JP2002244700A - Device and method for sound encoding and storage element - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a sound encoding device which excellently encodes a sound regardless of whether there is background noise. SOLUTION: According to the discrimination result of a subframe discrimination means 107, a sound source code book search means 105 discriminatingly uses an A-b-s sound source code book search means 202 or a spectrum area sound source code book search means 203 to make a sound source code book search in a frequency area when the sound has much background noise.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] The present invention relates to a speech coding apparatus,
The present invention relates to a speech encoding method and a storage element for storing an algorithm of the speech encoding method, and particularly to a speech encoding device, a speech encoding method, and a storage element that are not affected by the presence or absence of background noise.

[0002]

2. Description of the Related Art In recent years, mobile phones, Internet phones,
2. Description of the Related Art As applications using voice coding technology for coding voice signals at a low bit rate, such as DSVD modems, have become widespread, demands for higher quality voice coding technology have increased. The mainstream of speech coding systems at 6 kbps to 16 kbps is CELP (Code Excited Linear Prediction Coding: Code Excited Linear P
A system using the concept of CELP has been used in the above-described various applications and various standards such as digital mobile phones.

[0003] CELP is a sound source of speech by linear prediction.
Using the vocal tract separation model, the vocal tract information is quantized, and the sound source information is further vector-quantized to realize low bit rate encoding. Specifically, first,
The input audio is divided into frames, which are processing units. Further, the frame is analyzed, and a formant parameter expressing a formant component of the voice corresponding to the vocal tract information is extracted and quantized.

[0004] The quantization of the excitation information is often performed on a sub-frame shorter than the frame, and is often performed through a two-step process using two types of codebooks. The first stage is an adaptive codebook search using an adaptive codebook representing pitch information of speech. The adaptive codebook generates a predetermined number of pitches (adaptive code vectors) in which the excitation information up to the immediately preceding subframe is repeated by the pitch length and has a subframe length by a predetermined number of pitches. It is composed of further,
Each adaptive code vector is passed through a vocal tract filter composed of the quantized formant parameters to generate a composite vector. Next, a distance between the combined vector and the input subframe is measured, and an optimal adaptive code vector, that is, an optimal pitch is determined to give the shortest distance. The technique of analyzing using a synthesized sound in this way is called an “analysis by synthesis (Abs) method”.

The second stage is an excitation codebook using a fixed codebook composed of a predetermined number of fixed vectors (excitation code vectors) called an excitation codebook (sometimes called a noise codebook). Search (sometimes called a random codebook search). The excitation codebook includes a learning codebook configured by learning, a random codebook configured by random numbers, an algebraic codebook configured algebraically using several pulses, or a combination thereof. There are types. The Abs method is also generally used for the excitation codebook search. In the Abs method, first, a target signal for excitation codebook search is generated by subtracting a synthesized vector obtained by passing the determined optimal adaptive code vector through a vocal tract filter from the subframe vector. Next, similarly to the adaptive code vector, each excitation code vector is passed through the vocal tract filter to generate a synthesized vector. Further, a distance between a target signal of the excitation codebook search and the combined vector is measured, and an optimal excitation code vector is determined to give the shortest distance.

[0006] As described above, CELP is a sound source of speech.
Since the vocal tract model is used, the method can efficiently encode human speech, but has a disadvantage that it is weak to music signals other than human speech, background noise signals, and the like. However, applications using voice coding, such as the above-mentioned mobile phones, are often used in the real human environment, and measures against background noise are particularly desired.

As a countermeasure against the background noise of CELP,
(1) A method for reducing the noise level of an input signal using noise suppression, (2) A method for reducing unnaturalness by performing CELP coding on noise, and (3) A source codebook according to the characteristics of input speech , Etc., are proposed.

[0008] As an example of (1), a document (SFBo)
11, "Suppression of Acoustic.
c Noise in Speech Usage S
Practical Subtraction ”, IEEE
Trans. 0n ASSP, Vol. 27, no.
2, pp. 113 (Spectrum Subtract) described in pp. 113-120 (1979)).
ion) method and literature (JD Gibson, B. Ko).
o, S. D. Gray, “Filtering of
Colored Noise for Speech
Enhancement and Coding ”, I
EEE Trans. on SP, Vol. 39, N
o. 8, pp. 1732-1741 (1991)) and a noise suppression method using a Kalman filter method or the like. These methods are:
Although it is effective to reduce the level of noise to some extent, since noise cannot be completely removed, the unnaturalness caused by coding noise with CELP cannot be eliminated.

