JPH0786952A - Predictive encoding method for voice - Google Patents

Abstract

PURPOSE:To provide a high-quality decoded voice at the destination of transmission even when voices provided with different frequency characteristics are encoded and transmitted. CONSTITUTION:A predictive coefficient decision part 2 and a predictive coefficient quantization part 4 set a predictive coefficient to a synthetic filter 3. A pitch cycle vector and a noise waveform vector are outputted from an adaptive code book 5 and a noise code book 7, and gains are respectively multiplied at gain parts 6 and 8. The outputs of the gain parts 6 and 8 are added and supplied to the synthetic filter 3 later, and a resultant speech vector is synthesized. Concerning distortion provided by subtracting the resultant speech vector from an input speech vector to which power is quantized, the sense of hearing is weighted so that the degree of weighting can be adaptively controlled based on the frequency characteristic of the input speech vector, power is calculated later, the pitch cycle vector and the noise waveform vector are selected from the adaptive code book 5 and the noise code book 7 so as to minimize this power, and the gains of the gain parts 6 and 8 are set.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice predictive coding method which is used for digital mobile communication such as a car telephone and which codes a voice with high efficiency.

[0002]

2. Description of the Related Art In recent years, in the technical field of digital mobile communication and the like, various high efficiency coding methods have been used for the purpose of effectively utilizing radio waves. Among these high-efficiency coding methods, as a high-efficiency coding method for coding speech at a coding rate of about 8 kbit / s, there are a code-driven linear prediction (CELP) coding method and a vector addition-driven linear prediction method. (VSELP) encoding method, multi-pass encoding method, and the like.

The details of the CELP coding method are described in, for example, "Co Co.," written by MR Schroeder and BSAtal.
de-Excited Linear Prediction (CELP): High-quality
Speech at Very Low Rates "(Proc. ICASSP '85, 25.1.
1, pp. 937-940, 1985, for details of the VSELP coding method, see, for example, "Vector Sum Excited Linear Prediction (VSELP)" by IAGerson and MA Jasiuk.
Speech Coding at 8kps "(Proc. ICASSP '90, S9.3, p
p. 461-464, 1990) for details of the multi-pass encoding method, see, for example, Kazunori Ozawa and Taku Araseki, “9.6-4.8 kbit / s multi-pass speech using pitch information”. Encoding system "(The journal of journals (D-II), J72-D-
II, 8, pp. 1125-1132, 1989) respectively.

FIG. 1 is a block diagram showing a configuration example of a speech coding apparatus using a conventional CELP coding method.
Input audio data generated by sampling an analog audio signal at a sampling frequency of 8 kHz is input from the input terminal 1. In the prediction coefficient determination unit 2, a plurality of samples of the input voice data input from the input terminal 1 are combined into one vector as one frame (hereinafter,
Input speech vector), a linear prediction analysis is performed on this input speech vector, and the transfer function {1 / A (z)}
The prediction coefficient (linear predictive coding (LPC) coefficient or line spectrum pair (LSP) coefficient) of the synthesis filter 3 having is calculated and determined. As a result, the prediction coefficient quantization unit 4 quantizes the prediction coefficient and sets it in the synthesis filter 3.

The adaptive codebook 5 is configured to store a plurality of pitch period vectors corresponding to the pitch period of a voiced section of speech. From the adaptive code book 5, the pitch period vector selected and extracted by the distortion power calculation unit 12 described later is multiplied by the gain set by the distortion power calculation unit 12 in the gain unit 6, and the gain period 6 Is output.

On the other hand, the noise codebook 7 stores in advance a plurality of noise waveform vectors (for example, random number vectors) corresponding to unvoiced sections of speech. The noise waveform vector selected and extracted from the noise codebook 7 by the distortion power calculation unit 12 described later is multiplied by the gain set by the distortion power calculation unit 12 in the gain unit 8 and output from the gain unit 8. To be done. The output vector of the gain unit 6 and the output vector of the gain unit 8 are added by the adder 9
Are added together, the output vector of the adder 9 is supplied to the synthesizing filter 3 as a drive vector, and the synthesizing filter 3
In, a voice vector (hereinafter referred to as a synthesized voice vector) is synthesized based on the set prediction coefficient.

In addition, after the power of the input speech vector is calculated in the power quantizing unit 10, the power is quantized, whereby the quantized power of the input speech vector is used as the input speech vector. The pitch period vector and are normalized. Then, the subtractor 11 subtracts the synthesized speech vector from the input speech vector that is normalized and output from the power quantization unit 10 to obtain distortion data.

