configuration is shown in FIG.

In FIG. 4, high frequency components, ie, formants, are removed from the audio signal input from the input terminal 10 by a low-pass filter 90 whose rejection frequency is set to about 900 Hz. The pitch is extracted by applying the output of the low-pass filter 90 to the pitch extraction logic circuit 60' via the autocorrelation function calculation circuit 30.

In this way, the pitch is extracted by removing high frequency components using a low-pass filter and emphasizing the fundamental frequency components corresponding to the pinch.

[Problem that the invention seeks to solve]

However, the circuit configuration using the above-described low-pass filter is not suitable for a system that processes voice signals via a telephone line. This is because, in such an audio signal, the fundamental frequency component corresponding to the pitch is significantly attenuated by the bandpass filter, and even when the high frequency component is removed by the low-pass filter, the fundamental frequency component is not emphasized much.

Conversely, the purpose of suppressing formant components is not achieved, and low-order formants are emphasized by suppressing high-order formants and fundamental frequency components. This had the problem of causing pinch extraction errors.

Furthermore, when considering application to pinch synchronization analysis, a low-pass filter is placed in front of the autocorrelation function calculation circuit, so it is necessary to provide a separate autocorrelation function calculation circuit without a filter for spectrum calculation. be. This had the problem of being an uneconomical configuration.

[Means for solving problems]

As shown in Fig. 1, the above problem is caused by removing formants from the longest period or zero crossing interval given by the integration result of the autocorrelation function of the output of the autocorrelation function calculation circuit 30 that calculates the autocorrelation function of the input signal. The maximum point within this criterion is determined by the formant from the output of the formant removal criterion creation means 40, the output of the formant removal criterion creation means 40, and the output of the autocorrelation function calculation circuit 30. This problem is solved by the method of detecting the maximum point of the autocorrelation function given by the formant of the present invention, which includes a formant detection means 50 that determines and detects the maximum point.

[Effect] The time point indicating the maximum point of the autocorrelation function is expressed as BMAX(k), = 1 to . Furthermore, let the autocorrelation function be Ra(i),
The function obtained by integrating it is expressed as Ra (i)=SGN (4R(j)) (2). Here, SGN (X) represents the polarity of X.

Now, in the above equation, the integration has a so-called low-pass filtering effect in which the alternating current component decreases and the direct current component remains, and the blocking frequency decreases as the integral interval i increases. As a result, high frequency components are suppressed, low frequency components are emphasized, and the vibration of the function Ra (i) asymptotically approaches the fundamental frequency. That is, the period of vibration KMAX is the fundamental period (pitch) PI
Approaching TCH, K M A X → PITCH,
Letting the longest period be KAMAX, KAMAX≦PIT
It is CH.

On the other hand, since the maximum point of the autocorrelation function representing a period less than the fundamental period is given by the formant, K A M
The maximum point existing below A% can be determined to be the maximum point given by formant.

That is, B satisfies BMAX (k) < KAMAX ■
MAX (k) is the maximum point given by the formant.

Further, the maximum points Po, Pl, 'Pyu, -m- of the autocorrelation function corresponding to the pitch and points that are integral multiples thereof are arranged in order with the pitch as an interval. And the interval is K A M
Since A is larger than KAMAX from the definition of
The adjacent Po and Pl are never included.

Conversely, this is before and after Pl + Kt'.

The maximum point of the autocorrelation function included in the interval of M A
This means that it is the maximum point given by the formant.

Of course, this is not limited to Pl, but also Pa,
This holds true for P2, P3, -m-.

Therefore, when PITCIIX is estimated as the pitch, the maximum point of the autocorrelation function (most likely to be Bi-7chi), the area around it, that is, PITCIIX -K A
M A X, and PITCIIX + K A M A
The maximum point included in X is BMAX (PX) = PIT
With the exception of CIIX, it can be determined that the formants are given.

The above can also be achieved by the method described below.

That is, in the formant removal criterion creation means 40,
The longest positive or negative interval of Ra (i) (hereinafter referred to as KBMAX), that is, the longest zero-crossing interval of Ra (i) is selected as the criterion.

In this case, KBMAX<KAMAX ■, so K
As in the case of AMAX, BMAX (k)<KBMAX ■PITCH
X-KBMAX<BMAX(k)<PITCHX
BMAX (k) that satisfies 10 K B M A X can be determined to be a formant.

That is, BMAX(k) that satisfies the above equations 0 and 0 is determined to be a formant, and is detected by the formant detection means 50.

〔Example〕

FIG. 2 is a block diagram showing the circuit configuration of an embodiment of the present invention.

FIG. 3 is a waveform diagram obtained by an embodiment of the present invention,
The horizontal axis shows time, and the vertical axis shows (1) dimensionless in the case of an autocorrelation function waveform, and (2) voltage in the case of an audio signal waveform.

The same reference numerals indicate the same objects throughout the figures.

In FIG. 2, the autocorrelation function of the audio signal input from the input terminal IO and the time point at which its pitch is estimated are obtained by the autocorrelation function calculation circuit 30,
and is transferred to the formant detection circuit 50'. The zero crossing detection circuit 70 integrates the above autocorrelation function and detects the point of change in polarity. The change point is a zero crossing, which is transferred to the longest zero crossing interval detection circuit 80. In response to this, the longest zero-crossing interval M detection circuit 80 determines the longest interval between the changing points, and sends this as KBMAX to the formant detection circuit 50'. The formant detection circuit 50' is KBMA
Formant detection is performed using X, and the above-mentioned equation 0 and equation 0. The pitch extraction logic circuit 60 uses the detection result to extract the pinch from the maximum point of the autocorrelation function.

FIG. 3 (1) shows a waveform diagram of the autocorrelation function and Ra (i) obtained by the embodiment of the present invention.
Formant detection is performed in the section of J:DKBMAX below ITCIIX.

〔Effect of the invention〕

As described above, according to the present invention, the autocorrelation function is calculated without a low-pass filter, so the calculation result can be used as is for spectrum calculation.

Furthermore, since there is no need to emphasize low-frequency components, even audio that has been band-limited by a bandpass filter can be used without being aware of the band-limiting.

[Brief explanation of the drawing]

Fig. 1 is a principle diagram of the present invention, Fig. 2 is a block diagram showing the circuit configuration of an embodiment of the present invention, Fig. 3 is a waveform diagram obtained by the embodiment of the present invention, and Fig. 4 is a circuit of a conventional example. FIG. 2 is a block diagram showing the configuration. In the figure, 10 is an input terminal, 20 is an output terminal, 30 is an autocorrelation function calculation circuit, 40 is a formant removal criterion creation means, 50 is a formant detection means, 50' is a formant detection circuit, 60.60' is a pinch extraction logic circuit, 70 is the zero crossing detection circuit, 80 is the longest zero crossing interval detection circuit, 90 is the low-pass filter A. J each, 148〉Professional...
J7Z No. 2 Figure 1! aJ annex n times J Raku, the composition is 1st floor...12th figure閤(2)

Claims (1)

[Claims] Formant removal in which the longest period or zero-crossing interval given by the integration result of the autocorrelation function output from the autocorrelation function calculation circuit (30) that calculates the autocorrelation function of the input signal is used as a criterion for formant removal. The output of the judgment criterion creation means (40), the formant removal judgment criterion creation means (40), and the autocorrelation function calculation circuit (30)
) formant detecting means (50) for determining and detecting that the maximum point within the criterion is the maximum point given by the formant, based on the output of the formant. How to detect points.