KR19990024267A - Speech Waveform Coding Using Pisola by Formant Similarity Measurement - Google Patents

Abstract

The main purpose of speech coding depends on factors such as speech compression for transmission over band-limited channels, high quality speech synthesis to maintain clarity and naturalness, and processing speed. In general, a speech encoding method may be classified into a waveform encoding method, a signal source encoding method, and a hybrid encoding method.

The method proposed in the present invention is a new voice encoding method using a pitch-synchronous overlap add (PSOLA) technique. First, the basic period is fixed at every pitch period, and the formant similarity between the basic pitch and the adjacent pitch is measured. Then, if a certain threshold is satisfied, it is compressed and stored or transmitted in pitch units. As a result, even when the compression ratio is about 45%, MOS 4 or more was obtained.

Description

Speech Waveform Coding Using Pisola by Formant Similarity Measurement

In general, quantized speech samples are coded using B bit codewords. Sampling at rates tells you how much information you need to send or store With the B bit Denotes the product of. In general coding, the sound quality is maintained Need to be lowered. However, the waveform coding method mainly focuses on reducing the number of quantization bits per sample since the sampling rate of the speech signal is already determined by Nyquist sampling theory in order to preserve the waveform shape of the speech signal. According to the method of reducing the number of bits, ADPCM (Adaptive Differential PCM), ADM (Adaptive Delta Modulation) and the like have been proposed. Waveform encoding can maintain high sound quality and individuality, but requires a lot of memory due to the huge amount of data. In recent years, the manufacturing technology of the chip dedicated to digital signal processing and the analysis and synthesis algorithm of the waveform encoding method have been well developed, and the standardization of ADPCM having a data rate of 32kbps, 16kbps, etc. has been realized. However, there is a problem that a separate DSP chip must be used when using the ADPCM.

The present invention proposed a new waveform encoding method for compressing speech using a method of measuring form similarity between a reference pitch waveform and a neighboring pitch waveform in pitch units and transmitting or storing only pitch information and amplitude information when the similarity is high. In compression, the similarity was measured by the area of the cross normalized AMDF waveform, and the PSOLA technique was used for the synthesis.

In case of using the proposed method, it can be synthesized using a general-purpose chip, which can solve the problems of ADPCM described above.

1.NAMDF (Normalized AMDF)

As a method of measuring the pitch of the current frame, NAMDF may be defined and used as follows.

Where s (n) is the voice signal and N is the window period to obtain the NAMDF. When NAMDF is obtained by gradually increasing the delay factor d, the NAMDF becomes almost zero whenever the delay factor becomes an integer multiple of the voice pitch in the frame.

Figure 1 Error function comparison of primary and secondary functions

Figure 2 (a) Speech Waveform (b) NAMDF Waveform

In Figure 1 With graph The graph shows the graph at the peak point when autocorrelation function and AMDF are taken. Looking at the zero position, we find that finding the exact peak value of the autocorrelation function is more difficult than finding the peak value of AMDF. For this reason, it is possible to obtain an error during pitch search by obtaining an incorrect peak value during pitch search, and AMDF has been applied for a long time to emphasize periodicity instead of autocorrelation function. AMDF also has the advantage of not using multiplication. However, in normalization, one division can maintain the advantages of NAMDF because it does not significantly affect the overall computation. Figure 2 shows a waveform of NAMDF applied to an audio signal.

In the present invention, the pitch is searched using NAMDF and the similarity measurement interval is determined. In addition, the change of the peaks in one section can be measured using the Cross NAMDF method. In the present invention was applied to formant similarity measurement using the Cross NAMDF method.

2. Measuring the similarity of formants

Observing the voiced sound section shows that the formant changes little by little even in the interval where the pitch is kept constant. This formant's information appears between one pitch period

It depends on the number, shape, size and location of the peaks. Therefore, the characteristics of the peaks appearing in the reference pitch and the adjacent pitch period were compared to measure the similarity of formants.

To compare the characteristics of the peaks appearing in one period, Cross NAMDF was performed on the reference and one pitch waveforms. Cross NAMDF is shown in Equation 2.

here Is the waveform of the pitch period Is the waveform of the pth period. N is the window size Wow The smaller of the lengths. d is a delay factor.

Figure 3 Cross NAMDF Waveform

Figure 3 shows the result of Cross NAMDF between the reference pitch period and the waveform of the pth period. However, it is symmetrical by repeating the same pitch in two cycles. The area of the obtained waveform is calculated as in Equation 3.

here Is the area of the Cross NAMDF waveform of the p-th waveform. The similarity is measured by comparing the obtained area with the area of the NAMDF waveform at the reference pitch period.

Similarity measurement is shown in Equation 4.

Is the area of the NAMDF waveform of the reference waveform, Is the area of the Cross NAMDF waveform of the reference waveform and the neighboring waveform obtained as in Equation 3. D (p) represents the formant similarity of the p-th waveform, and the smaller the value, the more similar the reference waveform is to the p-th waveform. Figure 4 shows the change in compression rate as the threshold of D changes.

