JPH08202399A - Post-processing method for decoded speech - Google Patents

Abstract

(57) [Abstract] [Purpose] To improve the quality of decoded speech in a speech coding method based on linear prediction analysis. In a speech coding method based on linear prediction analysis, the coefficient of a high-frequency emphasis filter that emphasizes the brightness of speech in a post filter used to improve the quality of decoded speech is adaptively controlled.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to improving the quality of decoded speech in a speech coding method based on linear prediction analysis.

[0002]

2. Description of the Related Art CEL based on linear prediction analysis is currently used as a typical low-to-low bit rate speech coding method.
A method called P (Code Excited Linear Prediction) is widely used. However, even with this method, sound quality deterioration due to the low bit rate is inevitable. Therefore, in order to improve the quality of decoded speech separately from the speech coding method, pitch enhancement, formant enhancement, high-frequency enhancement processing, etc. are often performed on the decoded speech.

Japanese method of digital mobile communication (PD
In C), VSELP (Vector Sum Excited Linea)
An improved speech coding method of the CELP method based on linear prediction analysis called r Prediction) is adopted as a standard. FIG. 5 shows the configuration of a speech decoder in this system. In this figure, 30 is a long-term filter state, 31 is a codebook,
32 is an adaptive pitch pre-filter, 33 is a synthesis filter,
34 is an adaptive spectrum filter, 35 is an adder, 36 is a multiplier, 37 is a long-term prediction coefficient, and 38 is a quantization gain.

In this system, the excitation signal is generated by multiplying the 40-sample vector stored in the codebook 31 and the output of the long-term filter 30 for producing a harmonic structure according to the fundamental period of the sound by each gain and adding the gain. It After the excitation signal is processed by the adaptive pitch pre-filter 32 for emphasizing the harmonic structure of the voice, the synthesis filter 33 adds a spectrum envelope to synthesize the voice. Further, the adaptive spectrum filter 34 enhances the voice formant and the intelligibility (brightness).

FIG. 6 shows the internal structure of the adaptive spectrum filter. Reference numeral 41 is a formant emphasis filter, 42 is a high-pass filter for emphasizing high frequencies, and 43 is a power gain correction unit for adjusting input / output power. The high-pass filter 42 that performs the high-frequency emphasis processing is a filter with a fixed coefficient.

It is known that the vowel part of a voice signal generally has a harmonic structure of a fundamental period. Therefore, CE
In the LP system, a long-term predictor is used so as to reproduce this harmonic structure. However, since the fundamental period slightly changes, the low-frequency harmonic structure is reproduced with relatively high accuracy, but the high-frequency harmonic structure is not reproduced with high accuracy. For this reason, even in a portion where the high-frequency harmonic structure is not accurately reproduced, the high-frequency band is emphasized by the high-pass filter 43 having a fixed coefficient of the adaptive spectrum filter 34, and it is heard as annoying noise and distortion.

[0007]

SUMMARY OF THE INVENTION It is an object of the present invention to improve the quality of decoded speech in a speech coding method based on linear prediction analysis without causing annoying noise in a high frequency range. .

[0008]

In order to achieve the above object, the present invention adaptively adjusts the coefficient of a high-pass filter in a high-pass emphasis process used to improve the clarity (brightness) of decoded speech. It is configured to control.

[0009]

According to the present invention, the high-frequency harmonic structure is not accurately reproduced in the voice encoding process, and the high-frequency range is higher than that of the decoded voice including the annoying noise generated in the high-frequency part of the voice. By controlling the coefficient of the high-frequency emphasis filter so as to reduce the degree of emphasis, it is possible to obtain decoded speech with clear sound quality as a whole without emphasizing the noise portion.

[0010]

Embodiments of the present invention will now be described with reference to the drawings. FIG. 1 shows an adaptive post filter which is a first embodiment according to the present invention. In FIG. 1, reference numeral 1 is a formant enhancement filter, 2 is a coefficient-variable high-pass filter, 3 is a power gain correction unit, and 4 is a filter coefficient control unit. Generally, in an analysis / synthesis speech coding method using linear prediction such as CELP, noise due to quantization is masked in a large portion of speech power spectrum. Therefore, it is considered that the emphasized wave structure in the high frequency region is more difficult to be reproduced because the quantization noise is masked in the high frequency region in the high frequency region where the power is strong. Therefore, the first embodiment of the present invention controls the coefficient of the high-pass filter for high-frequency emphasis by the high-frequency power of the decoded speech.

In FIG. 1, the filter coefficient control unit calculates the ratio of the high frequency power to the power of the input voice, and controls the filter coefficient according to the calculated value. FIG. 2 shows the configuration of the filter coefficient control unit. In FIG. 2, 10 is a high-pass filter, 11 and 12 are power calculation units, 13 is a power ratio calculation unit, and 14 is a filter coefficient determination unit. Further, FIG. 3 shows an embodiment of the high pass filter 2. 15 in FIG.
Is a 1-sample delay device, 16 is a multiplier, 17 is an adder,
18 is a coefficient of this filter.

In the first embodiment, the coefficient of the high pass filter is controlled in VSELP subframe units. This is a 5 msec (4
This is because the residual signal of the audio signal is vector-quantized for each section of (0 sample). The formant of the decoded speech input to the adaptive post filter is emphasized by the formant emphasizing filter 1. In parallel with this, the decoded speech is input to the filter coefficient control unit 4. In the filter coefficient control unit 4, the high-frequency component of the input signal is the high-pass filter 10
Taken out by. Then, the power calculation units 11 and 12
The output of the high-pass filter and the power of the input are calculated respectively.

Next, the power ratio calculating unit 13 calculates the power ratio, and the filter coefficient calculating unit 14 determines the coefficient of the high pass filter 2 to be used for the subframe based on the power ratio. Filter coefficient determination unit 1 in the first embodiment
No. 4 operates as follows in order to suppress distortion due to a rapid change in coefficient between subframes.

