JP3099852B2 - Excitation signal gain quantization method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for quantizing a gain of an excitation signal, and more particularly to a code-excited linear predictive encoding (Code Excited Linea) for a speech signal sequence with a small amount of information.
The present invention relates to an efficient excitation signal gain quantization method capable of reducing waveform distortion due to r prediction (CELP).

[0002]

2. Description of the Related Art In the technical field of transmitting or storing digital mobile radio communications, voice storage services and other information, various high-efficiency systems are used to efficiently utilize radio waves and other information transmission media or storage media. A speech coding method is employed. The CELP encoding method is an effective method for encoding audio sampled at a sampling frequency of 8 kHz at about 8 kbit / s. The method, in summary, comprises a linear predictive synthesizer that selects the shape and gain of a plurality of excitation signal vectors such that the perceptual distortion between the synthesized signal and the input speech signal is minimized. The code is transmitted by controlling the excitation source.

Hereinafter, a CELP encoding method will be described with reference to FIG. As a method of encoding audio efficiently,
The original sound is divided into sections called frames, which have a fixed interval of about 5 to 50 ms, and the sound of one frame is a signal about the envelope shape of the frequency spectrum and an excitation signal for driving a linear filter corresponding to the envelope shape. And a method of encoding each of them separately. In this case, as a method of encoding the excitation signal, the excitation signal is separated into a periodic component considered to correspond to a fundamental frequency (or a pitch period) of the voice and another component (in other words, an aperiodic component). There is a known encoding method. One type of encoding method of the excitation signal is a code excitation type linear prediction encoding method (CELP). In this encoding method, as shown in FIG. 1, a parameter representing an envelope shape of a frequency spectrum of an input speech X input to an input terminal 11 is calculated in a linear prediction analysis unit 12. This analysis typically employs linear predictive analysis. The linear prediction parameter is encoded by a linear prediction parameter encoding unit 13, the encoded output A is decoded by a linear prediction parameter decoding unit 14, and the decoded linear prediction parameter a ′ is It is set as 15 filter coefficients. By supplying an excitation signal (vector) E to be described later to the linear prediction synthesis unit 15, a reproduced synthesized speech X 'can be obtained.

Here, the excitation signal (vector) E will be described. The codebook 16 is configured to store and use a large number of fixed excitation vectors, or to always hold the determined excitation vector of the immediately preceding frame. A segment having a length L corresponding to a certain period (pitch period) is cut out from the excitation vector, and the cut-out vector segments are connected repeatedly until the frame length T is reached, and a code corresponding to the period component of the speech is obtained. The vector is output. By changing the cutout length L given to the codebook 16 as a periodic code (represented by the same symbol L as the cutout length), a code vector corresponding to a different periodic component can be output. Hereinafter, the code vector output from the codebook is referred to as an adaptive code vector.

The codebook 17 is a random number codebook, and one or more codebooks are provided. The following description is for a case where _two random number codebooks 17 ₁ and 17 ₂ are provided. Each random number codebook 17 ₁ , 17 ₂ is usually based on white Gaussian random noise, and various built-in vectors having a length L for one frame are stored in advance independently of the input voice, and given random number codes Vectors specified by C (C ₁ , C ₂ ) are read out and output as code vectors corresponding to the non-periodic components of speech. Hereinafter, the code vector output from the random number code book 17 is referred to as a random number code vector.

Each of the code vectors output from codebook 16 and codebook 17 is gain-adjusted in gain quantization section 20. In other words, each code vector is individually
_The gains are adjusted by the gains g ₀ , g ₁ , and g ₂ output from the codebook 23 in ₀ , 21 ₁ , and 21 ₂ , and these results are added in the adder 22. The codebook 23 switches or creates the gains g ₀ , g ₁ , g ₂ according to the given gain code G. The addition output E of the addition unit 22 is supplied to the linear prediction synthesis unit 15 as an excitation vector candidate, and the synthesis unit 15 outputs a synthesized reproduction voice X ′. Input terminal 1
This synthesized speech X ′ with respect to the input speech X input from
Is calculated by the distortion calculator 18. Based on the criterion for minimizing the distortion d, the auditory sensation correction unit 19 first searches the cutout length L in the codebook 16 and determines the optimal code vector of the codebook 16. Next, a random code vector is determined from the codebook 17, and the optimum gains g ₀ , g ₁ , and g ₂ of the gain quantization unit 20 are further determined. According to the above procedure, the distortion d
Is searched for, and the excitation vector E of the current frame is selected as an excitation vector candidate at that time.
Is determined. A periodic code L indicating the cut-out length of the codebook 16 when the distortion d is minimized, and codebooks 17 ₁ and 17 ₂
Random numbers C ₁ and C ₂ indicating the respective code vectors of
A gain code G indicating ₀ , g ₁ , g ₂ and a linear prediction parameter code A are output as encoded outputs and transmitted or stored.