Fig. 4 Compression rate of pitch period according to threshold

(First bar (D = 0)-number of pitch cycles transmitted, others (D0)-Number of pitch cycles compressed)

(a) When D = 1 as the threshold (45.6%)

(b) When D = 2 as the threshold (30.4%)

(c) When D = 5 as the threshold (23.9%)

3. Voice Synthesis by PSOLA Technique

In the present invention, a PSOLA method having low spectral distortion and complexity when reconstructing a voice signal is suitable. PSOLA synthesis is performed using the transmitted or compressed waveform, amplitude information, and pitch information.

Figure 5 shows the process of synthesis using the PSOLA technique. (a) is the original audio waveform, (b) is a picture showing pitch information, and (c) is one periodic waveform for synthesis. (d) is a waveform synthesized using the PSOLA method.

Fig. 5 Example of processing in pitch units

4. Experiment and Results

To simulate the method proposed in the present invention, a 16-bit A / D converter capable of a microphone input to IBM-PC / Pentium-150MHz was interfaced and stored in 16-bit quantized at a sampling rate of 11 kHz. In the simulation, 256 units of pitch analysis frame were used and coded by pitch period unit.

Figure 6 is a block diagram of the method proposed in the present invention. The transmitter first finds the pitch using the NAMDF method for one frame. The pitch is determined by the distance to the valley closest to the zero point in Figure 2 (b). One period corresponding to the obtained pitch is defined as a reference waveform and stored or transmitted. (The first reference waveform is the waveform corresponding to the first pitch period.) Extract the amplitude information of the reference waveform and perform a NAMDF of only the reference waveform to obtain a reference area. The reference area is newly obtained when the similarity exceeds the threshold and the reference waveform changes.

When the area of the reference waveform is obtained, the pitch is advanced by the pitch of the processed waveform, a new frame is taken, NAMDF is performed, the pitch is obtained, and the amplitude information is extracted. After that, the waveform as much as the obtained pitch is cross NAMDF as in the reference waveform and equation (3.1), and the area A (p) is obtained by Equation 3. Similarity D (p) is measured with the obtained area and reference area as shown in Equation 4.

If the similarity is less than the threshold, compress and repeat the above process. If the similarity is greater than the threshold value, the reference area is recalculated using the period as the reference waveform, and the above process is repeated. The synthesizer recovers the PSOLA method using the transmitted waveform, pitch information and amplitude information.

The compression rate can be adjusted by changing the threshold at the transmitter. When the voice is compressed in this way, the results according to the compression ratio are shown in Table 1. As can be seen from Table 1, the compression performance was achieved by 45% of the total speech, and the MOS of about 4.1 was obtained.

Figure 6 Block diagram of the proposed method

Figure 7 Coding for 'A' Voice

(A) Speech waveform (B) Waveform transmitted

(C) Transmitted pitch information (C) Transmitted amplitude (change) information

There is a problem that a DSP chip should be used when applied to a prototype that processes speech using ADPCM, which is a typical method of waveform encoding. This weakens the price competitiveness of the product. Therefore, the present invention proposed a new method that can be synthesized by a general-purpose chip by encoding the waveform in a pitch unit completely different from the conventional waveform compression method.

First, the pitch is searched by NAMDF to obtain a reference waveform, and the similarity is measured for each pitch section. A threshold of similarity is determined to determine whether to compress the waveform. When compressing, only amplitude and pitch information is stored or transmitted. As a result, it was found that MOS 4.1 was maintained even after compressing to about 45% of the total voice.

In the speech encoding method proposed in the present invention, only voiced sound is compressed, but if compression is performed on unvoiced sound and silence, a higher compression rate can be obtained. The feature of the speech encoding method proposed by the present invention is that the algorithm is very simple. Therefore, when the voice data is compressed and transmitted or stored by using the method proposed in the present invention in the field to be commercialized using the voice encoding method, it can be commercialized using a low-cost general purpose chip and thus may have external competitiveness.

Claims (1)

In waveform encoding method of voice signal, pitch is measured by pitch period, and the similarity of formant is measured for each pitch period. The part with high similarity is compressed and the part with low similarity is transmitted. After the search, the waveform is divided by the pitch period from the voiced sound of the voice signal by the searched pitch, and then the reference pitch period is determined first. After that, the CROSS AMDF is applied to the reference pitch period and subsequent pitch periods. The waveform information is regarded as having high formant similarity, and only the pitch period value and the energy value are transmitted, and the rest are not transmitted. If it does not exceed the threshold value, the formant similarity is regarded as small and all the waveform information is transmitted. To form and measure formant similarity in the same way for subsequent pitch periods. Determining whether to transmit one after the decoding operation, the decoding method using the blood Solar (PSOLA) method with only the pitch value and an energy value for the transmitted compressed pitch period based on the pitch period based on transmission.