CN1716380A - Audio frequency splitting method for changing detection based on decision tree and speaking person - Google Patents

Abstract

The audio splitting method based on the decision tree and the talker change detection includes the first adaptive silencing detection to find out the mute in audio frequency and the coarse splitting of audio frequency signal with the mute; the subsequent mutation detection for fine splitting of audio frequency signal and classifying the audio segment into phonetic part and non-phonetic part with the decision tree; and the final detecting the talker change point in the phonetic segment to obtain the final splitting result. The present invention performs phonetic detection via combining two methods of mute detection and mutation detection and adopting phonetic/non-phonetic decision tree to raise the accuracy of phonetic detection, and performs the talker change detection in phonetic segment with saving in calculation time.

Description

Change the audio frequency splitting method of detection based on decision tree and speaker

Technical field

The present invention relates to signal Processing and pattern-recognition, mainly is a kind of audio frequency splitting method that changes detection based on decision tree and speaker.

Background technology

Speaker's retrieval technique is meant utilizes signal Processing and mode identification method, retrieval speaker dependent's technology in a large amount of audio documents.Speaker's retrieval technique need solve two problems, and promptly who is speaking and when is speaking.Whom common speaker's retrieval solves in the problem of speaking by speaker Recognition Technology, then needs to use audio segmentation when in a minute.

Dividing method commonly used has based on cutting apart of bayesian information criterion and cutting apart based on the KL2 distance.The dividing method of bayesian information criterion determines whether cutting apart by Bayes's value of calculating " two section audio features are obeyed same Gaussian distribution " and " two section audio features are obeyed two Gaussian distribution respectively " two hypothesis.But bayesian information criterion often only is confined to cutting apart between the speaker, lacks robustness for the irregular situation of characteristic distribution such as noise.The arithmetic speed of bayesian information criterion is slower in addition, is unfavorable for real-time processing.

Based on the dividing method of the KL2 distance KL2 distance of MFCC relatively, and relatively come to determine speaker's change with empirical value.But come from the window of mobile regular length based on the voice segments that the algorithm of KL2 distance is used for computed range, make distance value and unreliable.

Summary of the invention

The present invention will solve the existing defective of above-mentioned technology, a kind of audio frequency splitting method that changes detection based on decision tree and speaker is provided, change by detecting voice and speaker, realize audio segmentation one-tenth is belonged to the voice segments of different people, be used for the audio segmentation of speaker's retrieval.

The technical solution adopted for the present invention to solve the technical problems: at first utilize adaptive silence detection to find out quiet in the audio frequency, and utilize these quiet audio frequency to be carried out coarse segmentation, detect to segment according to sudden change then and cut, and come to carry out the classification of speech/non-speech to cutting apart the audio fragment that obtains with decision tree, the detection speaker changes a little between sound bite at last, is changed by the speaker a little to obtain final segmentation result.

The technical solution adopted for the present invention to solve the technical problems can also be further perfect.Described silence detection is calculated each frame energy after audio frequency being divided frame, determine quiet by adaptive energy threshold value and time threshold.Described sudden change detects to determine catastrophe point by the distance between the distribution of calculating energy and zero-crossing rate.Described decision tree is the decision rule of one group of precondition, the section feature and the rule of correspondence of audio-frequency fragments is judged successively, by the final value decision audio fragment type of decision tree.Described speaker changes detection for distance between the adjacent voice segments and adaptive threshold are compared, and determines that the speaker changes a little.

The effect that the present invention is useful is: detect two kinds of methods in conjunction with silence detection and sudden change, and adopt the speech/non-speech decision tree to carry out speech detection, utilize advantage separately to improve the speech detection accuracy.Between voice snippet, carry out the speaker again and change detection, compare general needs in twos the clustering algorithm of computed range more save computing time.

Description of drawings

Fig. 1 is a speech/non-speech categorised decision tree topology structural drawing of the present invention;

Fig. 2 is a method flow diagram of the present invention;

Embodiment

The invention will be described further below in conjunction with drawings and Examples: this audio frequency splitting method that changes detection based on decision tree and speaker was divided into for six steps:

The first step: audio frequency pre-service

The audio frequency pre-service is divided into sample quantization, zero-suppresses and floats, three parts of pre-emphasis and windowing.

