KR100309873B1 - A method for encoding by unvoice detection in the CELP Vocoder - Google Patents

Abstract

PURPOSE: An encoding method using detection of voiceless sound in a code-excited linear prediction is provided to reduce the amount of calculation and a transmitting rate by detecting and encoding the part of a voiceless sound in a voice sound. CONSTITUTION: A variation degree of spectrum is tracked by using an LSP. The first frame of a starting part of a voiceless sound part and the last frame of the last part of the voiceless sound part are detected by using interval information and interval dispersion of LSP. The first frame of the starting part of the voiceless sound part is encoded as 1. The last frame of the last part of the voiceless sound part is encoded as 0. In a decoding process, the encoded first frame is decoded by using LSP information of the next frame. The encoded last frame is decoded by using LSP information of a previous frame.

Description

A method for encoding by unvoice detection in the CELP Vocoder}

Since the method of transmitting the voice information of the past has been applied to a limited number of users and a specific field, the transmission rate is not considered very much and only if the voice of high quality can be sent to the receiver. However, due to the development of multimedia and mobile communication compared to the past, the services provided to a specific group or individual are now available to a large number of the public, and the number has also increased exponentially. As a result, the transmission rate, which has been applied until now, cannot satisfy the number of user groups, and the degradation of sound quality that occurs when the transmission rate is reduced and the number of users on the same channel has also been a problem. Under such a background, a voice encoder, that is, a coder / decoder (Vocoder) has been developed.

Speech coding methods for storing or transmitting speech signals can be classified into three types: waveform coding, signal source coding, and hybrid coding. The waveform encoding method removes only the excess components of the waveform itself without separating the components of the speech signal, encodes them, transmits them, and synthesizes them again. The waveform encoding method maintains the high sound quality and the speaker's personality, while the large amount of data is required to maintain the transmission waveform, so that the transmission rate is high and a large memory is required. On the other hand, the signal source encoding method uses a method of separately separating and encoding the excitation component and the filter component of the speech signal based on the speech model of speech, so that the transmission bandwidth is small and the memory usage is small. However, due to the cumulative problem of analysis and synthesis, synthetic sound quality is greatly degraded in nature and clarity. Hybrid coding combines the memory efficiency and low bit rate of signal source coding with the advantages of maintaining the high quality of waveform coding. Formant information uses linear predictive coding. VELP, MELP, CELP methods and the like have been proposed.

Currently, G.723.1 is adopted as the standard vocoder for ITU-T multimedia, and G.723.1 has a dual rate structure of 5.3 / 6.3kbps. In case of 5.3kbps among two transmission rates, ACELP (Algebraic CELP), which is one of CELP (Code Excited Linear Prediction) series speech incubation method, is used.

G.723.1, a CELP type vocoder, has a double transmission rate of 5.3 / 6.3kbps, and is used as an internet phone and other mobile communication vocoder that is currently commercialized, and provides excellent sound quality compared to low data rates. However, in G.723.1, the audio signal is encoded by dividing it into only mute and voice parts, so there is no separate processing in unvoiced sound. In other words, the unvoiced sound is also transmitted through the same analysis as the voiced sound, which requires more computation and loss in terms of transmission rate.

The present invention is to solve the above problems by using a LSP obtained from the analysis stage to detect and encode a portion of the unvoiced voice in the speech to reduce the amount of calculation, and thus to reduce the transmission rate.

1 is a block diagram of a conventional CELP speech coder.

Figure 2 is a graph showing the characteristics of the LSP used for component separation in the present invention, Figure 2a shows the spectrum for unvoiced sound, Figure 2b shows the spectrum for voiced sound, Figure 2c shows the LPC envelope and LSP of unvoiced sound 2D is a diagram showing an LPC envelope and an LSP of a voiced sound.

3 is a diagram showing an LSP extracted from a CELP encoder, and FIG. 3a is a graph showing a speech waveform for 'She', and FIG. 3b is a graph showing an LSP extracted from a CELP.

4 is a diagram showing parameters used for component separation of the present invention. FIG. 4a shows a speech waveform for 'She', FIG. 4b shows a spectral variation diagram, and FIG. 4c shows a number of formants. 4d shows the dispersion of LSP intervals.

5 is a diagram showing unvoiced sound portions detected by the component separation algorithm of the present invention.

6 is a diagram showing an unvoiced sound portion detected by the component separation algorithm of the present invention.