[0007]

As for the quantization of the shape of the excitation signal vector and the quantization of the gain, a method of scalar-quantizing a gain corresponding to each excitation signal vector and a method of quantizing a plurality of excitation signals are described. There is a vector quantization method that collectively quantizes a gain corresponding to a vector. Here, regarding the characteristics of these quantization methods, the scalar quantization method requires a small amount of memory but has difficulty in reducing the waveform distortion, while the vector quantization method has This is suitable for reducing waveform distortion, but requires a codebook with a large memory amount.

Then, when the gain is vector-quantized,
As shown in FIG. 2, a plurality of codebooks 23 are used to switch and select the gain, and the characteristics of the input speech as a result of analyzing the input speech by the feature analysis unit 30 , that is,
Information on whether the input voice is voiced or unvoiced
There is also a method of adaptively switching and using these codebooks 23 using information, power, pitch period, and other characteristics . By doing so, the waveform distortion can be reduced,
The memory capacity of the codebook 23 is generally large, and there is a problem that the provision of a plurality of codebooks further increases the total memory capacity.

An object of the present invention is to provide an efficient method for quantizing the gain of an excitation source, which can reduce waveform distortion due to CELP coding with a small amount of information.

[0010]

The input speech is subjected to linear prediction analysis 12 at a predetermined sampling period, and a synthesized signal X 'having passed through a linear prediction synthesis unit 15 comprising prediction coefficients and an input speech X are input.
In the gain quantization method of the excitation signal used in the linear prediction encoding method for determining the shapes and gains of the plurality of excitation signals that minimizes the distortion between the first and second excitation signals, two-stage gain quantization is performed. in the gain quantization 20 _first stage quantization tables 31 adaptively changed in accordance with the characteristic 30 of the input speech, and the first excitation signal sequence is passed through a linear prediction synthesis unit
The synthesized signal x ′ after synthesis and the target signal sequence y
Signal sequence d = (y−gx ′) at the time of
The difference signal sequence (y-gx ') is obtained by obtaining a gain value g that minimizes the energy or distortion, and subtracting a sequence obtained by multiplying the excitation signal sequence x' by the gain g from the target signal sequence y. In the two-stage gain quantization, the difference signal sequence (y-gx ') in the first stage is set as a new target signal sequence, and the gain of the excitation signal is vector-quantized collectively for a plurality of new target signals. The construction method was constructed.

[0011]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS The gain quantization method of an excitation signal according to the present invention, in short, quantizes gains for a plurality of excitation signals in two stages. In the second stage, the gain is vector-quantized collectively so as to supplement the result of the gain quantization in the first stage.

The first embodiment of the present invention will be described in detail with reference to FIG. First, in the first stage, the codebook 16
Performing scalar quantization in the gain quantizer 20 ₁ only for the first gain of the excitation vector obtained by selecting, as supplement the results of the gain quantization in the first stage in the second stage, the first Excitation vector gain and codebook 17
The a gain of the second excitation signal vector obtained by collectively selected in the vector quantization section 20 ₂ to implement the vector quantization.

The codebook 16 is usually an adaptive codebook configured to always hold the determined excitation vector of the immediately preceding frame, and the first excitation vector is the adaptive codebook 1
6 and the second excitation signal vector is an output from codebook 17 which is a random codebook. The quantization table 31 is for switching the gain g, and a very small amount of information is sufficient. For example, the input voice, which is a property of the voice analyzed by the feature
It is adaptively switched and used in accordance with the information as to whether the sound is a voice or a voiceless sound, and the pitch cycle .

