JP2523286B2 - Speech encoding and decoding method - Google Patents

Description

TECHNICAL FIELD The present invention relates to a voice encoding method for converting a voice signal into a frequency domain and quantizing it while suppressing waveform distortion.

"Prior Art" As a coding method of this type, Japanese Patent No. 1258025 "adaptive conversion coding method of speech" is known. According to this method, as shown in FIG. 3, the audio signal input from the input terminal 11 is generally an analog audio signal sampled at a constant cycle and the sampled value is a digital signal. It is input to the linear prediction analysis unit 12 in units, subjected to linear prediction analysis, and its prediction coefficient and power are vector quantized by the vector quantization unit 13 and output as auxiliary information. The auxiliary information is inversely quantized and decoded by the local decoding unit 14, and the decoded output controls the filter coefficient of the inverse filter 15. The audio signal from the input terminal 11 is input to the inverse filter 15, and the prediction residual in the linear prediction analysis unit 12 is extracted. This prediction residual signal
In 16, the signal is converted into a frequency domain sample by discrete cosine transform, discrete Fourier transform or the like, and then a certain number of low frequency components of the frequency domain signal are taken out from the low frequency component to form a vector. These are quantized by the vector quantizer 18 collating with the codebook in vector units.

It is assumed that the spectrum envelope of the input voice signal is in the state shown in FIG. 4A, for example, four of which are divided into bands F _{1 to} F _4, and vector quantization is performed on the prediction residual signal. At this time, in this example, the amplitude is large in the low frequency band F _{1 and} small in the high frequency bands F ₃ and F ₄ , and the linear prediction synthesis filter is controlled in the frequency domain on the decoder side so as to have such an amplitude. There is a need to. That is, at this time, the amplification factor of the sample point in the frequency domain in the linear prediction synthesis filter is different on the decoder side. That is, in order to minimize the final waveform distortion, it is necessary to control the quantization distortion in vector units.
Actually, each vector (frequency domain) in the vector quantizer 18 from the spectrum envelope value decoded in the local decoder 14
The adaptive information allocation unit 19 calculates and determines the optimum information allocation for each analysis frame. For example, in the case of FIG. 4A, the most bits are allocated to the band F _{1 as} shown in FIG. 4B,
The bands F ₃ and F ₄ are the smallest, and F ₁ is in the middle. This is adaptive information allocation, which is a very effective means for reducing waveform distortion.

The codes quantized by the vector quantizers 13 and 18 are sent to the decoding side. On the decoding side, as shown in FIG. 3, the code from the vector quantization unit 13 is the spectrum envelope reproduction unit 21.
Is decoded to regenerate the spectrum envelope, and the filter coefficient of the linear prediction synthesis filter 22 is controlled by the prediction coefficient indicating this. From the spectrum envelope reproduced together with this, the bit allocation in the adaptive information allocation unit 19 is calculated, whereby the residual spectrum reproduction unit 23 extracts the quantization code of the vector quantization unit 18 for each band, and The code is decoded to obtain the frequency component of the prediction residual. The inverse transform unit 24 transforms this into a prediction residual signal in the time domain and supplies it to the linear prediction synthesis filter 22 as a sound source signal. The filter output is obtained at the output terminal 25 as a synthesized voice signal.

However, when putting this type of digital coding into practical use, an error may occur in the transmission path. This is an unavoidable problem especially in mobile radio applications. However, coding using adaptive information allocation suffers great damage from code errors. In digital transmission, for example, as shown in FIG. 5, each analysis frame is transmitted. Each frame includes a code indicating a spectrum envelope, that is, a code from the quantizing unit 13, auxiliary information 31 including power information as necessary, and a code from the quantizing unit 18, that is, waveform information 32.

On the decoding side, auxiliary information 31
Is decrypted. As a result, the information indicating the spectrum envelope is determined, and the method of allocating the information (bits) to the waveform information 32 that is different for each frame is known. The waveform information 32 is decoded according to the allocation promise.