1, sample quantization

A), sound signal is carried out filtering, make its nyquist frequency F with sharp filter _NBe 4KHZ;

B), audio sample rate F=2F is set _N

C), to sound signal S _a(t) sample by the cycle, obtain the amplitude sequence of digital audio and video signals $s\left(n\right)=s_a\left(\frac{n}{F}\right);$

D), s (n) is carried out quantization encoding, the quantization means s ' that obtains amplitude sequence (n) with pulse code modulation (pcm).

2, zero-suppress and float

A), calculate the mean value s of the amplitude sequence that quantizes;

B), each amplitude is deducted mean value, obtain zero-suppressing that to float back mean value be 0 amplitude sequence s " (n).

3, pre-emphasis

A), Z transfer function H (the z)=1-α z of digital filter is set ^-1In pre emphasis factor α, α desirable 1 or slightly little value than 1:

B), s " (n) by digital filter, obtain the suitable amplitude sequence s (n) of high, medium and low frequency amplitude of sound signal.

4, windowing

A), calculate frame length N (32 milliseconds) and the frame amount of the moving T (10 milliseconds) of audio frame, satisfied respectively:

$\frac{N}{F}=0.032$

$\frac{T}{F}=0.010$

Here F is an audio sample rate, and unit is Hz;

B), be that N, the frame amount of moving are T with the frame length, s (n) is divided into～the audio frame F of series _m, each audio frame comprises N audio signal samples;

C), calculate the hamming code window function:

D), to each audio frame F _mAdd hamming code window:

$\mathrm{\ω}\left(n\right){\×F}_m\left(n\right)\⇒\left\{F_m^{\′}\left(n\right)\right|n=\mathrm{0,1},\·\·\·,N-1\}.$

Second step: feature extraction

Feature extraction on the audio frame comprises energy, the extraction of zero-crossing rate and Mel cepstrum coefficient (MFcc).

1, the extraction of energy:

$E=\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{s}^2\left(n\right)$

2, the extraction of zero-crossing rate:

$\mathrm{Zcr}=\frac{1}{2N-1}\underset{n=1}{\overset{N-1}{\mathrm{\Σ}}}\left|\right[\mathrm{sgn}\left(s(n+1)\right)-\mathrm{sgn}\left(s\left(n\right)\right)\left]\right|$

3, the extraction of MFCC:

A), the exponent number p of Mel cepstrum coefficient is set;

B), be fast fourier transform FFT, time-domain signal s (n) is become frequency domain signal X (k).

C), calculate Mel territory scale:

$M_i=\frac{i}{p}\×2595\log (1+\frac{8000/2.0}{700.0}),(i=\mathrm{0,1,2},...,p)$

D), calculate corresponding frequency domain scale:

$f_i=700\×e^{\frac{M_i}{2595}\ln 10}-1,(i=\mathrm{0,1,2},...,p)$

E), calculate each Mel territory passage φ _JOn the logarithm energy spectrum:

$E_j=\underset{k=0}{\overset{\frac{K}{2}-1}{\mathrm{\Σ}}}{\mathrm{\φ}}_j\left(k\right){\left|X\left(k\right)\right|}^2$

Wherein $\underset{k=0}{\overset{\frac{K}{2}-1}{\mathrm{\Σ}}}{\mathrm{\φ}}_j\left(k\right)=1.$

F), be discrete cosine transform DCT

The 3rd step, silence detection

1, the calculating of energy threshold

Detect the quiet significant limitation that has with unified energy threshold because the audio power under the various environment differs greatly, but voice and quiet between the relativeness of energy size be constant, so can calculate adaptive threshold:

Threshold(E)＝min(E)+0.3×[mean(E)-min(E)]

Wherein, Threshold (E) is the adaptive energy threshold value, and min (E) is the minimum value of each frame energy, and mean (E) is the mean value of each frame energy.

2, quiet section detection

The energy and the energy threshold T of each audio frame are compared, and the frame that is lower than threshold value is quiet frame.Continuous quiet frame is formed one quiet section.

3, the calculating of zero-crossing rate threshold value

Threshold(Zcr)＝0.5×mean(Zr _i)，i∈{i|E _i＜Threshold(E)}

Wherein, Threshold (Zcr) is a self-adaptation zero-crossing rate threshold value, mean (Zcr _i) be the zero-crossing rate mean value of quiet frame.