7 shows unvoiced portions detected by the component separation algorithm of the present invention.

In order to solve the above problems, in the method of encoding component separation through LSP parameters in a CELP vocoder, the spectral gradient is tracked with the LSP to find a large change, and the dispersion and spacing information of the LSP is simultaneously Using the first one frame of the part where the voice starts or the last one of the part where the voice ends, ending in the unvoiced sound, and using 1 bit in the encoding for the detected frame. The part detected as the first frame is encoded as 1, the part detected as the last frame ending with unvoiced sound is encoded as 0, and in decoding, the frame encoded as 1 is decoded using the LSP information of the subsequent frame and encoded as 0. The frame is decoded using the LSP information of the previous frame.

Hereinafter, with reference to the accompanying drawings will be described in detail the present invention.

1. Principle of CELP Type Speech Coder

Fig. 1 shows the structure of G.723.1, which is a CELP vocoder which is conventionally used.

The CELP type encoder 100 is configured to select one of the signals obtained by passing two time-varing linear recursive filters of the input excitation signal sequence stored in the codebook to optimize a given fidelity determination.

The CELP encoder 100 analyzes a speech signal obtained as an input, extracts necessary parameters, and uses the analysis by synthesis method to synthesize a speech signal and compare the input speech signal with a very good sound quality. great. However, since the speech must be synthesized and compared each time, it has a very complicated structure.

2. LSP Extraction from G.723.1

In the CELP vocoder G.723.1, a 10th-order linear predictive analysis (10) is used. In each subframe, the 180 sample window is located at the center of each subframe. Hamming window is used at this time.

Linear predictive coefficients (LPC) are calculated using Levinson-Durbin recursion. This LPC coefficient is used to make the short-perceptual weighted filter 20. The LPC synthesis filter is defined as follows.

Where i is the subframe index and is defined between 0 and 3.

The LPC coefficient \ {alpha_j \} _ j = 1..10 is converted to the LSP coefficient \ {p'_j \} _ j = 1..10 by a search during interpolation by unit circle and zero crossing (LSP quantizer ( 30)).

3. Detection of unvoiced sound using LSP in CELP

LSP has constant spectral sensitivity, low spectral distortion in low bit rate coding, and good linear interpolation of parameters.

2 shows LPC and LSP spectrum characteristics of unvoiced and voiced sounds.

2A and 2B show an unvoiced voice and a voiced sound spectrum, respectively, and FIG. 2C shows an LPC envelope and an LSP spectrum of unvoiced sound. It can be seen that the LSP representing the bandwidth of the resonance peak does not appear at the low frequency of FIG. 2C. FIG. 2D shows the spectrum of the LPC envelope and LSP of the voiced sound, where a narrow gap LSP representing three resonance peaks can be seen. In this unvoiced LSP spectrum, narrowly spaced spectra appear at higher frequencies, with two or fewer. That is, the unvoiced sound has a formant located at a high frequency and mainly two or less, and the voiced sound has three or more formants and is distributed from a low frequency to a high frequency.

Figure 3 shows the shape of the LSP obtained through analysis in the CELP for the speech waveform 'She'. It is a 10th order unquantized coefficient per frame, which can have a maximum value of 256 (2pi) and has a normalized frequency resolution of pi / 256. In FIG. 3, up to 11 frames represent unvoiced sound and thereafter voiced sound. Prior to 11 frames, the first formant does not appear and each LPC interval does not show a relatively large change. And even though the formant appears, it can be seen that it is two or less. However, after 11 frames, not only the first formant but also three or more formants appear at very narrow intervals. With this feature, it can be estimated that the unvoiced sound and the voiced sound can be distinguished by the information of the gap between the number of formants and the 10 LSP coefficients.

In the present invention, the unvoiced start frame or the unframed end frame of the unvoiced sound is detected from the silent to unvoiced section and applied to the rate reduction. The criterion uses the resonance peaks described above, that is, the variance of the formant number and the spacing of the LSP coefficients, and in addition, the spectral gradient is measured first.

P_i ^ T is the LSP coefficient of the i-th frame and P_j ^ T is the LSP coefficient of the j-th frame that is the i-th previous frame. Take the coefficients of two frames and measure the change in the spectrum as follows:

Distance (i) has a large value at the beginning of the unvoiced sound in the silence and the section transitioning from the unvoiced sound to the silence. Of course, it also has a large value in other formant changes.

