JP2940835B2 - Pitch frequency difference feature extraction method - Google Patents

Abstract

PURPOSE:To leave a correct pitch frequency difference feature quantity even if plural peaks appears in the autocorrelation function of a predicted residue for detecting a pitch frequency. CONSTITUTION:An input voice is converted into a digital signal and a residue calculation part 3 calculates the autocorrelation function of the predicted residue, frame by frame. A residue correlation calculation part 4 finds the cross- correlation function between the autocorrelation function of the predicted residue found in each frame and the autocorrelation function of a predicted residue which is seven frames precedent and obtains the pitch frequency difference feature quantity from the peak value of the cross-correlation function.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention is used in a word speech recognition device, a continuous speech recognition device, a speaker recognition device, or the like that requires difference information of pitch frequency, which is one of the elements of speech prosody information. And a pitch frequency difference feature amount extraction method for acquiring information about accents and intonations.

[0002]

2. Description of the Related Art Conventionally, in order to obtain pitch frequency difference information, the following method has been employed. The short-time spectrum f (r) ^h of the input signal is converted to the spectrum envelope f (r)
The spectral fine structure obtained by normalizing has a harmonic structure. This normalized spectrum f (r) ^h
The correlation function obtained by Fourier-expanding / f (r) represents the property of the sound source by the correlation function of the prediction residual. (For example, Sadahiro Furui, “Digital Speech Processing,” Tokai University Press (1985)) Therefore, first, the autocorrelation function of the prediction residual in the frame at time t is calculated, and after detecting the peak position of the correlation value, the pitch is calculated. Determine the frequency. FIG. 3A
Shows an example of the autocorrelation function of the prediction residual of the voiced stationary part, where the horizontal axis is the time delay width and the vertical axis is the correlation value. The sharp peak in the figure represents the pitch frequency component, and the time delay width of the peak position becomes the pitch period, and the pitch frequency can be determined. Similarly, the pitch frequency in the frame at time s is determined. Finally, the pitch frequency at time t and time s
The difference information is obtained from the difference from the pitch frequency.

However, in voiced and unvoiced transitional portions and sentence endings of voices, since the pitch structure of voices is disturbed, a peak indicating a true pitch frequency component in the autocorrelation function does not appear remarkably, and the double pitch frequency and half pitch frequency do not appear. The peak indicating the pitch frequency component may exceed the true peak. FIG. 3B shows that the peak of the double pitch component exceeds the peak of the true pitch component,
This is an example in which an incorrect pitch frequency is calculated. When the wrong peak is selected and the pitch frequency is determined, the true pitch information is completely lost. Therefore, in order to obtain correct pitch frequency difference information, when the obtained pitch frequency becomes discontinuous with the pitch frequency values of the preceding and following frames, the second peak and the third peak of the autocorrelation function of the prediction residual are detected. Then, it was necessary to verify whether or not it included a pitch frequency continuous with the pitch frequency of the surrounding frame. However, it is difficult to describe general rules for this verification, and there has been a problem that in addition to automatic correction, correction by inspection is necessary.

[0004]

According to the present invention, the process of uniquely determining the pitch frequency in each frame is avoided, and the autocorrelation function of the prediction residual of speech at a certain time t (frame) and the time t A cross-correlation function of the prediction residual of the voice and the auto-correlation function at a time s (a frame several frames away) from the voice is calculated as a pitch frequency difference feature vector. Calculating this cross-correlation function means that the data sequence of the auto-correlation function of the two prediction residuals is
Since the correlation function is calculated and shifted as a feature vector element by shifting by one point (one reference delay time in the autocorrelation function), the element of the feature vector takes a large value when the peak positions overlap each other, Difference information is obtained from the peak value. In this way, the information on the peak position of the autocorrelation function of the mutual prediction residuals can be faithfully reflected on the pitch frequency difference feature vector.

[0005]

FIG. 1 shows an embodiment for calculating a pitch frequency difference feature quantity according to the present invention. The audio input from the input terminal 1 is converted into a digital signal in the A / D converter 2. The digital signal is converted to 1
The autocorrelation function of the prediction residual is calculated every frame (for example, every 10 milliseconds). The autocorrelation function of the prediction residual is calculated by Fourier expansion of a normalized spectrum f (r) ^h / f (r) obtained by normalizing the short-time spectrum f (r) ^h with the spectrum envelope f (r). Alternatively, after obtaining the prediction residual waveform, the correlation function may be directly calculated.

