CN101009096A - Fuzzy judgment method for sub-band surd and sonant - Google Patents

Abstract

This invention relates to sub band voiced blur judgment method, which is characterized by the following steps: using current band voiced parameters extraction method to remove remaining four sub band voiced degree besides first sub band; multiplying the self-related function on base sound circle parameters of the first filter signal with one set gain factor as first sub band voiced degree; if the product is super one, then making it one; quantizing vectors on each sub voiced sound to get each sub band voiced belonging degree to combine excitation signals.

Description

The method of sub-band surd and sonant fuzzy judgment

Technical field

The invention belongs to the speech coding technology field, particularly low rate parametric speech coding technology.

Background technology

Voice coding in communication system, voice storage-playback, have in the consumer product of phonetic function and be widely used.International Telecommunication Union, some regional organizations and some countries had formulated a series of voice compression coding standards in succession in the last few years, were that 1.2kb/s has obtained gratifying voice quality to 16kb/s in code rate.Domestic and international research mainly concentrates on the following high-quality speech compressed encoding of 1.2kb/s speed at present, is mainly used in radio communication, secret communication, high capacity voice storage playback etc.Pumping signal is synthetic extremely important in the low rate voice coding, is the synthetic important means of present pumping signal and be with pumping signal synthetic more.The synthetic logical voiced sound degree parameters of band that mainly are fixed against of pumping signals of being with more, the key step of being with logical voiced sound degree parameter to ask at present is as follows:

(1) divides frame in chronological order to the input speech signal sampling point;

(2) present frame is extracted the pitch period parameter;

(3) present frame is extracted surplus spectral amplitude parameter;

(4) the present frame voice signal is carried out filtering through the Butterworth filter on 56 rank, its passband is respectively 0-500,500-1000,1000-2000,2000-3000 and 3000-4000Hz;

(5) according to the voice signal behind first sub-band filter the pitch period parameter of asking in the step (2) is further improved, the gene periodic quantity after the improvement is

(6) ask for value Vbp on the pitch period parameter position that the autocorrelation function of voice signal is asked for behind first sub-band filter in step (5) _{N, 1}

(7) ask for the pitch period parameter that the autocorrelation function of the autocorrelation function of voice signal behind all the other 4 sub-band filters and its time envelope is asked in step (5)

Locational value, and each subband got higher value among both;

Vbp _n，i＝max{Vbp _n，i′，Vbp _n，i″}i＝2，3，4，5

Wherein, Vbp _{N, i}The autocorrelation function of the i subband of ' expression present frame (n frame) exists

Locational value; Vbp _{N, i}" autocorrelation function of the temporal envelope of the i subband of expression present frame (n frame) exists

Locational value.

(8) if the value Vbp that asks in the step (6) _{N, 1}Less than 0.6, then the value in the step (7) all is revised as 0, that is, and Vbp _{N, i}=0 i=2,3,4,5, otherwise remain unchanged;

(9) value of asking in step (7) and the step (8) is done binaryzation, promptly make comparisons with threshold value 0.6 respectively,, think that then current sub is a voiced sound, represent that with 1 its band leads to the voiced sound degree, otherwise be voiceless sound, represent that with 0 its band leads to the voiced sound degree, that is: if greater than 0.6

(10) the logical voiced sound degree of band of each subband of obtaining in the step (9), surplus spectral amplitude parameter in the step (3) and the pitch period parameter in the step (5) are used for together synthetic pumping signal.

Above-mentioned prior art has adopted 0,1 judgement for the expression of subband voiced sound degree, and promptly current sub is not that voiced sound is exactly a voiceless sound.In fact not having clear and definite boundary between the voiced sound of subband and the voiceless sound, simply is the not nature that voiceless sound or voiced sound can cause the speech frame transition with sub-band division.

As shown in Figure 1, when the logical voiced sound of expression band was spent, original technology adopted simple 0,1 judgement, and this can make the synthetic sense of low code check parametric speech coding strengthen, and naturalness descends.