The method of quantizing the gain of the excitation signal is as follows. First. First, the shape of the first excitation vector is determined. This process uses the adaptive codebook 16. In the normal case, it is almost equivalent to obtaining the pitch period. Second.
Here, carrying out the gain quantization of the first stage in the gain quantization section 20 ₁ of the first stage. In this processing, a synthesized signal x 'obtained by passing only the first excitation signal through the linear prediction synthesizing unit 15 and synthesizing is subtracted from a response from the immediately preceding frame.
For an input voice y as a target signal, a difference signal sequence (y−
gx '), the gain g which minimizes the energy or distortion is determined. In this case, even if the level of the input voice of the frame is, for example, 0, the effect of the input voice of the immediately preceding frame extends to the frame, so that the level of the frame is slightly smaller than the response from the immediately preceding frame. It has a level. Therefore, the result of subtracting the response from the immediately preceding frame from the apparent input voice to obtain the input voice y of the true frame is used as the target signal. It is sufficient that the quantization table 31 of the gain g of the first excitation signal has a very small amount of information, and the distortion is reduced by adaptively switching and using it according to the characteristics of the voice analyzed by the feature analysis unit 30. be able to. Voiced and unvoiced information, power, and pitch cycle can be considered as switching parameters. Since the quantization in the first stage is scalar quantization, even if a large number of quantization tables are used, the amount of memory in the quantization table itself is very small. do not do.

Third. Determine the shape v of the second excitation vector. Specifically, the codebook 17 of random numbers
And then passed through the linear prediction synthesis unit 15
Later, the difference signal sequence generated in the first stage quantization
The distortion is calculated in comparison with (y−gx ′). The distortion of this result
Select the vector v which was the smallest of Fourth. The vector quantization of the gain of the second stage is carried out in the vector quantization section 20 ₂ of the gain of the second stage. In the gain quantization in the second stage, the vector quantization is performed collectively on the gain of the first excitation vector and the gain of the second excitation signal vector so as to supplement the result of the gain quantization in the first stage. . This
Here, " How to compensate for the result of gain quantization in the first stage
The meaning of "ni" is as follows. Excitation signal
In the gain quantization, generally, the total gain is 8 bits.
It is necessary to perform quantization with a small number of bits.
Scalar quantization in the first stage
Because you can only give a bit
Ri, the quantization of the gain quantizer 20 ₁ of the gain of the first stage scan
The steps are extremely coarsely set in four steps. That
Above, quantization is performed according to the characteristic parameters of the input signal.
Although distortion can be reduced by applying
The gain of the first excitation signal is 1
It doesn't mean that it depends on 00%.
Only by performing quantization, the amount of gain
A child error occurs. Hence, the second excitation signal
The second-stage vector quantization combined with
Reduces overall distortion by compensating and correcting the gain quantization error
You do it.

The vector quantization in the second stage is, specifically,
It is implemented as follows. That is, the difference signal sequence in the gain quantization of the first stage (y-gx ') as a new target signal sequence, the vector quantization in the gain vector quantizer 20 ₂ collectively gain more new target signal I do. This place
In this case, although not explicitly shown in FIG.
The difference signal sequence (y−gx ′) obtained by the color quantization is
This is stored in the buffer element, and this is supplied to the distortion calculation unit 18.
The first excitation signal and the second excitation signal are
The signal is compared with the synthesized signal after passing through the
You. A difference signal sequence after synthesis represented by the following equation e is obtained,
Use a gain vector that minimizes the energy or distortion.
Gain code that is provided to obtain the vector quantizer 20 ₂
Select from book.

E = (y−gx ′) − (g ₁ x ′ + g ₂ v ′) where g ₁ and g ₂ : gain vector elements v ′: v of the second excitation signal are sent to the linear prediction synthesis unit. After passing
As described above, since the gain vector quantization method of the present invention absorbs a large fluctuation of the gain by the quantization in the first stage, the variation range of the code vector of the codebook of the quantization in the second stage is obtained. The variation range can be reduced as compared with the variation range in the case of normal vector quantization.

To summarize the above, in ordinary CELP
In general, the first excitation signal has a pitch periodicity
And the second excitation signal is noise-like without pitch periodicity.
It is something. I will explain this normal CELP
In addition, the gain of the first excitation signal is
The degree of periodicity of the pitch period of the
This reflects the magnitude of the increase in amplitude compared to the here
As described in paragraph number [0013],
Whether the input voice is a voiced sound
Or the information of unvoiced sound, power and pitch cycle
Can be Note that these parameters and information are
Is also sent to the demultiplexer as an auxiliary parameter,
Reference is also made to vessels.