However, if a code error occurs in the auxiliary information 31, not only the characteristics of the spectrum envelope cannot be transmitted correctly,
An error occurs in the bit allocation of the waveform information 32. If this happens, the information boundary for each vector of the waveform information 32 that occupies most of the information will be erroneous, and most of the information content will be lost in the entire frame.

Therefore, practically, the waveform information 32 is deleted and the auxiliary information 31 is deleted.
However, there was no choice but to add a sufficient error correction code. However, if a transmission path error of a certain level or more occurs, decoding cannot be performed and voice output cannot be obtained.

It is an object of the present invention to reduce waveform distortion by quantizing speech in the frequency domain with adaptive information allocation that is weak against channel code errors, and with fixed information amount allocation. It is to provide a speech encoding method.

"Means for Solving Problems" According to the first invention, a sample value sequence of a voice signal is set to one frame for every fixed number, an orthogonal transform is performed for each frame to obtain a spectrum, and the spectrum is quantized. In the coding method for encoding, the envelope of the spectrum is flattened, the flattened spectrum is rearranged according to a certain rule in the frequency domain, the rearranged spectrum is divided into a plurality of vectors, and each vector is associated with the corresponding spectrum. Vector quantization is performed on a weighted distance scale with an envelope value. The rearrangement rule is determined so that the weights, that is, the sums of the spectral envelope values are almost equal for each vector. The amount of information (the number of bits) for each vector is assigned to be constant. Unlike the conventional method of controlling the waveform distortion by adaptively allocating the information, the waveform distortion is controlled by the weighted matching in the vector with the fixed information allocation.

According to the second aspect of the present invention, the waveform information consisting of a plurality of indexes and the auxiliary information are input, and the code vector consisting of the spectrum value corresponding to each index of the waveform information is read out from the code book and read out within one frame. The spectral values in the multiple code vectors are rearranged in the frequency direction by a predetermined method, the spectral values of the rearranged code vectors are inversely orthogonally transformed to obtain a sound source signal, and the spectrum envelope value is obtained from the auxiliary information. Is reproduced, the filter coefficient is controlled by the reproduced spectrum envelope value, and the sound source signal is input to perform linear prediction synthesis to obtain a reproduced sound signal.

[Embodiment] FIG. 1 shows an embodiment of the second invention, and parts corresponding to those in FIG. 3 are designated by the same reference numerals. In the present invention, the spectrum (transform coefficient) is flattened, the flattened spectrum envelope is divided into a plurality of bands, and each is quantized. Flattening the spectrum is done by passing it through an inverse filter 15 in the time domain. Further, in this embodiment, the rearrangement blocking unit 41 rearranges the samples.
This is to make the average value of the spectrum envelope value corresponding to the sample for each of the plurality of vectors as constant as possible. The method of rearrangement is, for example, that the spectral envelope value (A in FIG. 2) is sequentially numbered 1 to 4 with a constant value as shown in FIG. 2, and the vector # 1 collects each sample of the number 1, that is, Samples having the same number are collected and set as vectors # 1 to # 4. If you do this with a small enough number,
Both the portion with a large spectral envelope and the portion with a small spectral envelope are distributed to each vector, and as a result, the average values of the spectral envelope values of each vector are almost equal.

Next, weighted vector quantization of the spectrum envelope value Wi is performed in the weighted vector quantization unit 42 for each vector created in this way. That is, a large weight is given to a sample having a large level in each vector. Specifically, the input vector in the frequency domain , The codebook vector that minimizes D in the following equation Is searched and the index is transmitted as information (code).

m is the dimension number of the vector, and the weight Wi may be the same for each vector in the same analysis frame as shown in FIG. 2C, but the weight Wi changes for each frame depending on the spectrum envelope value. However, the amount of information (the number of bits) allocated to each vector may be fixed regardless of the spectrum envelope value as shown in FIG. 2D. The weight Wi is the local decoding unit 14
It is determined by using the decoded spectrum envelope.

Also, the vector quantization codebook may be learned with a scale with an average weight Wi or uniform weight,
It may be Gaussian noise itself whose power is normalized, or may be selected from noise according to its weight.