FIG. 4A illustrates a speech waveform for 'She', and FIG. 4B is a spectrum change diagram obtained using Equation 3. FIG. 4b, it can be seen from FIG. 4 that a large change occurs at the beginning of the unvoiced sound in silence. In the present invention, since the section to be detected is an unvoiced section such as the four frames, a large value appearing in the remaining frames does not have an important meaning. FIG. 4C is a graph showing the number of formants, and 4 frames have two or less formants. 4D is a graph showing the dispersion of the LSP intervals. Since unvoiced sounds appear less than voiced sounds, a threshold value is applied to them so that a section equal to 4 frames is detected.

4. Experiment and Results

The equipment used for the experiment was IBM-PC / Pentium (150) system and interfaced with 16-bit AD / DA converter for commercial input and output of voice signal and inputted data at 11kHz and 8kHz sampling rate. The length of one frame for each sample was 240 samples, and the subframes were processed in units of 60 samples. In order to measure the performance of the treatment result, the following representative sentences were used as samples by five speakers of various age groups.

Speaking 1:: Insu's little boy likes a genius boy.

Voice 2: / Jesus told the lesson of creation.

Voice 3: / Soongsil University, Information and Communication, Voice Communication Research Team./

Voice 4: / The human challenge of flying through the sky is endless./

Voice 5: Weather forecast

In the 5.3kbps speech coder of G.723.1, the unvoiced detection algorithm is implemented in C-language. In order to compare the data rates, two processes were performed. First, the LSP obtained after the LPC analysis of FIG. 1 is used to determine whether it is an unvoiced sound section to be detected. In this process, a frame having a spectral gradient of 20 or more, a number of formants of 2 or less, and a dispersion of LSP intervals of 20 or less were detected. By using the same algorithm, we confirmed whether the frame detected by the decision logic is the beginning of the unvoiced sound in the section that changes from silent to unvoiced or the end of the unvoiced sound in the section that changes from unvoiced to silent.

5 to 7 show unvoiced frames detected by the proposed decision logic. 5 and 6 show results detected at a portion starting with an unvoiced sound, and FIG. 7 shows results detected at a portion ending with an unvoiced sound.

The detected frame skips both the analysis and the encoding process and encodes a new frame type by adding 1 bit, which is different from the existing silent and voice activity. That is, one frame may be used to encode an interval of a start frame of an unvoiced sound (eg, a transition from silent to unvoiced) as 1, and an interval of an unvoiced end frame (such as a transition from unvoiced to silent) to 0. . The decoder reads the one bit and decodes the data using the previous frame.

Voice samples 1) and 2) produced both voice 1) through 5) and voice samples 3) were the voices of the weather forecasters. The average error rate is 0.18%.

The analysis stage of the CELP speech coder, which is a low transmission vocoder, performs linear prediction on the spectrum as an LPC, transforms it into an LSP, and transmits it. This makes it suitable for maintaining sound quality since it is resistant to errors and easy to linear interpolation.

The present invention proposed a method of reducing the transmission rate by detecting a portion of the unvoiced sound in the speech using the LSP obtained from the analysis stage. First, the unvoiced sound section is found to reduce the transmission rate while minimizing degradation of sound quality by using characteristics indicating the number of formants of LSP, reflectance characteristics of spectrum and dispersion information of LSP interval. The bit rate can be reduced by using simple coding and decoding techniques. As a result of applying the method of the present invention to 5.3kbps of the CELP vocoder G.723.1, the average separation error rate was 0.18% and the transmission rate was reduced by 10.9%, and the sound quality showed subtle degradation of about 0.034.

Claims (1)

A method of encoding by performing component separation through the LSP parameter in a CELP vocoder, wherein the spectral gradient is tracked with the LSP to find a large change, and the part of the voice starting using the interval information of the LSP and the variance of the interval simultaneously Detecting the first one frame of the part starting with the unvoiced sound or the last one ending with the unvoiced sound from the part where the voice ends, and the part detected as the first frame starting with the unvoiced sound by using 1 bit for encoding the detected frame Is 1, the part detected as the last frame ending in unvoiced sound is encoded by 0, and when decoding, the frame encoded by 1 is decoded using the LSP information of a subsequent frame, and the frame encoded by 0 is a preceding frame. Encoding using the LSP information of the apparatus.