As an example of (2), JP-A-11-242
No. 499, the method of separating and coding speech and background noise using the SS method, and a document (Omuro, Maino, "Improvement of quality of speech with background noise added in low bit rate speech coding") , IEICE Technical Report, S
P98-145 (1999)) reduces the unnaturalness caused by coding background noise by CELP by estimating the background noise level and positively adding background noise to decoded speech. And the literature (Okazaki, Takahashi, "Low-rate C with post-noise smoother"
Improvement of ELP noise section quality ", a method of smoothing a reproduced voice in a spectral domain to reduce unnaturalness, described in Acoustic Society of Japan, pp. 237-238 (1998. 3)). In these methods, unpleasant sounds due to CELP coding of background noise are greatly reduced, and unnaturalness is reduced to some extent. However, since background noise is approximated by random noise, The background noise in the real environment is altered, and when the background noise is meaningful, for example, for identification of a place, there arises a problem that necessary information is not transmitted.

As (3), Japanese Patent Application Laid-Open No. 8-123493
As a sound source codebook described in Japanese Unexamined Patent Publication, a multi-pulse type codebook effective for voiced sounds and a random noise codebook effective for noise and unvoiced sounds are provided. There is a method of changing the ratio. In CELP, as described above, generally A
Perform a codebook search using the -bs method. The Abs method compares a synthesized voice with a target voice on a time axis. In the SS method, the power of the estimated noise spectrum is subtracted from the intensity of the discrete Fourier transform value of the input signal, and background information is removed by using the phase information of the input signal as it is, and the characteristic amount of the noise component is used. It can be considered that the intensity of the spectrum is more important than the similarity on the time axis.

[0011]

As described above, CELP using a learned excitation codebook is a suitable method for efficiently encoding human speech without background noise or background music. When noise or music is added to the background, there is a problem that quality is deteriorated. In addition, when a source codebook made of Gaussian noise without learning is used, there is no degradation due to background noise or background music, but the quality is poor for speech coding in a noise-free environment. there were. The present invention solves this problem in a relatively simple manner, and uses a method of switching the excitation codebook according to the characteristics of the input speech. In the excitation codebook search, if the background noise level is low, the learned excitation If the background noise level is high using the codebook, Abs on the time axis
Speech coding method and apparatus for realizing such a method by performing a search in the spectral domain without using the method, so that the speech can be satisfactorily encoded even when there is no background noise or when background noise is included. It is an object to provide

[0012]

In order to achieve the above object, the present invention relates to a speech encoding apparatus for encoding a digital speech signal, comprising: a frame dividing means for dividing the digital speech signal into a plurality of frames; A subframe dividing unit that divides the frame into subframes shorter than a frame, a formant parameter extracting unit that extracts and encodes the formant parameters of the frame from the frame, and a formant parameter of the subframe and the frame. Using an adaptive codebook search means for extracting and encoding the pitch period of the subframe using, and an excitation codebook including the subframe, the formant parameter of the frame, the pitch period, and a plurality of excitation code vectors. Extracting the noise source component of the subframe and coding An excitation codebook searching means for performing the analysis, and a subframe determining means for determining the property of the subframe. Further, the excitation codebook searching means extracts and encodes a noise source component using an analysis method by synthesis. "Analysis method by synthesis (Abs)" excitation codebook search means;
Spectrum domain excitation codebook search means for extracting and encoding an excitation codebook search component in a spectrum domain, wherein the excitation codebook search means includes an Abs excitation It is characterized in that either one of the codebook searching means and the spectral domain excitation codebook searching means is used properly. Accordingly, an excitation codebook search unit is switched by determining background noise or the like, and a speech encoding device that satisfactorily encodes speech regardless of whether background noise is present or background noise is provided. Can be.