4, quiet section zero-crossing rate correction

Check each frame zero-crossing rate successively from each two ends of quiet section,, then be considered as the initial or voiceless sound when finishing of syllable, shift out quiet section if be higher than threshold value.

5, smoothing processing

Being lower than quiet section of 10 frames (0.1 second) is regarded as the pause in short-term between continuous speech and casts out.

The 4th step, voice are cut apart

After the silence detection, sound signal is divided into continuous quiet section and non-quiet section.For avoiding noise jamming, need to length greater than 10 seconds non-quiet section further cut apart.

1, the parameter estimation of energy and zero-crossing rate distribution

In non-quiet section of need further cut apart, be that window is long with 50 frames, 10 frames are the window step-length, calculate the energy and the zero-crossing rate X of 50 frames in each window ²The parameter that distributes:

$p\left(x\right)=\frac{x^a{e}^{-\mathrm{bx}}}{b^{a+1}\mathrm{\Γ}(a+1)},(x\≥0)$

Wherein $a=\frac{{\mathrm{\μ}}^2}{{\mathrm{\σ}}^2}-1,b=\frac{{\mathrm{\σ}}^2}{\mathrm{\μ}}$ , μ is a mean value, σ is a variance.

Distance calculation

The distance definition of each window is as follows:

$D\left(i\right)=1-\frac{\mathrm{\Γ}(\frac{a_{i-1}+a_{i+1}}{2}+1)}{\sqrt{\mathrm{\Γ}(a_{i-1}+1)\mathrm{\Γ}(a_{i+1}+1)}}\frac{2^{\frac{a_{i-1}+a_{i+1}}{2}+1}{b}_{i-1}^{\frac{a_{i+1}+1}{2}}{b}_{i+1}^{\frac{a_{i-1}+1}{2}}}{{(b_{i-1}+b_{i+1})}^{\frac{a_{i-1}+a_{i+1}}{2}+1}}$

a _I-1, a _I+1, b _I-1, b _I+1The parameter a of window before and after being respectively, b

3, sudden change detects

In each has the window of maximum value distance D (i), once more each frame is calculated same distance, getting the maximum frame of distance is cut-point.

The 5th step, speech/non-speech classification

We come non-quiet section classification with the decision tree that trains: voice or non-voice.The corresponding section feature of each node of decision tree.The topological diagram of decision tree is seen accompanying drawing one.

Selected section feature is as follows:

A), high zero-crossing rate ratio HZCRR:

$\mathrm{HZCRR}=\frac{1}{2N}\underset{n=0}{\overset{N-1}{\mathrm{\Σ}}}[\mathrm{sgn}(\mathrm{Zcr}\left(n\right)-1.5\×\stackrel{\‾}{\mathrm{Zcr}})+1]$

The relative noise of HZCRR distribution center of voice segments and music etc. are higher.HZCRR is higher than f _hSection be regarded as voice segments.

B), low-yield ratio LRMSR:

$\mathrm{LRMSR}=\frac{1}{2N}\underset{n=0}{\overset{N-1}{\mathrm{\Σ}}}[\mathrm{sgn}(0.5\×\stackrel{\‾}{E}-E\left(n\right)+1]$

The relative noise of LRMSR distribution center of voice segments and music etc. are lower.LRMSR is lower than f _lSection be regarded as voice segments.

C), fundamental frequency Mean F

The fundamental frequency of audio section can be estimated with zero-crossing rate:

MeanF＝max[Zcr(n)]×F/N，

Wherein F is a sample frequency 8000, and N is frame length 32ms.The fundamental frequency of voice distributes narrow than non-voice, so fundamental frequency is higher than f _fSection be regarded as non-speech segment.

D), no zero-crossing rate space-number NZCRR

The number of times that zero-crossing rate is zero frame appears in the definition section of being of NZCRR the inside.Continuous no zero passage frame is only calculated once.

General voice can not have or not zero-crossing rate at interval, so NZCRR is lower than f _nSection be regarded as voice segments.

E), energy variance VarRMS

The variance of the definition section of the being self-energy of Var RMS.

The energy variation of voice is little a lot of than music etc., so variance is less than f _vSection be regarded as voice segments.

F), f _h, f _l, f _f, f _n, f _vValue all obtain by decision tree training.The language material of training decision tree comprises about 30 minutes voice and about 30 minutes noise, and wherein voice are the voice that 20 speakers (10 male sex, 10 women) enroll under office environment; Noise comprises 5 minutes white noises, 5 minutes Gaussian noises, 10 minutes music and 10 minutes neighbourhood noises.