Summary of the invention

The objective of the invention is for overcoming the weak point of prior art, changing old is not that voiceless sound is exactly the simple decision method of voiced sound, proposes the method for the logical pure and impure sound fuzzy judgment of band, strengthens the naturalness of voice.

The method of the sub-band surd and sonant fuzzy judgment that the present invention proposes may further comprise the steps:

(1) divide frame in chronological order to the input speech signal sampling point:

(2) present frame is extracted the pitch period parameter;

(3) present frame is extracted surplus spectral amplitude parameter;

(4) the present frame voice signal is carried out filtering through the Butterworth filter on 56 rank, obtain passband and be respectively 0-500,500-1000,1000-2000,5 subband signals of 2000-3000 and 3000-4000Hz;

(5) voice signal according to first subband further improves the pitch period parameter of asking in the step (2), and the gene periodic quantity after the improvement is

(6) ask for the pitch period parameter that the autocorrelation function of first subband voice signal is asked in step (5) Locational value Vbp _{N, 1}

(7) ask for the pitch period parameter that the autocorrelation function of the autocorrelation function of all the other 4 subband voice signals and its time envelope is asked in step (5) Locational value, and respectively each subband is got higher value among both:

Vbp _n，i＝max{Vbp _n，i′，Vbp _n，i″}i＝2，3，4，5

Wherein, Vbp _{N, i}' expression present frame, promptly the autocorrelation function of the i subband of n frame exists

Locational value; Vbp _{N, i}" the expression present frame, promptly the autocorrelation function of the temporal envelope of the i subband of n frame exists

Locational value;

(8) if the value Vbp that asks in the step (6) _{N, 1}Less than 0.6, then the value in the step (7) all is revised as 0, i.e. Vbp _{N, i}=0 i=2,3,4,5, otherwise remain unchanged;

(9) with the value Vbp that asks in the step (6) _{N, 1}Multiply by the gain factor of a setting, this gain factor is set at 1.2, if the back result that multiplies each other then makes it equal 1, that is: greater than 1

$\mathrm{Vb}{p}_{n,1}=\left\{\begin{array}{ccc}1& \mathrm{if}& \mathrm{Vb}{p}_{n,1}\×1.2>1\\ \mathrm{Vb}{p}_{n,1}\×1.2& \mathrm{if}& \mathrm{Vb}{p}_{n,1}\×1.2\≤1\end{array}\right.$

(10) with the value asked in step (8) and the step (9) the voiced sound degree of membership as each subband, merging becomes a vector $\stackrel{\‾}{\mathrm{Vbp}}=({\mathrm{Vbp}}_{n,1},{\mathrm{Vbp}}_{n,2},{\mathrm{Vbp}}_{n,3},{\mathrm{Vbp}}_{n,4},{\mathrm{Vbp}}_{n,5}),$ Carry out vector quantization together; Vector quantization adopts the method that the code word in the code book is searched for entirely to obtain the optimum quantization code word:

Wherein

Represent input vector to be quantified,

Code word vector in the expression code book, C represents code book, and i is the index value of code word vector in code book, and the distortion measure that Er () function representation is specific adopts the minimum weight square error here, promptly $\mathrm{Er}(\stackrel{\‾}{\mathrm{Vbp}},\stackrel{\‾}{{\mathrm{Vbp}}_i})=\underset{k=1}{\overset{5}{\mathrm{\Σ}}}{W}_k{({\mathrm{Vbp}}_k-{\mathrm{Vbp}}_{i,k})}^2,$ Vbp wherein _kBe k component of vector to be quantified, Vbp _{I, k}Be k component of i code word vector in the code book, W is the weighting factor vector, gets W=[16 here, 8,4,2,1]; Obtain the voiced sound degree of membership of each subband of quantizing after the quantification

(11) the voiced sound degree of membership of each subband of obtaining in the step (10), surplus spectral amplitude parameter in the step (3) and the pitch period parameter in the step (5) are used for together synthetic pumping signal.