Synthetic filter for gain quantization apparatus for excitation signal
The immediately preceding excitation signal is normalized so that the power is 1
Shall be. At this time, if the input voice is voiced,
Although the degree of periodicity is high, there is a gain close to 1
On the other hand, in the case of unvoiced sounds, the periodicity of the pitch period is low.
Therefore, it is often less than 0.5. At this time, the gain
To indicate the mixing ratio of the first excitation signal and the second excitation signal
And becomes a numerical value between 0 and 1. In addition, power is big
Time is often voiced, and the gain of the first excitation signal is
Close to 1.

As described above, a signal having a pitch periodicity
The gain of a certain first excitation signal is equal to the input
Since it often depends on the characteristic parameters of the signal, the input
Distortion by performing quantization according to the characteristics of the signal
In many cases, it can be reduced. But the first excitation
Signal gain 100% dependent on input signal characteristic parameters
Does not mean that quantization
If gain quantization error occurs only by applying
And therefore combined with the first stage gain quantization
By using the vector quantization of the second stage together,
This is to reduce distortion.

Here, a conventional example of the gain quantization method and its generation will be described.
The difference between the effects of the bright gain quantization method will be explained.
You. Consider performing 8-bit quantization on the entire gain
You. Calculation amount and memo of the entire excitation signal gain quantizer
The amount is proportional to the number of elements in the codebook.
A comparison between the present invention and the conventional example will be made. First excitation signal
And two second excitation signals, the second excitation signal
When the original vector quantization is performed, for each excitation signal
It requires 2 ⁸ = 256 code vectors. Therefore,
The required number of elements is 256 × 2 = 512.

As in the conventional example, the characteristic pattern of the input signal is
The codebook, which is a set of code vectors, is represented by n
If the type is prepared and switched, the excitation signal
The number of elements of the entire gain quantization apparatus is 512 × n. This
In the case of, the distortion can be made extremely small. Against this
Then, in the first embodiment of the present invention, the total gain is 8
Consider implementing bit quantization. Scalar quantization
Give 2 bits and give vector quantization 6 bits
Shall be. Scalar quantization consists of 2 ² = 4 steps.
This step width is adjusted according to the characteristics of the input speech.
When n types are prepared and switched, a table with n × 4 elements
You just need a bull. Next, vector quantization
In this quantization, this quantization has 2 ⁶ × 2 = 128
A codebook with the number of elements is required, which is a feature of voice input.
The number of elements of the entire gain quantizer is
(N × 4 + 128 = 4n + 128).

In the embodiment of the present invention, the number of elements is (4n
+128), which is a vector quantity
In comparison with (512n), which is the number of elements
It is much smaller, but has the same effect of distortion reduction as this conventional example.
To come. That is, the entire operation of the gain quantization apparatus for the excitation signal is performed.
The amount of processing and memory is proportional to the number of elements as described above.
However, the number of elements is much smaller than in the conventional example.
The present invention suffices because the amount of arithmetic processing and memory is extremely small.
Without distortion, it is possible to obtain an excellent distortion reduction effect.
You.

A second embodiment of the present invention will be described with reference to FIG. In this case, prior to performing the vector quantization of gains in the vector quantization section 20 ₂ of the gain, triangular window level for each excitation vector in time domain 40
To separate the excitation vector signal into a total of four signals. By multiplying the excitation vector by the triangular window 40 of the level and performing the pre-operation for emphasizing the first half and the second half of the frame in this manner, distortion is reduced as compared with the first embodiment in which the excitation vector is directly vector-quantized. Can be made smaller.

A third embodiment of the present invention will be described with reference to FIG. As in the second embodiment, each excitation vector is multiplied by a triangular window 40 in the time domain, and these are separated into a total of four signal systems. The quantization of the gain in the first stage controls the quantization level according to the characteristics of the audio signal. Alternatively, a plurality of quantization tables are provided and these are switched and used. On the other hand, in the gain quantization in the second stage, the gain is collectively vector-quantized so as to supplement the result of the gain quantization in the first stage. In this case, a gain vector that minimizes distortion is selected from the codebook and transmitted in the codebook regardless of the characteristics of the voice.