Next, the distortion power calculator 12 calculates the power of the distortion data, and the pitch period vector and the noise waveform are respectively calculated from the adaptive codebook 5 and the noise codebook 7 so that the power of the distortion data becomes the smallest. The respective vectors are selected and the gains of the gain units 6 and 8 are set. Thereby, in the code output unit 13,
Information (code) and gain selected for each of the prediction coefficient, the power of the input speech vector, the pitch period vector and the noise waveform vector are converted into a code of a bit sequence and output, and these codes are transmitted. .

By the way, in the distortion power calculation unit 12,
When evaluating distortion data, which is the difference between a synthetic speech vector and an input speech vector, if the distortion data is evaluated to be the minimum, that is, the SN is maximized, the quantization noise is uniform on the frequency axis. Will be distributed in. Further, the audio signal has much power in the low frequency range, but the power decreases as the frequency increases. Therefore, if the quantization noise is uniformly distributed on the frequency axis, the quantization noise level is relatively higher than the speech level in the high frequency range, which causes deterioration of the coded speech.

Therefore, in the prior art, as shown in FIG.
After the distortion data is weighted based on the spectrum of the input speech vector using, the power calculation unit 15 evaluates the distortion data. That is, in the low frequency range where the voice power is high, the quantization noise is inaudible because it is masked by the voice even if the quantization noise level becomes a little higher than in the case of the uniform distribution. On the contrary, in the high range, weighting is performed so that the distribution becomes lower than the uniform distribution. In FIG. 2, e is distortion data output from the subtractor 11, and e ′ is weighted distortion data.

The transfer function W of the perceptual weighting filter 14
(Z) is represented by the equation (1).

[Equation 1] here,

[Equation 2]

[Equation 3] In the expressions (1) to (3), the coefficient α _i is the non-quantized LPC coefficient obtained by the prediction coefficient determination unit 2. In addition, for coefficients γ ₁ and γ ₂ , 0 <γ ₂ <γ ₁ <1
The value of is used. Further, since influenced the characteristics of perceptual weighting filter 14 by a factor gamma ₁ and gamma _2, the values of these coefficients gamma ₁ and gamma ₂ are determined empirically by listening.

[0012]

By the way, the voice of a telephone has conventionally been transmitted to the International Telegraph and Telephone Consultative Committee (CCIT).
Voices with IRS characteristics standardized in T) have been used. However, recently, with the spread of small electret microphones, voices having frequency characteristics different from the IRS characteristics (hereinafter referred to as NON-IRS characteristics) have also been used. When encoding voices having such different frequency characteristics, the auditory weighting filter 14 described above is used.
The optimal values of the coefficients γ ₁ and γ ₂ of are naturally different.

However, in the above-described conventional predictive coding method for speech, as described above, the values of the coefficients γ ₁ and γ ₂ of the perceptual weighting filter 14 are irrespective of the frequency characteristics of the input speech. It was fixed to a certain value empirically by listening to it. Therefore, when speech having a frequency characteristic that does not match the values of the coefficients γ ₁ and γ ₂ of the auditory weighting filter 14 is encoded and transmitted, there is a problem that a decoded speech of good quality cannot be obtained at the transmission destination. there were. The present invention has been made under such a background, and a predictive coding method for a voice capable of obtaining a decoded voice of good quality at a transmission destination even when voices having different frequency characteristics are coded and transmitted. The purpose is to provide.

[0014]

The invention according to claim 1 is
Input speech is subjected to linear prediction analysis to calculate prediction coefficients, the prediction coefficients are set in a synthesis filter, and the synthesis filter stores an adaptive codebook in which a plurality of pitch period vectors are stored and a plurality of noise waveform vectors. Speech coded by utilizing the pitch period vector and the noise waveform vector respectively selected from the selected noise codebook and driving in frame units composed of a plurality of samples of the input speech to synthesize synthesized speech. In the predictive coding method, the distortion is selected in order to select the pitch period vector and the noise waveform vector from the adaptive codebook and the noise codebook so that the distortion between the synthesized speech and the input speech is minimized. When auditory weighting is performed, the degree of weighting is adaptively controlled based on the frequency characteristics of the input voice. It is set to. The invention according to claim 2 is characterized in that, in the invention according to claim 1, the degree of weighting is adaptively controlled by using a Percoll coefficient.

[0015]

According to the present invention, since the degree of perceptual weighting applied to the distortion between the synthesized voice and the input voice is adaptively controlled based on the frequency characteristic of the input voice, voices having different frequency characteristics Even when encoded and transmitted, the decoded voice of good quality can be obtained at the transmission destination.