As in FIG. 5, the index from the codebook for each frame is used as waveform information, and this and the quantized prediction coefficient and auxiliary information of power are transmitted. As shown in FIG. Since information, waveform information, and the amount of information (bit number) for each band in the waveform information are fixedly known in advance, these are separated and a vector is reproduced using the same codebook, and the waveform As is clear from FIG. 2, the decoded spectrum signal of information is rearranged in the original order for each frame by the rearranging unit 44, inverse orthogonally transforms it, and supplies it to the linear prediction synthesis filter 22 to output speech. To get

When reproducing the vector at the decoding side, the boundary of information for the vector can be fixed regardless of the spectral envelope. Also, the weight Wi is automatically determined once the index is determined, and it is not necessary to send information about the weight Wi. Therefore, even if the auxiliary information is incorrect, the input signal of the synthesis filter 22, that is, the original residual signal is correctly decoded. In general, the residual signal alone is fairly intelligible, and even if the characteristics of the synthesizing filter 22 are not properly controlled, a voice output of a certain quality can be obtained.

The coding of this embodiment can also be applied to the coding method for performing pitch prediction in the linear prediction analysis and the coding of Japanese Patent Application No. 59-173903. Also, instead of an inverse filter, the input audio signal may be divided by its spectral envelope.

[Advantages of the Invention] As described above, according to the present invention, waveform distortion can be controlled by weighted vector quantization, and the overall waveform distortion can be reduced to the same extent as adaptive information allocation. At the same time, since the fixed information is assigned, there is an advantage that the waveform information can be decoded independently of the auxiliary information. Therefore, even if the auxiliary information cannot be transmitted correctly due to a channel code error,
Waveform information can be decoded, and the quality of output voice is minimally degraded. Therefore, this is an extremely effective coding method for mobile radio applications in which transmission path code errors cannot be avoided.

[Brief description of drawings]

FIG. 1 is a block diagram functionally showing an example of the speech encoding method of the present invention, and FIG. 2 is a spectrum envelope, rearrangement, its vector, quantization weight in the vector, and information distribution of the vector. FIG. 3 is a block diagram functionally showing a conventional adaptive information allocating speech coding method, FIG. 4 is a diagram showing an example of spectrum envelope and vector information distribution, and FIG. FIG. 6 is a diagram showing an example of information arrangement within one frame when the adaptive information is allocated.

Claims (3)

(57) [Claims] 1. A speech coding method in which a sample value sequence of a speech signal is set to one frame for each fixed number, orthogonal transformation is performed for each frame to obtain a spectrum, and the spectrum is quantized. Processing, rearrangement of the flattened spectrum for each frame in a frequency direction in a predetermined manner in a predetermined manner, and a plurality of vectors for the rearranged flattened spectrum for each frame. And the distance to the code vector of the codebook is calculated by weighting each of these multiple vectors with the spectrum envelope value, and the code vector with the smallest distance is selected from the codebook. And a process for outputting the number of the selected code vector and the auxiliary information of the spectral envelope value. Encoding method. 2. The rearrangement of the spectra is performed by sequentially giving a cyclically repeated number to each sample of the spectrum to be rearranged in one frame, and collecting the samples of the same number to form one vector. The speech coding method according to claim 1, wherein the speech coding method is performed as follows. 3. A process of inputting waveform information consisting of a plurality of indexes and auxiliary information and reading a code vector consisting of a spectrum value corresponding to each index of the waveform information from a codebook, and reading within one frame. A process of rearranging the spectrum values in the plurality of code vectors thus obtained in the frequency direction by a predetermined method, a process of inverse orthogonal transforming the spectrum values of the reordered code vectors to obtain a sound source signal, and the above-mentioned auxiliary Speech decoding including a process of reproducing a spectrum envelope value from information, a process of controlling a filter coefficient by the reproduced spectrum envelope value, and a process of inputting the sound source signal and performing linear prediction synthesis to obtain a reproduced voice signal. Method.