In any of the embodiments described above, signal normalization can be performed prior to gain quantization using the power of the input speech, the prediction gain of the linear prediction, and the power of the excitation vector, It is possible to reduce the step width of the scalar quantization and the variation width of the code vector of the vector quantization. Usually, the quantization of the gain in the first stage may be a quantization of a relatively small number of bits. It is possible to adopt quantization (0-bit quantization) that adaptively changes only the average value.

[0027]

As described above, according to the present invention, the adaptive quantization in the first stage performs quantization in accordance with the characteristics of speech that changes from frame to frame, and all quantization is performed in the quantization in the second stage. By performing vector quantization in consideration of the excitation signal, waveform distortion can be reduced.

That is, since the gain vector quantization method of the present invention absorbs a large change in gain by the quantization in the first stage, the variation range of the code vector of the codebook of the quantization in the second stage is changed to a normal vector. The variation range can be smaller than the variation range in the case of quantization. Therefore, even if there is a code error, the effect of the code error can be reduced by protecting only the code bits in the first stage.

The distortion can be reduced without increasing the amount of calculation and the amount of memory as compared with the conventional vector quantization method of gain as shown in FIG. Also, as compared with the vector quantization method of gain that adaptively switches between codebooks of vector quantization as shown in FIG. 2, the amount of memory of the codebook can be significantly reduced without increasing distortion. Can be.

Further, the quantization table for the gain g needs only a very small amount of information, and the distortion can be reduced by adaptively switching and using it according to the characteristics of the voice analyzed by the feature analysis unit. . Although voiced and unvoiced information, power, and pitch period can be considered as switching parameters, the quantization in the first stage is scalar quantization. The increase in the total amount of memory is not a problem because it is very small.

[Brief description of the drawings]

FIG. 1 is a diagram showing a basic configuration of a CELP encoding method having a plurality of excitation signals.

FIG. 2 is a diagram showing a conventional example of a CELP encoding method for adaptively quantizing a gain.

FIG. 3 is a diagram showing a first embodiment of a gain quantization method for an excitation signal according to the present invention;

FIG. 4 is a diagram showing a second embodiment of the excitation signal gain quantization method of the present invention.

FIG. 5 is a diagram showing a third embodiment of the gain quantization method of the excitation signal according to the present invention.

[Explanation of symbols]

Reference Signs List 12 Linear prediction analysis 15 Linear prediction synthesis unit 16 Codebook 17 Codebook 20 _1- gain quantization unit 20 _2- gain vector quantization unit 30 Feature analysis unit 31 Quantization table

──────────────────────────────────────────────────続 き Continuation of front page (72) Inventor Naka Omuro 1-6, Uchisaiwaicho, Chiyoda-ku, Tokyo Nippon Telegraph and Telephone Corporation (56) References JP-A-4-1800 (JP, A) JP-A Hei 2-144598 (JP, A) (58) Field surveyed (Int. Cl. ⁷ , DB name) G10L 11/00-21/06 H03M 7/30 H04B 14/04 JICST file (JOIS)

Claims (1)

(57) [Claims] 1. A plurality of excitation signal shapes for minimizing distortion between an input speech and a synthesized signal after passing through a linear prediction synthesis section composed of prediction coefficients and performing linear prediction analysis of the input speech at predetermined sampling periods. In the gain quantization method of the excitation signal used in the linear prediction coding method for determining the gain and the gain, two-stage gain quantization is performed. In the first-stage gain quantization, the quantization is performed in accordance with the characteristics of the input speech. the table adaptively change the quantization table adaptively changed according to the characteristics of the input speech, the first excitation
After form if it the oscillating signal sequence is passed through a linear prediction synthesis unit
Difference signal when the synthesized signal x 'and the target signal sequence y are set
Signal sequence d = (y−gx ′), and its energy, that is, distortion
Is determined, and the target signal sequence y
, A difference signal sequence (y-gx ′) is generated by subtracting a sequence obtained by multiplying the excitation signal sequence x ′ by the gain g, and the difference signal sequence (y-gx ′) in the first stage is obtained in the second stage of gain quantization. ) Is a new target signal sequence, and the gains of a plurality of new target signals are collectively vector-quantized.