The 6th step, speaker change detection

Each speaker's phonetic feature has all formed specific distribution in feature space, can describe speaker's individual character with this distribution.Different speakers' distribution is also different, so can detect speaker's change with the similarity between the characteristic distribution.Here we use T ²Distance is calculated the MFCC characteristic distance between each voice segments.

1, T ²Distance calculation

Change in order to detect the speaker, need to calculate the T between per two adjacent voice segments ²Distance.T ²Distance definition

As follows:

$T^2=\frac{\mathrm{ab}}{a+b}{({\mathrm{\μ}}_1-{\mathrm{\μ}}_2)}^T{\mathrm{\Σ}}^{-1}({\mathrm{\μ}}_1-{\mathrm{\μ}}_2)$

A wherein, the length of the b section of being, μ ₁, μ ₂Be the mean value of MFCC in each section, ∑ is common covariance matrix.

2, adaptive threshold calculates

By comparing T ²Whether distance and threshold value can detect and exist the speaker to change.The computing formula of adaptive threshold is as follows:

T＝μ+λσ

Wherein μ is a mean distance, and σ is a distance variance, and λ is a penalty coefficient, is set as-1.5 here.

3, merge

If the distance between two voice segments is less than threshold value, these two voice segments are regarded as belonging to same speaker so, these two voice segments can be merged into one.If exist quietly between these two voice segments, this section is quiet so also will merge.If have non-voice between two voice segments, then nonjoinder.This is in order to prevent the interference of noise.

Experimental result:

This method is tested on 1997 Mandarin Broadcast News Speech Corpus (Hub4-NE) news broadcast voice.This sound bank comprises CCTV, the news broadcast of KAZN and VOA, and about 40 hours T.T.s, wherein about 10 hours content is music or noise.

We use on this storehouse simultaneously based on the dividing method of bayesian information criterion with based on the dividing method of KL2 distance and have carried out same experiment, are used for comparing with this method.These two kinds of methods all are directly to change with speaker characteristic MFCC search speaker between fixing window long (1 second).

The likelihood score and the parameter use number that compare the parameter estimation of two hypothesis based on the method for bayesian information criterion.1: two window of wig belongs to same speaker, and feature is obeyed same Gaussian distribution; Suppose that 2: two windows belong to different speakers, feature is obeyed two Gaussian distribution respectively.If suppose that Bayes's value (likelihood score deducts the penalty term number of parameters) of 2 is higher, then thinking has the speaker to change.

The KL2 distance is to be used for the method that the speaker is cut apart.By the KL2 between the speaker characteristic that calculates two sections voice distance and with threshold ratio, exist the speaker to change with detection.

We carry out the assessment of five aspects to the result of partitioning algorithm:

1) cut-point false drop rate: the cut-point of mistake accounts for the ratio that detects cut-point

2) cut-point loss: nd cut-point accounts for the ratio of actual cut-point

3) pure voice ratio: detect the ratio that pure voice segments total length accounts for the actual speech total length

4) voice segments recall rate: the actual speech section ratio that is detected

5) error rate such as retrieval: the value when false rejection rate equates with wrong acceptance rate in speaker's retrieval

The definition of pure voice segments is the voice segments that only comprises speaker's voice.The voice segments that comprises noise or a plurality of speaker's voice is impure voice segments.Pure voice ratio is the ratio that pure voice segments total length accounts for whole voice length.The voice segments recall rate is meant the voice segments ratio that is detected corresponding pure voice segments.These two indexs can better be weighed the effect of segmentation effect to speaker's retrieval, are replenishing of false drop rate and loss.Error rates such as retrieval be on the basis of cutting apart as a result, be the speaker retrieve experiment etc. error rate.This index is used for weighing the final effect of partitioning algorithm.

Experimental result is as follows:

Algorithm	False drop rate	Loss	Pure voice ratio	Recall rate	Etc. error rate
						BIC	25.87％	13.37％	72.39％	85.42％	15.91％
KL2	25.50％	14.42％	71.69％	83.72％	25.84％
						This method	26.48％	10.47％	80.40％	89.29％	12.94％

Each method is as follows working time:

Algorithm	Processing time (second)	Speed (minute audio frequency/second)
			BIC	2190	1.08
KL2	1331	1.78
			This method	883	2.69

Experimental machine device configuration CPU is AMD Athlon (tm) XP2500+, in save as 512M ddr400.