[0016]

Embodiments of the present invention will be described below with reference to the drawings. In the present invention, the configuration of the speech coding apparatus is almost the same as that in FIGS. 1 and ₂ , but the value of the coefficient γ ₂ of the auditory weighting filter 14 that constitutes the distortion power calculation unit 12 is as shown below. It is adaptively controlled by each of the first to third embodiments.

(1) First Embodiment (Intra-frame Processing) As a result of the audition when determining the values of the coefficients γ ₁ and γ ₂ of the perceptual weighting filter 14, the coefficient γ ₁ and the coefficient γ ₁ and The values of γ ₂ are γ ₁ = 0.9 and γ ₂ =
0.6 is preferred. On the other hand, γ ₁ = 0.9 and γ ₂ = 0.4 are preferable for voices having NON-IRS characteristics. Therefore, the coefficient γ ₂ is changed according to the input voice so as to be optimal for both the IRS characteristic and the NON-IRS characteristic.
You can control the value of. That is, γ ₂ = 0.6 is set when the input voice has the IRS characteristic, and γ ₂ = 0.4 is set when the input voice has the NON-IRS characteristic.

Further, voice and NO having IRS characteristics
As a result of analyzing voices having N-IRS characteristics, it was found that the two have a large difference in the distribution of the appearance probability of the first-order coefficient k ₁ of the PARCOR coefficient (Parcor coefficient). That is, this PARCOR coefficient k ₁ must be -1.
PARKOR coefficient k which exists in the range of <k ₁ <1 but is obtained by analyzing voice having NON-IRS characteristics
₁ tends to concentrate near the value +1 and, on the other hand, I
PARCOR obtained by analyzing voice having RS characteristics
The coefficient k ₁ has no such tendency.

Here, in FIG. 3, a voice having an IRS characteristic is analyzed in the vicinity of a value +1 when statistics of the value of the PARCOR coefficient k ₁ obtained by processing the actual voice data by distribution of the appearance probability are analyzed. Distribution of the appearance probability of the PARCOR coefficient k ₁ (curve a) obtained, and the PARCOR coefficient k ₁ obtained by analyzing the voice having the NON-IRS characteristic
The distribution (curve b) of the appearance probability of is shown. As can be seen from FIG. 3, the voice having the NON-IRS characteristic is PARCO.
Although there are many R coefficients k ₁ > 0.9, the number of PARCOR coefficients k ₁ > 0.9 is reduced in speech having IRS characteristics.

Therefore, the value of the coefficient γ ₂ of the perceptual weighting filter 14 is adaptively controlled by the input voice using the feature of the difference of the PARCOR coefficient k ₁ . That is, the PARCOR coefficient k ₁ obtained by analyzing the input voice
Is greater than or equal to the threshold value Th (for example, tH = 0.9), the value of the coefficient γ ₂ of the auditory weighting filter 14 is set to 0.4, and the value of the coefficient k ₁ is set to the threshold value. When it is smaller than Th, the value of the coefficient γ ₂ is set to 0.6. The PARCOR coefficient k ₁ can be obtained when the prediction coefficient determination unit 2 performs the linear prediction analysis. Also,
As already described in the prior art, the audio coding is performed on a frame-by-frame basis. Therefore, in this embodiment, the adaptive control of the coefficient γ ₂ of the auditory weighting filter 14 is also performed on a frame-by-frame basis.

(2) Second Embodiment (Interframe Processing) In the prediction coefficient determination unit 2 shown in FIG. 1, the input speech data is subjected to linear prediction analysis in frame units, and the prediction coefficient of the synthesis filter 3 is calculated. However, in the consonant part or the unvoiced section of the input voice data, the linear prediction analysis for each frame is not always effective. Therefore, NON
Even in a voice having the -IRS characteristic, the value of the PARCOR coefficient k ₁ does not always concentrate around the value +1 in frame units. Further, if the value of the coefficient γ ₂ of the perceptual weighting filter 14 greatly changes in units of frames, the continuity of decoded speech is lost, which is not desirable.

Therefore, in this embodiment, the coefficient γ ₂
As shown in the equation (4), the coefficient γ ₂ of the input voice data of the frame (current frame) currently being processed and the input voice data of each of the M frames processed in the past are expressed as It is expressed by the sum of the coefficients γ ₂ .