In a speech encoding method for encoding a digital audio signal, a frame dividing step of dividing the digital audio signal into frames, and a sub-frame division dividing the frame into sub-frames shorter than the frame. A formant parameter extracting step of extracting and encoding the formant parameters of the frame from the frame; and an adaptive extracting and encoding a pitch period of the subframe using the subframe and the formant parameters of the frame. Extracting and encoding a noise source component of the subframe using a codebook search step, the subframe, the formant parameter of the frame, the pitch period, and an excitation codebook including a plurality of excitation code vectors. Excitation codebook search step; And a sub-frame determination step of determining the quality, further, the excitation codebook search process, to extract and encode the noise source components using analysis by synthesis A-
bs excitation codebook search step, and a spectral domain excitation codebook search step of extracting and encoding an excitation codebook search component in a spectrum domain, wherein the excitation codebook is determined according to the result of the subframe determination step. In the search step, one of the Abs excitation codebook search step and the spectral domain excitation codebook search step is selectively used. Accordingly, an excitation codebook search process is switched by discriminating background noise and the like, and a speech encoding method that satisfactorily encodes speech when there is no background noise or when background noise is included is provided. Can be.

[0014]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A speech coding apparatus and a speech coding method according to the present invention will be described below in detail with reference to the accompanying drawings.

FIG. 1 is a block diagram showing an embodiment of a speech coding apparatus according to the present invention. This speech coding apparatus is a speech coding apparatus for coding a digital speech input signal based on CELP, and includes a frame dividing means 101, a subframe dividing means 102, a formant parameter extracting means 103, an adaptive codebook. A search unit 104, an excitation codebook search unit 105, a gain quantization unit (gain codebook search unit) 106, and a subframe determination unit 107 are provided. Further, excitation codebook search means 105 includes target signal calculation means 201, Abs excitation codebook search means 202, and spectral domain excitation codebook search means 203.

The digital audio signal is divided by the frame dividing means 101 into processing units called frames. Examples of the frame length include values such as 20 to 30 ms. A DC component may be removed by passing through a high-pass filter before and after frame division. Further, the frame is divided by the sub-frame dividing means 102 into processing units called sub-frames. The number of subframes per frame includes a value of 2 to 6, for example.

Further, the frame is analyzed by formant parameter extracting means 103, and the formant parameter of the frame is extracted. As formant parameters, LPC (Linear Predi)
ction Coding, linear prediction) coefficient, LSP
(Line Spectrum Pair, line spectrum pair), LSF (Line Spectrum Fre)
frequency, line spectrum frequency), LPC cepstrum coefficient, and reflection coefficient. As a method of extracting the formant parameters, a linear prediction analysis can be cited. As a method of the linear prediction analysis, there is a method of calculating an autocorrelation function of a frame and calculating LPC coefficients by a Levinson-Durbin recursive method. Before calculating the autocorrelation function, a window function such as a Hamming window or a Hanning window may be applied to the frame.

Next, the formant parameters of the frame are quantized and transmitted or stored. Examples of the quantization method of the formant parameter of the frame include scalar quantization, vector quantization, multi-stage vector quantization, split vector quantization, and predictive quantization. When quantizing the formant parameters, LSP or L
It is preferable to use a parameter with high quantization efficiency such as SF.

Next, a formant parameter of the sub-frame is calculated using the quantized formant parameter of the frame. As a method of calculating the formant parameter of the sub-frame, there is a method of obtaining the formant parameter of the current and past frames by interpolation from the quantized formant parameters. The interpolation method includes linear interpolation and quadratic interpolation.

Next, in the adaptive codebook searching means 104, an adaptive codebook search corresponding to the extraction of the pitch period component of the subframe is performed for each of the subframes. In this adaptive codebook search, a predetermined number of pitch candidates are prepared in advance. Pitch candidates include an integer pitch that is an integral multiple of the sampling unit and a non-integer pitch. The following processing is performed on all of the pitch candidates.

First, an adaptive code vector is generated. The adaptive code vector is generated by cutting out the pitch length from the excitation vector up to the immediately preceding subframe and arranging it repeatedly until the subframe length is reached.
Next, a synthesized speech vector is generated by using the adaptive code vector and the formant parameter of the subframe. As a method of generating a synthesized speech vector, a method of applying a linear filter constituted by the formant parameters of the subframe to the adaptive code vector may be used.

Next, when the synthesized speech vector is multiplied by an arbitrary gain, a shortest distance closest to the subframe is calculated. In the calculation of the shortest distance, a method of performing an auditory weighting so as to minimize an auditory error may be introduced. The above processing is performed for all the pitches in the pitch candidate, and the adaptive code vector and the pitch cycle having the shortest distance are set as the adaptive code vector and the pitch cycle of the subframe. The code assigned to the pitch period of the subframe is transmitted or stored.