If εi> TH1, αi = αi-1 + Δα (Condition 1) If εi ≦ TH2, αi = αi-1−Δα (Condition 2) However, if αi ≦ C1, αi = C1 (Condition 3) αi ≧ C2 Then αi = C2 (Condition 4) where αi is the coefficient of the high-pass filter 2 of the i subframe, Δα is the step size of coefficient update, TH1, TH
2 is a threshold value for condition determination, and C1 and C2 are values that determine the lower and upper limits of the filter coefficient. In this case, the lower and upper limits of 0 to -0.4 are sufficient. Also, P
ti is the output of the power calculation unit 12, and Phi is the power calculation unit 11
, Ε i is the output of the power ratio calculation unit 13 and is defined as in the following equation (1). ε i = Pti / Phi (1) The high pass filter 2 filters the output of the formant enhancement filter 1 using the coefficient determined by the filter coefficient control unit 4. Then, in the power gain adjusting section,
Smoothly adjust the change of input / output power of the adaptive post filter to obtain the adaptive post filter output.

Next, a second embodiment of the present invention will be described. The second embodiment focuses on the harmonic structure of the voice, and emphasizes the high band of the decoded voice when the harmonic structure in the middle range to the high range is reproduced with relatively high accuracy. FIG. 4 shows the configuration of the filter coefficient control unit in the second embodiment. In FIG. 4, 20 is an amplitude spectrum calculation unit, 21 is a logarithmic transformation unit, 22 is a low-middle band inverse Fourier transform unit, 23 is a middle-high band inverse Fourier transform unit, 24 is a peak comparison unit, and 25 is a filter coefficient determination unit.

The filter coefficient control unit in the second embodiment obtains the amplitude spectrum of the input signal from the same input signal as in the first embodiment by using the Fourier transform in the amplitude spectrum calculation unit 20. The amplitude spectrum is logarithmically converted by the logarithmic conversion unit 21, and the low-middle and middle-high frequency spectra are obtained by the inverse Fourier transform units 22 and 23. This operation is generally an operation for obtaining a cepstrum, and in the second embodiment, the harmonic structure in the low-middle band and the middle-high band of the decoded speech is detected in the cepstrum region and is controlled by the detection result. .

The detection of the fundamental period of the harmonic structure of voice is realized by obtaining the peak of the pitch period of the cepstrum in the range in which it may exist. For this reason, when the harmonic structure is accurately reproduced up to the high range, even if the cepstrum is separately obtained from the low-middle range and the mid-high range, the peak positions thereof are almost the same. Therefore, in the second embodiment, the inverse Fourier transform units 22, 2
The peak comparator detects the respective peaks of the low-middle-range cepstrum and the middle-high range cepstrum within the range in which the pitch interval of the voice may exist, and compares the peak positions and sizes. As a result, when a peak of a certain size is detected at substantially the same position, the filter coefficient determination unit 25 controls the coefficient of the high pass filter 2 in a direction to emphasize the high band. On the contrary, when the peak position is deviated or the size of the peak is small, the filter coefficient determination unit 25 controls the coefficient of the high pass filter 2 in a direction not to emphasize the high band.

[0018]

As described above, according to the present invention, the coefficient of the high-pass filter for enhancing the intelligibility (brightness) of the decoded speech is set in the speech coding method using the linear prediction analysis. By adaptively controlling, the intelligibility (brightness) can be controlled without emphasizing annoying noise and distortion in the high frequency range, and high-quality decoded speech can be obtained.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of an adaptive post filter according to the present invention.

FIG. 2 is a block diagram showing a coefficient control unit of a high pass filter according to the first embodiment of the present invention.

FIG. 3 is a diagram showing an embodiment of a high-pass filter.

FIG. 4 is a block diagram showing a coefficient control unit of a high pass filter according to a second embodiment of the present invention.

FIG. 5 is a block diagram showing a decoder of a speech coding method using linear prediction.

FIG. 6 is a block diagram of a conventional adaptive post filter.

[Explanation of symbols]

 1 Formant enhancement filter 2 High pass filter with variable coefficient 3 Power gain correction unit 4 Filter coefficient control unit 10 High pass filter 11, 12 Power calculation unit 13 Power ratio calculation unit 14 Filter coefficient determination unit 15 1 Sample delay device 16 Multiplier 17 Adder 18 Filter coefficient 20 Amplitude spectrum calculation unit 21 Logarithmic transformation unit 22 Low-mid inverse Fourier transform unit 23 Mid-high inverse Fourier transform unit 24 Peak comparison unit 25 Filter coefficient determination unit 30 Long-term filter state 31 Codebook 32 Adaptive pitch pre-filter 33 Synthesis filter 34 Adaptive spectrum filter 35 Adder 36 Multiplier 37 Long-term prediction coefficient 38 Quantization gain 41 Formant enhancement filter 42 High-pass filter 43 Power gain correction unit

─────────────────────────────────────────────────── ─── Continuation of the front page (51) Int.Cl. ⁶ Identification code Office reference number FI technical display location H03M 7/30 Z 9382-5K

Claims (3)

[Claims] 1. A speech coding method based on linear prediction analysis, wherein high-frequency emphasis filter processing in which coefficients are adaptively controlled is applied to speech decoded by the decoder. Post-treatment method. 2. The post-processing method for decoded speech according to claim 1, wherein the control of the coefficients of the high-frequency emphasis filter is controlled based on the high-frequency power of the speech signal. 3. The post-processing method for decoded speech according to claim 1, wherein the control of the coefficients of the high-frequency emphasis filter is controlled based on the harmonic structure of the decoded speech signal.