[Equation 4] In the equation (4), n is the frame number (the frame number of the current frame is n), and γ ₂ (n) is the PARCOR obtained by analyzing the input voice data of the frame of the frame number n.
The coefficient γ ₂ , M determined by the coefficient k ₁ is an order (for example,
M = 3), W _i is a weighting coefficient. Thus, the coefficient γ ₂
By expressing the sum in the form of a sum, a rapid change of the coefficient γ _{2 in} a frame unit can be avoided, so that the continuity of the decoded speech is not lost.

(3) Third Embodiment By the way, in an actual use situation, whether the input voice data input to the voice encoding device has the IRS characteristic,
Alternatively, it is unclear whether it has NON-IRS characteristics. However, for example, since the frequency characteristic of the voice is determined by the telephone used, if the model of the telephone is determined, it is determined whether the input voice data has the IRS characteristic or the NON-IRS characteristic. It is unlikely that the two characteristics will be reversed.

Therefore, in this embodiment, the value of the coefficient γ ₂ of the auditory weighting filter 14 is set to (5) based on the distribution of the occurrence probability of the PARCOR coefficient k ₁ shown in FIG.
It is adaptively controlled as shown in the equation.

[Equation 5] In _{(5), γ 2min = 0.4, γ 2max} = 0.6,
α ₁ = 0.05, α ₂ = 0.01, β = 0.001, a _{1
= 0.97, a 2 = 0.95,} a 3 = 0.90, the initial value γ
₂ (0) = 0.6. Further, in the equation (5), n
Is the frame number.

When the input voice data has the IRS characteristic, the coefficient γ ₂ has a zero probability that a value of a ₁ = 0.97 or more appears, and a value of a ₂ = 0.95 or more appears. The probability is also very low. Therefore, the coefficient γ ₂ (n) hardly changes. That is, when the input voice data has the IRS characteristic, the coefficient γ ₂ (n) is maintained at 0.6.
On the other hand, when the input voice data has the NON-IRS characteristic, the coefficient γ ₂ has a very high probability that a value of a ₂ = 0.95 or more appears, and if α ₁ = 0.05, 4 frames. Later, γ ₂ (n) = 0.4. Even if the value of the coefficient γ ₂ is 0.9 or less depending on the voice section, since the increase is as small as β = 0.001, it does not change rapidly. Further, if there is a frame in which the value of the coefficient γ ₂ exceeds 0.05, the coefficient γ ₂ (n) converges to 0.4. that's all,
Although the embodiments of the present invention have been described in detail with reference to the drawings, the specific configuration is not limited to these embodiments, and even if there are design changes and the like within the scope not departing from the gist of the present invention. Included in the present invention.

[0026]

As described above, according to the present invention,
Since the coefficient of the perceptual weighting filter can be adaptively controlled according to the frequency characteristics of the input speech, even if the speech having different frequency characteristics is encoded and transmitted, a decoded speech of good quality can be obtained at the transmission destination. There is an effect that can be.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of a speech coding apparatus to which a speech predictive coding method according to first to third embodiments of the present invention and a conventional example is applied.

FIG. 2 is a block diagram showing an example of a configuration of a distortion power calculation unit 12.

[Fig. 3] PARCO due to difference in frequency characteristics of input voice
Is a diagram showing an example of the distribution of the probability of occurrence of R factor k _1.

[Explanation of symbols]

 1 Input Terminal 2 Prediction Coefficient Determining Section 3 Synthesis Filter 4 Prediction Coefficient Quantization Section 5 Adaptive Codebook 6,8 Gain Section 7 Noise Codebook 9 Adder 10 Power Quantization Section 11 Subtractor 12 Distortion Power Calculation Section 13 Code Output Section 14 Auditory Weighting Filter 15 Power Calculation Unit

Claims (2)

[Claims] 1. A predictive coefficient is calculated by performing linear predictive analysis on input speech, the predictive coefficient is set in a synthesis filter, and the synthesis filter is provided with an adaptive codebook in which a plurality of pitch period vectors are stored. A speech is generated by synthesizing a synthesized speech by driving in a frame unit composed of a plurality of samples of the input speech with a pitch period vector and a noise waveform vector respectively selected from a noise codebook in which a noise waveform vector is stored. In the predictive coding method for speech, the pitch period vector and the noise waveform vector are selected from the adaptive codebook and the noise codebook so that the distortion between the synthesized speech and the input speech is minimized. Therefore, when the distortion is auditorily weighted, the degree of the weighting is adaptively controlled based on the frequency characteristic of the input voice. Predictive coding method of speech, characterized by. 2. The degree of weighting is adaptively controlled by using a Percoll coefficient.
Predictive coding method of the described speech.