Further, the sub-frame discriminating means 107
In, the characteristics of the subframe are determined. In the present embodiment, the degree of background noise is determined. As a method of determining the degree of the background noise included, an autocorrelation function of the subframe is calculated, and 0 is calculated.
The ratio between the next autocorrelation function and the largest autocorrelation function excluding the zeroth order is equal to or less than a predetermined value, and the value of the zeroth order autocorrelation function (that is, the intensity of the subframe) is equal to or smaller than a predetermined value. When the value is equal to or more than the value, there is a method of determining that a large amount of background noise is included.

Further, the excitation codebook search means 105 extracts and encodes a noise source component using an excitation codebook composed of a plurality of excitation code vectors. As a sound source codebook, a random codebook, Alge
Examples include a codebook composed of a plurality of small pulses represented by a braic (algebraic) codebook, and a learning codebook composed of learning.

First, the target signal calculating means 201 subtracts a synthesized speech vector synthesized from the adaptive code vector of the subframe and the formant parameter of the subframe from the subframe, thereby obtaining a noise source search. A target signal vector is generated.

Next, based on the result of the discrimination by the subframe discriminating means 107, which one of the spectrum domain excitation codebook searching means 203 and the Abs excitation codebook searching means 202 performs the codebook search? Is determined.
If it is determined that a large amount of background noise is included, the target signal vector is sent to the spectral domain excitation codebook search means 203, and a codebook search is performed. If it is determined that a large amount of background noise is not included, the target signal vector is sent to the Abs excitation codebook search means 202, and a codebook search is performed.

Abs excitation codebook search means 202
In the first, the same distance calculation as in the adaptive codebook search means 104 is performed on the excitation code vector, and when the synthesized speech vector of each excitation code vector is multiplied by an arbitrary gain, the target The shortest distance closest to the signal vector is calculated. In this case, the distance calculation may be performed after orthogonalizing the synthesized speech vector of the excitation code vector with respect to the synthesized speech vector of the adaptive code vector.

Next, the excitation code vector that minimizes the shortest distance is determined. In the spectral domain excitation codebook searching means 203, first, as in the adaptive codebook searching means 104 and the Abs excitation codebook searching means 202, Of the excitation code vector is calculated. Further, the synthesized speech vector and the target signal are transformed into a frequency domain. Examples of the conversion method to the frequency domain include, but are not limited to, discrete Fourier transform, discrete cosine transform, and wavelet transform.

Further, the similarity between the target signal and the synthesized speech vector in the frequency domain is calculated. Examples of the similarity include, but are not limited to, a small distance when both vectors are normalized, and a small distance weighted using a masking frequency reflecting human auditory characteristics. Next, the excitation code vector that maximizes the similarity is determined. The code assigned to the excitation code vector determined here is transmitted or stored.

Further, in the gain encoding means 106,
The gain component is encoded. The gain component extraction method includes scalar quantization for separately quantizing the gain component of the adaptive code vector and the gain component of the noise code vector, and vector quantization for quantizing both of them at the same time. . The gain component quantized here is transmitted or stored. Further, in the speech encoding device of the present invention, when it is determined that a large amount of background noise is included,
Adaptive switching of the codebook configuration according to the characteristics of the input subframe may be performed, such as reducing the number of bits in the adaptive codebook, not performing the adaptive codebook search, or switching the excitation code vector. . The speech encoding apparatus of the present invention has been described above. However, the present invention also relates to a speech encoding method used in the speech encoding apparatus and a computer-readable storage element that stores an algorithm of the speech encoding method. It is assumed that.

[0031]

As described above, according to the first aspect of the present invention, in the speech coding apparatus, the excitation codebook searching means uses the Abs excitation code in accordance with the result of determination by the subframe determination means. In the case of speech containing a large amount of background noise, the sound source containing the background noise is searched for in the case of speech containing a lot of background noise. It is possible to provide a speech coding device that can perform good coding.

According to a second aspect of the present invention, a target speech coding apparatus is realized with a simpler configuration by estimating a background noise level of a subframe by a subframe discriminating means and using the background noise level for subframe discrimination. can do.

According to a third aspect of the present invention, a spectral domain excitation codebook search means converts an input target signal and a speech code vector into a frequency domain, and determines an optimal excitation code vector based on the similarity in the frequency spectrum. I will decide
A target speech encoding device can be realized with a simpler configuration.

