JPH0234400B2 - ONSEINOPITSUCHICHUSHUTSUHOHO - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a voice pitch extraction method, and more particularly to a method for obtaining voice pitch periods required for voice analysis, voice synthesis, voice recognition, etc. in a short time.

Speech can be viewed as repeating waveforms with very little difference in waveform. The fundamental period of voiced sound, which is this repetition, is the pitch of the voice, and is usually 2.5ms.
It has been found that it is about ~12ms. Accurate extraction of the pitch of speech is an important matter when analyzing and synthesizing speech, and especially in speech synthesis, the pitch period is an important parameter for data compression and synthesized sound quality. Therefore, the pitch period is required to be accurate.

Conventional methods for determining the pitch period include searching for the peak value of the audio waveform,
A method of obtaining it from autocorrelation coefficients has been proposed. However, although the former has the advantage of short calculation time, it has the disadvantage of poor extraction accuracy in the transient state where the amplitude is at its maximum or decreases, and the latter has the disadvantage that the autocorrelation coefficient Since the time delay correlation coefficient including the pitch period must be calculated in order to obtain the Pitch period, the calculation time is long, and since there are many product operations, a multiplier is required, which complicates the hardware configuration. have

SUMMARY OF THE INVENTION In view of the above-mentioned drawbacks, it is an object of the present invention to provide a voice pitch extraction method that shortens calculation time and simplifies the circuit configuration on the premise of obtaining a highly accurate pitch period.

In order to achieve this object, the present invention adds the above-mentioned audio signals excluding the DC component from each point in time when the waveform of the audio signal crosses zero in one analysis interval stored in the input audio storage means, and the addition result is The method is characterized in that the time until the value becomes equal to or less than a predetermined threshold is determined for each calculation from each of the above-mentioned points, and the pitch is obtained from the maximum value of each of the times.

Here, embodiments of the present invention will be described with reference to the drawings. In FIGS. 1 and 2, an audio signal is input to an input audio storage means 1 and is stored for one analysis period including several pitch periods. The zero-crossing detecting means 2 determines the point in time when the audio signal changes from negative to positive (the zero-crossing point in this embodiment) from the stored audio signal, and the adding means 3 calculates the zero-crossing point from negative to positive.
Output to. The addition means 3 sequentially reads out the audio signals stored in the input audio storage means 1 from the zero-crossing point and adds them until the addition result reaches a predetermined threshold, for example, zero.

For example, in the case of the audio signal shown in FIG _.
T ₂ , T ₃ , T ₄ , and T ₅ are sequentially obtained, and calculations are performed from each time point T ₁ to _{T 5} until the addition value becomes zero.

The fact that the sum is added until the sum reaches a predetermined threshold value is based on the phenomenon that the average value within one pitch period of the audio signal waveform is close to zero. That is, if we add the audio signals from the negative to positive zero crossing point, we get
At each zero crossing, addition of positive and negative values is repeated, and the added value gradually decreases and approaches zero. In this case, the threshold value can be within a certain range with zero as a reference.

The addition time measuring means 4 is a time period from when the addition means 3 starts addition until the added value reaches a predetermined threshold value.
Find it and temporarily memorize it. In FIG. 2, the times a, b, c, d, and e from the time points T ₁ , T ₂ , T ₃ , T ₄ , and T ₅ at which the addition starts until the added value becomes zero are measured.

The addition time determined by the addition time measuring means 4 and the value stored in the maximum value storage means 6 are compared by the maximum value determination means 5, and the larger value is stored in the maximum value storage means 6 as the maximum value. . In this case, the value stored by the maximum value storage means 6 is initially a minimum value that is considered to be zero or a pitch period, and the larger value is determined to be the maximum value when the addition time is measured.

The above procedure is carried out for the audio signal within one analysis section, and the maximum value finally determined is used as a pitch period candidate. In the audio signal of FIG. 2, the pitch period candidate is C.

Normally, the pitch period is distributed within a certain range, for example between 2.5ms and 12ms, as mentioned above, so
Accuracy can be improved by determining whether or not the value found as a pitch period candidate falls within this predetermined range. Further, the above-mentioned one analysis interval may be about twice the normal pitch period maximum value of 12 msec.

Furthermore, the audio signal to be analyzed needs to be passed through a high-pass filter to remove the DC component and allow only the AC component to participate in the addition.

Furthermore, since the pitch frequency is about 80Hz to 400Hz, accuracy can be improved by passing the audio signal through a low-pass filter of about 1KHz.

In the above example, the addition was performed using the zero-crossing point from negative to positive as the reference, but the zero-crossing point from positive to negative may be used as the reference.

As embodied in the above embodiments, according to the present invention, the pitch period is calculated by calculating the addition time using the zero crossing of the audio signal as a reference, so that the pitch period can be calculated with high accuracy and the time delay phase can be calculated as in the conventional method. Since there is no calculation of relational numbers and the calculation is centered on addition, the calculation time is greatly reduced, and a multiplier is no longer required, making the hardware configuration simple.

[Brief explanation of drawings]

1 and 2 show an embodiment of the voice pitch extraction method according to the present invention. FIG. 1 is a block diagram showing the principle, and FIG. 2 is an audio signal waveform diagram for explaining the embodiment. In the drawing, 1 is input voice storage means, 2 is zero crossing detection means, 3 is addition means, 4 is addition time measurement means,
5 is a maximum value determining means, and 6 is a maximum value storage means.

Claims (1)

[Claims] 1 Add the audio signals excluding the DC component from each point in time when the waveform of the audio signal crosses zero in one analysis interval stored in the input audio storage means, and add the audio signal until the addition result becomes equal to or less than a predetermined threshold. A voice pitch extraction method in which the time is determined for each calculation from each point in time, and the pitch is obtained from the maximum value of each time.