Characteristics of the present invention are the logical voiced sound degree parameter of the band in the low rate parametric speech coding have been adopted the method for fuzzy judgment.Original technology has adopted 0,1 judgement to being with logical voiced sound degree parameter, has increased the synthetic sense of voice, has reduced naturalness.The present invention adopts the relative theory of fuzzy mathematics, decides the voiced sound degree of membership of current sub with autocorrelation function.And consider that first subband often is subjected to The noise bigger, so its voiced sound degree of membership will be multiplied by a gain factor, the more accurate voiced sound degree of reasonably having described each subband.

This method can improve the naturalness of synthetic speech.The most suitable 600～800b/s the low rate of this method parametric speech coding will be realized on signal processor chip DSP.

Description of drawings

Fig. 1 is the logical voiced sound degree decision method FB(flow block) of the band of prior art.

The logical voiced sound degree fuzzy decision method FB(flow block) of band that Fig. 2 proposes for the present invention.

Embodiment

The logical voiced sound degree fuzzy decision method of the band that the present invention proposes reaches embodiment in conjunction with the accompanying drawings and further specifies as follows:

Method flow of the present invention may further comprise the steps as shown in Figure 2:

(1) divides frame in chronological order to the input speech signal sampling point;

(2) present frame is extracted the pitch period parameter;

(3) present frame is extracted surplus spectral amplitude parameter;

(4) the present frame voice signal is carried out filtering through the Butterworth filter on 56 rank, obtain passband and be respectively 0-500,500-1000,1000-2000,5 subband signals of 2000-3000 and 3000-4000Hz:

(5) voice signal according to first subband further improves the pitch period parameter of asking in the step (2), and the gene periodic quantity after the improvement is

(6) ask for the pitch period parameter that the autocorrelation function of first subband voice signal is asked in step (5)

Locational value Vbp _{N, 1}

(7) ask for the pitch period parameter that the autocorrelation function of the autocorrelation function of all the other 4 subband voice signals and its time envelope is asked in step (5)

Locational value, and respectively each subband is got higher value among both:

Vbp _n，i＝max{Vbp _n，i′，Vbp _n，i″}i＝2，3，4，5

Wherein, Vbp _{N, i}' expression present frame, promptly the autocorrelation function of the i subband of n frame exists

Locational value; Vbp _{N, i}" the expression present frame, promptly the autocorrelation function of the temporal envelope of the i subband of n frame exists

Locational value;

(8) if the value Vbp that asks in the step (6) _{N, 1}Less than 0.6, then the value in the step (7) all is revised as 0, i.e. Vbp _{N, i}=0 i=2,3,4,5, otherwise remain unchanged;

(9) with the value Vbp that asks in the step (6) _{N, 1}Multiply by the gain factor of a setting, this gain factor is set at 1.2, if the back result that multiplies each other then makes it equal 1, that is: greater than 1

${\mathrm{Vbp}}_{n,1}=\left\{\begin{array}{ccc}1& \mathrm{if}& {\mathrm{Vbp}}_{n,1}\×1.2>1\\ {\mathrm{Vbp}}_{n,1}\×1.2& \mathrm{if}& {\mathrm{Vbp}}_{n,1}\×1.2\≤1\end{array}\right.$

(10) with the value asked in step (8) and the step (9) the voiced sound degree of membership as each subband, merging becomes a vector $\stackrel{\‾}{\mathrm{Vbp}}=({\mathrm{Vbp}}_{n,1},{\mathrm{Vbp}}_{n,2},{\mathrm{Vbp}}_{n,3},{\mathrm{Vbp}}_{n,4},{\mathrm{Vbp}}_{n,5}),$ Carry out vector quantization together; Vector quantization adopts the method that the code word in the code book is searched for entirely to obtain the optimum quantization code word:

Wherein

Represent input vector to be quantified, Code word vector in the expression code book, C represents code book, and i is the index value of code word vector in code book, and the distortion measure that Er () function representation is specific adopts the minimum weight square error here, promptly $\mathrm{Er}(\stackrel{\‾}{\mathrm{Vbp}},\stackrel{\‾}{\mathrm{Vb}{p}_i})=\underset{k=1}{\overset{5}{\mathrm{\Σ}}}{W}_k{(\mathrm{Vb}{p}_k-\mathrm{Vb}{p}_{i,k})}^2,$ Vbp wherein _kBe k component of vector to be quantified, Vbp _{I, k}Be k component of i code word vector in the code book, W is the weighting factor vector, gets W=[16 here, 8,4,2,1]; Obtain the voiced sound degree of membership of each subband of quantizing after the quantification

(11) the voiced sound degree of membership of each subband of obtaining in the step (10), surplus spectral amplitude parameter in the step (3) and the pitch period parameter in the step (5) are used for together synthetic pumping signal.