According to a fourth aspect of the present invention, in the speech coding apparatus, in the excitation codebook search step, the Abs excitation codebook search step and the spectrum domain excitation In the case of speech that contains a lot of background noise, the sound source codebook search is performed in the frequency domain to properly encode speech containing background noise. A possible audio coding method can be realized.

According to the invention of claim 5 of the present invention, the target speech coding apparatus is realized with a simpler configuration by estimating the background noise level of the subframe in the subframe discriminating step and using it for discriminating the subframe. can do.

According to a sixth aspect of the present invention, in the spectral domain excitation codebook search step, the input target signal and the speech code vector are transformed into the frequency domain, and the optimal excitation code vector is determined based on the similarity in the frequency spectrum. I will decide
A target speech encoding method can be realized with a simpler configuration.

According to a seventh aspect of the present invention, since the algorithm of the speech encoding method according to any one of the fourth to sixth aspects is stored in a storage element readable by a computer, these methods can be easily performed. Can be implemented on a computer.

[Brief description of the drawings]

FIG. 1 is a block diagram of a speech encoding device according to the present invention.

[Explanation of symbols]

 Reference Signs List 101 frame dividing means 102 subframe dividing means 103 formant parameter extracting means 104 adaptive codebook searching means 105 excitation codebook searching means 106 gain quantization means (gain codebook searching means) 107 subframe discriminating means 201 target signal calculating means 202 A -Bs excitation codebook search means 203 spectrum domain excitation codebook search means

Claims (7)

[Claims] 1. A speech encoding apparatus for encoding a digital speech signal, comprising: a frame dividing unit for dividing the digital speech signal into a plurality of frames; and a subframe for dividing the frame into subframes shorter than the frame. Dividing means; formant parameter extracting means for extracting and encoding the formant parameters of the frame from the frame; and extracting and encoding the pitch period of the subframe using the subframe and the formant parameters of the frame. An adaptive codebook search unit, and an excitation source for extracting and encoding a noise source component of the subframe, using an excitation codebook including the subframe, the formant parameter of the frame, the pitch period, and an excitation code vector. Codebook searching means; Sub-frame discriminating means for discriminating properties; and said excitation codebook searching means extracts and encodes a noise source component using an analysis method based on synthesis. Excitation codebook search means, and spectrum domain excitation codebook search means for extracting and encoding an excitation codebook search component in the spectral domain, wherein the excitation codebook search means , A speech coding apparatus characterized in that either one of Abs excitation codebook searching means and spectrum domain excitation codebook searching means is selectively used. 2. The speech coding apparatus according to claim 1, wherein the subframe discriminating unit estimates a background noise level of the subframe and uses the estimated background noise level for discriminating the subframe. 3. The spectral domain excitation codebook search means converts an input target signal and a speech code vector into a frequency domain, and determines an optimal excitation code vector based on a similarity in a frequency spectrum. The speech encoding device according to claim 1. 4. A speech encoding method for encoding a digital audio signal, comprising: a frame dividing step of dividing the digital audio signal into frames; and a subframe dividing step of dividing the frame into subframes shorter than the frame. A formant parameter extracting step of extracting and encoding the formant parameter of the frame from the frame; and an adaptive code for extracting and encoding the pitch period of the subframe using the subframe and the formant parameter of the frame. A book search step, an excitation source for extracting and encoding a noise source component of the subframe using the subframe, a formant parameter of the frame, the pitch period, and an excitation codebook including a plurality of excitation code vectors. Codebook search step; and properties of the subframe Further, the excitation codebook searching step includes: an Abs excitation codebook searching step of extracting and encoding a noise source component using an analysis method by synthesis; And a spectrum domain excitation codebook search step of extracting and encoding an excitation codebook search component in a region. In the excitation codebook search step, according to the determination result in the subframe determination step, the Abs
A speech coding method, wherein one of an excitation codebook search step and the spectral domain excitation codebook search step is selectively used. 5. The speech encoding method according to claim 4, wherein said sub-frame discriminating step estimates a background noise level of said sub-frame and uses it for discriminating said sub-frame. 6. The spectral domain excitation codebook search step converts an input target signal and a code vector into a frequency domain, and determines an optimal excitation code vector according to a frequency spectrum similarity. A speech encoding method according to claim 4 or claim 5. 7. A storage element readable by a computer, wherein the storage element stores an algorithm based on the speech encoding method according to claim 4. Description: