KR19990055873A - A Method for Reducing the Computation of G.723.1 Speech Coder by Negative Waveform Skip of Autocorrelation Function - Google Patents

Abstract

The present invention is to shorten the pitch search time in speech encoding. In the method of encoding a speech signal based on a pitch delay obtained without maximum correlation, a section in which a positive correlation value appears in pitch search is obtained. The method of reducing the calculation amount of the G.723.1 speech encoder by skipping the negative waveform of the autocorrelation function, skipping the next section by the period in which the positive correlation value appears.

Description

A Method for Reducing the Computation of G.723.1 Speech Coder by Negative Waveform Skip of Autocorrelation Function

The present invention relates to speech encoding, and more particularly, to an encoding method by pitch delay.

One of the coding methods that have recently been standardized and attracted attention is G.723.1. G.723.1 consists of a dual rate voice encoder of 5.3 / 6.3 kbps ACELP / MP-MLQ (Algebraic Code-Excited Linear Predictive / Multi-Pulse-Maximum Likelyhood Quantinization). In encoding a double rate speech coder, the input signal is sampled at 8 kHz and quantized with 16 bit linear PCM to have a 30 ms frame length with 240 samples. And a high pass filter is passed to remove the DC component. The transfer function of the high pass filter is shown in Equation 1.

One frame of 240 samples is divided into four subframes, and the 10th linear prediction (LP) coefficient is calculated using the autocorrelation method for the subframe, and the LP coefficient of the fourth subframe is LSP (Line spectrum). After converting to a pair) value and vector quantization using Predictive Split Vector Quantinization (PSVQ). Then, the estimated pitch (inverse of the fundamental frequency, which has the speaker's personality information and emotion) is found in the open loop method of 18 to 142 for two subframes (120 samples).

The closed loop pitch prediction is performed using the estimated pitch delay value, and the order of the predictor used here is 5th order. Finally, the excitation signal is modeled using ACELP at 5.3kbps and MP-MLQ at 6.3kbps. 1 and 2 show a coding and decoding block diagram of G.723.1.

The pitch delay search time requires the largest amount of computation after the codebook search time. Looking at the pitch delay search process in detail, the pitch delay search is calculated using the autocorrelation function according to the time delay for the speech signal f (n). In order to find the most desirable pitch delay in the pitch search method, the correlation equation (2) must be repeated for all possible pitch delays.

Where f [n] is the speech signal passing through the adaptive weighting filter, and j is the pitch delay value. This is a large amount of computation that requires multiplying and adding j times for every pitch delay j (the range of j from 18 to 142 samples). Therefore, the pitch search error should be able to shorten the pitch search time only when the pitch does not change.

Accordingly, the present invention is to propose a method for shortening the pitch search time.

1 is a block diagram of an encoder of G.723.1;

Fig. 2 is a block diagram of a decoder of G.723.1.

3 shows examples of conventional waveforms and waveforms according to the present invention, (a) shows an original sound waveform, (b) shows an autocorrelation waveform, and (c) shows a waveform of the proposed invention.

The autocorrelation function is used to find the pitch delay value. In this pitch delay search block, if the autocorrelation function is taken based on the positive waveform peaks, a positive value is obtained for each positive waveform peak, and a negative value is obtained for each negative waveform peak. Therefore, the correlation waveform also has a structure in which negative and positive peaks appear repeatedly over time.

Since the pitch delay that obtains the maximum correlation value in the pitch search process is considered as the pitch period, the correlation value applied to the pitch search needs only the case of forming positive peaks. In addition, if a positive peak of the correlation value appears, the interval does not need to calculate the correlation value since the positive peak width is the interval of the positive peak from the end of the peak. In the present invention, unnecessary pitch search time is reduced by using the property of the negative waveform skipping of the autocorrelation function.

Hereinafter, the present invention will be described in detail.

The procedure for finding the optimum pitch delay value is as follows. First, a pitch search through an open loop is performed twice per frame, that is, once in a subframe (0, 2) of four subframes (subframes 0, 1, 2, and 3) for a speech signal passing through an adaptive weighting filter. . Pitch search through open loop uses autocorrelation function as in Equation 2. At this time, the pitch delay value at which C _OL is maximum, that is, index j is transmitted. At this time, a correlation value is obtained close to 100% for each pitch period, and the degree of similarity depends on the periodicity of the waveform within the pitch search interval and the change in the waveform amplitude. However, in the case of voiced sound with strong pitch periodicity, more than 70% is usually obtained, and the correlation value shows a slow change with time delay.

Calculation of the correlation of the speech waveform through Equation 2 As shown in Figure 3 (b), the waveform has the following characteristics.

a) Slow change: Because of the high correlation between neighboring samples in the speech waveform, the correlation waveform peak representing this correlation also changes slowly.

b) Peak width: Since the energy of the first formant (lowest resonance peak) is the highest in the waveform of the voiced sound, the waveform change in one pitch period causes the damping oscillation to the structure of the first formant. The autocorrelation waveform for this also constitutes a waveform which maintains the period of the first formant as it is. Thus, the correlation waveforms have a constant width between pitch periods.

c) Correlation value: Voice waveform forms one pitch period by repeating alternating positive and negative waveforms. The correlation is obtained based on the positive waveform, and a positive value is obtained for each positive waveform, and a negative value is obtained for each negative waveform. On the contrary, on the basis of the negative waveform, a positive value is obtained for each negative waveform and a negative value is obtained for each positive waveform. Therefore, the correlation waveform also has a structure in which negative and positive waveforms appear repeatedly with time delay.

Since the pitch delay that obtains the maximum correlation value in the pitch search process is considered as the pitch period, the correlation value applied to the pitch search needs only a positive value. In addition, when the characteristics of the correlation and the positive interval starts to appear, the interval does not need to calculate the correlation value because the width of the positive interval exists from the portion where the positive interval ends. That is, when a positive correlation value is found as shown in FIG. 3 (c), a section having a negative correlation is assumed and a pitch section is skipped for this section without a pitch search. The number of times is counted again until a negative interval is obtained, and the number of times is skipped by the positive interval as in the previous method. The skipped interval is a correlation value of zero and the pitch is searched. Therefore, in the present invention, unnecessary pitch search time is reduced by skipping the negative waveform of the autocorrelation function.

In order to reduce the pitch search time of G.723.1, a new pitch search method using waveform characteristics of autocorrelation function is proposed. Table 1 shows the processing time of the entire encoding for the speech samples to find the pitch search time difference between the two processes. In the method proposed in the present invention, there was almost no deterioration in sound quality compared to the existing method, and time saving of about 31.3% was achieved. In this case, since the time measurement value is different depending on the computer type, only the relative time reduction rate is considered in the evaluation.

Therefore, it is possible to significantly reduce the calculation time without deterioration of the sound quality, so that it can be used to extend battery life and to implement real-time processing in mobile communication fields and other applications using voice compression.

Table 1. Comparison of results of the proposed method with the existing method

Claims (1)

In a method of encoding a speech signal based on a pitch delay at which a maximum correlation value is obtained, Find the interval where the positive correlation value appears in the pitch search. Search for the pitch by skipping the next section by the period in which the positive correlation value appears A method for reducing the calculation amount of a G.723.1 speech encoder by skipping negative waveforms of an autocorrelation function.