The specific embodiment of each step of said method of the present invention is described in detail as follows respectively:

Said method step (1) divides the embodiment of frame to be by the 8kHz frequency sampling, to remove the voice sampling point that power frequency is disturbed through high-pass filtering to the input speech signal sampling point in chronological order.Every 25ms, just 200 voice sampling points constitute a frame;

The embodiment of said method step (2) is: the pitch period parameter p of asking for present frame by the described method of linear prediction (MELP) speech coding algorithm standard of the 2400b/s of U.S. government mixed excitation _n

The embodiment of said method step (3) is: ask the surplus spectral amplitude parameter of present frame by the described method of linear prediction (MELP) speech coding algorithm standard of the 2400b/s of U.S. government mixed excitation, be designated as vector R, its dimension is k, R=[r ₁, r ₂..., r _k] k=10;

The embodiment of said method step (4) is: the described method of linear prediction (MELP) speech coding algorithm standard by the 2400b/s of U.S. government mixed excitation is carried out bandpass filtering to the present frame voice signal;

Embodiment in the said method step (5) is: the described method of linear prediction (MELP) speech coding algorithm standard by the 2400b/s of U.S. government mixed excitation is further improved the pitch period parameter of present frame voice signal, and the pitch period parameter after the improvement is

The embodiment of said method step (6) is: by the described method of linear prediction (MELP) speech coding algorithm standard of the 2400b/s of U.S. government mixed excitation ask for first subband voice signal autocorrelation function of present frame voice signal and The value at place is as the voiced sound degree of membership Vbp of this subband ₁

The embodiment of said method step (7) is: ask for behind the present frame voice signal 4 subband filtering signals and the envelope signal autocorrelation function exists by the described method of linear prediction (MELP) speech coding algorithm standard of the 2400b/s of U.S. government mixed excitation

The value at place, and get higher value among both as the voiced sound degree of membership Vbp of this subband _i

The embodiment of said method step (8) is: if Vbp ₁＜0.6, Vbp then _i=0, i=2,3,4,5;

The embodiment of said method step (9) is: Vbp ₁=Vbp ₁* 1.2, if Vbp ₁＞1, then make Vbp ₁=1;

The specific practice of the embodiment of said method step (10) is: to being with logical voiced sound degree parameter

Carry out vector quantization, $\stackrel{\‾}{\mathrm{Vbp}}=[{\mathrm{Vbp}}_1,{\mathrm{Vbp}}_2,{\mathrm{Vbp}}_3,{\mathrm{Vbp}}_4,{\mathrm{Vbp}}_5].$ The code book of vector quantization needs training separately, and training algorithm adopts the LBG algorithm iteration to generate.When code word training and search, distortion measure adopts the minimum weight square error, weighting factor W, promptly $\mathrm{Er}=\underset{i=1}{\overset{5}{\mathrm{\Σ}}}{W}_i{({\mathrm{Vbp}}_i-{\mathrm{Vb}\hat{p}}_i)}^2,$ Wherein, Er is a distortion measure, Vbp _iBe i component of trained vector or vector to be quantified, Be i component of vector in the code book, W is the weighting factor vector, W=[16,8,4,2,1]; The code word search adopts the method for full search to obtain the optimum quantization code word, promptly travels through code words all in the code book and gets the code word of distortion minimum as the last result who quantizes.

The specific practice of the embodiment of said method step (11) is: by the described method of linear prediction (MELP) speech coding algorithm standard of the 2400b/s of U.S. government mixed excitation, utilize the logical voiced sound degree parameter of band, pitch period parameter, surplus spectral amplitude parameter to carry out pumping signal and synthesize.