Next, the residual correlation calculator 4 calculates the cross-correlation function between the autocorrelation function of the prediction residual obtained for each frame and the autocorrelation function of the prediction residual several frames ahead (for example, 7 frames ahead). Ask. For example, sampling frequency 1
Consider male voice data sampled at 2kHz. In this case, one reference delay time of the autocorrelation function is 1/12000 second. Change range of pitch frequency of male voice from 50Hz to 3
Assuming that the frequency is 00 Hz, in the autocorrelation function a (n) of the prediction residual, the peak representing the 50 Hz pitch component is at the 240th point (12000/50 = 240), and the peak representing the 300 Hz pitch component is at the 40th point (12000). / 30
0 = 40). Accordingly, 201 data from a (40) to a (240) are cut out and newly set as x (n) (n = 0, 1,..., 200). Similarly, y (n) (n = 0, 1,..., 200) is obtained seven frames ahead, and a cross-correlation function Rxy between x (n) and y (n) is obtained.
(Τ), that is, a feature vector is obtained.

[0007] Rxy (τ) = (1 / T) · Σ t x (t) y (t + τ) Σt shows the addition of from t = 0 to t = T. Where T =
In 200, τ is set to −40 ≦ τ ≦ 40 in consideration of a difference between peak positions that can move between seven frames, that is, a pitch frequency that can change in seven frames. This is 5 in the low frequency range.
Covers changes from 0Hz to 60Hz. The 81-dimensional feature vector Rxy (τ) obtained in each frame is set as a pitch difference feature amount in the frame, and output from the result output unit 5. Of course, τ can be changed depending on the properties of the target audio data. In a system in which the number of dimensions is too large to handle in 81 dimensions, the feature vectors are examined, for example, two dimensions at a time, and elements with large numerical values are selected and merged into 41-dimensional feature vectors. Such a reduction in the number of dimensions may be performed.

[0008]

As described above, according to the present invention, the pitch information appearing in the autocorrelation function of the prediction residual can be faithfully incorporated into the feature quantity, so that a feature quantity with little information loss can be obtained. In addition, verification by inspection after automatic correction of pitch frequency is unnecessary. Particularly at the end of a sentence, at the end of a long word, or at the end of a phrase, it is effective because it is difficult to uniquely determine the correct pitch frequency.

For example, it is assumed that the autocorrelation function x (n) of the prediction residual as shown in FIG. 3A is obtained at time t, and the peak position is at the 100th point. At time s several frames ahead, the autocorrelation function y (n) of the prediction residual as shown in FIG. 3B is obtained.
Assume that two appear at the 0th point. These are simultaneously shown in FIG. 2A. At this time, the peak representing the true pitch component is the 95th peak (that is, the pitch frequency is 126 Hz).
Despite this, if the peak at the 110th point becomes larger, if the pitch frequency is determined at each time as in the past, the pitch frequency at the time s becomes 1
It becomes 09Hz. At this stage, the correct pitch information is completely lost. However, in the feature quantity according to the present invention, the cross-correlation function Rxy of the two autocorrelation functions is calculated. In the case of the above example, the peak at the 100th point at the time t shown in FIG. Rxy (-5), which overlaps with the peak at the point, and Rxy (10), where the peak at the 100th point at time t overlaps the peak at the 110th point at time s shown in FIG. Is a pitch difference feature amount that does not lose the information of the peak position.

[0010] Pitch difference information that provides information on accent and intonation improves recognition performance by being used in combination with phonemic information in word speech recognition and continuous speech recognition. Also, in speaker recognition, pitch difference information is effective as a feature representing individuality. As described above, pitch difference information including information on two peak positions in y (n) is obtained. However, when the subsequent processing is statistical processing, the correct pitch difference information is saved.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an example of a pitch frequency difference feature value extraction device to which the present invention is applied.

FIGS. 3A and 3B show examples of calculating a cross-correlation function of an auto-correlation function of two prediction residuals in FIGS. 3A and 3B, respectively.

FIG. 3A is a diagram illustrating an example of an autocorrelation function of a prediction residual in a voiced stationary part, and FIG. 3B is an autocorrelation function of a prediction residual in a frame of a voiced and unvoiced transient part. It is a figure which shows the example which the peak which shows a pitch frequency exceeds the peak which shows a true pitch frequency.

──────────────────────────────────────────────────続 き Continued on the front page (58) Field surveyed (Int.Cl. ⁶ , DB name) G10L 3/00-9/20 JICST file (JOIS)

Claims (1)

(57) [Claims] 1. a pitch frequency of a voice at a time t;
In a pitch frequency difference feature amount extraction method for providing difference information from the pitch frequency of the voice at time s after time t, the residual autocorrelation function of the voice at time t and the residual of the voice at time s A pitch frequency difference feature amount extraction method, wherein a difference autocorrelation function and a cross-correlation function are calculated, and the difference information is obtained from a peak value of the residual cross-correlation function.