Claims (3)

1, the method for sub-band surd and sonant fuzzy judgment is characterized in that this method may further comprise the steps:

(1) divides frame in chronological order to the input speech signal sampling point;

(2) present frame is extracted the pitch period parameter;

(3) present frame is extracted surplus spectral amplitude parameter;

(4) the present frame voice signal is carried out filtering through the Butterworth filter on 56 rank, obtain passband and be respectively 0-500,500-1000,1000-2000,5 subband signals of 2000-3000 and 3000-4000Hz;

(5) voice signal according to first subband further improves the pitch period parameter of asking in the step (2), and the gene periodic quantity after the improvement is

(6) ask for the pitch period parameter that the autocorrelation function of first subband voice signal is asked in step (5)

Locational value Vbp _{N, 1}

(7) ask for the pitch period parameter that the autocorrelation function of the autocorrelation function of all the other 4 subband voice signals and its time envelope is asked in step (5)

Locational value, and respectively each subband is got higher value among both:

Vbp _n，i＝max{Vbp _n，i′，Vbp _n，i″}i＝2，3，4，5

Wherein, Vbp _{N, i}' expression present frame, promptly the autocorrelation function of the i subband of n frame exists Locational value; Vbp _{N, i}" the expression present frame, promptly the autocorrelation function of the temporal envelope of the i subband of n frame exists

Locational value;

(8) if the value Vbp that asks in the step (6) _{N, 1}Less than 0.6, then the value in the step (7) all is revised as 0, i.e. Vbp _{N, i}=0 i=2,3,4,5, otherwise remain unchanged;

(9) with the value Vbp that asks in the step (6) _{N, 1}Multiply by the gain factor of a setting, this gain factor is set at 1.2, if the back result that multiplies each other then makes it equal 1, that is: greater than 1

${\mathrm{Vbp}}_{n,1}=\left\{\begin{array}{ccc}1& \mathrm{if}& {\mathrm{Vbp}}_{n,1}\×1.2>1\\ {\mathrm{Vbp}}_{n,1}\×1.2& \mathrm{if}& {\mathrm{Vbp}}_{n,1}\×1.2\≤1\end{array}\right.$

(10) with the value asked in step (8) and the step (9) the voiced sound degree of membership as each subband, merging becomes a vector $\stackrel{\‾}{\mathrm{Vbp}}=({\mathrm{Vbp}}_{n,1},{\mathrm{Vbp}}_{n,2},{\mathrm{Vbp}}_{n,3},{\mathrm{vbp}}_{n,4},{\mathrm{Vbp}}_{n,5})$ , carry out vector quantization together; Vector quantization adopts the method that the code word in the code book is searched for entirely to obtain the optimum quantization code word:

Wherein

Represent input vector to be quantified,

Code word vector in the expression code book, C represents code book, and i is the index value of code word vector in code book, and the distortion measure that Er () function representation is specific adopts the minimum weight square error here, promptly $\mathrm{Er}(\stackrel{\‾}{\mathrm{Vbp}},{\stackrel{\‾}{\mathrm{Vbp}}}_i)=\underset{k=1}{\overset{5}{\mathrm{\Σ}}}{W}_k{({\mathrm{Vbp}}_k-{\mathrm{Vbp}}_{i,k})}^2$ , Vbp wherein _kBe k component of vector to be quantified, Vbp _{I, k}Be k component of i code word vector in the code book, W is the weighting factor vector, gets W=[16 here, 8,4,2,1]; Obtain the voiced sound degree of membership of each subband of quantizing after the quantification

(11) the voiced sound degree of membership of each subband of obtaining in the step (10), surplus spectral amplitude parameter in the step (3) and the pitch period parameter in the step (5) are used for together synthetic pumping signal.

2, by the described sub-band surd and sonant fuzzy decision method of claim 1, it is characterized in that each frame comprises 180 or 200 voice sampling points in the described step (1).

3,, it is characterized in that synthetic pumping signal is directly used the voiced sound degree of membership of each subband after the quantification that obtains in the step (10) in the step (11) by the described sub-band surd and sonant fuzzy decision method of claim 1.

Images