Experimental result shows that with respect to BIC and KL2 method, the working time of this dividing method is the shortest, and cuts apart and obtain at most the longest pure voice, thereby the performance of raising speaker that can be best retrieval makes that its mistake such as grade is minimum.

Claims (8)

1, a kind of audio frequency splitting method that changes detection based on decision tree and speaker, it is characterized in that: at first utilize adaptive silence detection to find out quiet in the audio frequency, and utilize these quiet audio frequency to be carried out coarse segmentation, detect to segment according to sudden change then and cut, and come to carry out the classification of speech/non-speech to cutting apart the audio fragment that obtains with decision tree, the detection speaker changes a little between sound bite at last, is changed by the speaker a little to obtain final segmentation result.

2, the audio frequency splitting method that changes detection based on decision tree and speaker according to claim 1 is characterized in that: comprise the steps:

1), audio frequency is carried out pre-service: the audio frequency pre-service is divided into sample quantization, zero-suppresses and floats, three parts of pre-emphasis and windowing;

2), audio feature extraction: the feature extraction on the audio frame comprises energy, the extraction of zero-crossing rate and Mel cepstrum coefficient;

3), silence detection: audio frequency divided calculate each frame energy behind the frame, determine quiet by adaptive energy threshold value and time threshold;

4), voice cut apart: after the silence detection, sound signal is divided into continuous quiet section and non-quiet section, and length non-quiet section greater than 10 seconds further cut apart; It is to determine catastrophe point by the distance between the distribution of calculating energy and zero-crossing rate that sudden change detects;

5), speech/non-speech classification: come non-quiet section classification: voice or non-voice, the corresponding section feature of each node of decision tree with the decision tree that trains; Decision tree is the decision rule of one group of precondition, the section feature and the rule of correspondence of audio-frequency fragments is judged successively, by the final value decision audio fragment type of decision tree;

6), the speaker changes detection: detect speaker's change with the similarity between the characteristic distribution, be about to distance between the adjacent voice segments and adaptive threshold relatively, determine that the speaker changes a little.

3, according to claim 2ly change the audio frequency splitting method of detection based on decision tree and speaker, it is characterized in that: described audio frequency pre-service concrete steps are:

1), sample quantization:

A), sound signal is carried out filtering, make its nyquist frequency F with sharp filter _NBe 4KHZ;

B), audio sample rate F=2F is set _N

C), to sound signal s _a(t) sample by the cycle, obtain the amplitude sequence of digital audio and video signals $s\left(n\right)=s_a\left(\frac{n}{F}\right);$

D), s (n) is carried out quantization encoding, the quantization means s ' that obtains amplitude sequence (n) with pulse code modulation (pcm);

2), zero-suppress and float:

A), calculate the mean value s of the amplitude sequence that quantizes;

B), each amplitude is deducted mean value, obtain zero-suppressing that to float back mean value be 0 amplitude sequence s " (n):

3), pre-emphasis:

A), Z transfer function H (the z)=1-α z of digital filter is set ^-1In pre emphasis factor α, α desirable 1 or slightly little value than 1;

B), s " (n) by digital filter, obtains the suitable amplitude sequence s (n) of high, medium and low frequency amplitude of sound signal;

4), windowing:

A), calculate frame length N (32 milliseconds) and the frame amount of the moving T (10 milliseconds) of audio frame, satisfied respectively:

$\frac{N}{F}=0.032$

$\frac{T}{F}=0.010$

Here F is an audio sample rate, and unit is Hz;

B), be that N, the frame amount of moving are T with the frame length, s (n) is divided into a series of audio frame F _m, each audio frame comprises N audio signal samples;

C), calculate the hamming code window function:

D), to each audio frame F _mAdd hamming code window:

ω(n)×F _m(n){F _m′(n)|n＝0，1，…，N-1}。

4, according to claim 2ly change the audio frequency splitting method of detection based on decision tree and speaker, it is characterized in that: the concrete steps of described audio feature extraction are:

1), the extraction of energy:

$E=\underset{n=1}{\overset{N}{\mathrm{\Σ}}}{s}^2\left(n\right)$

2), the extraction of zero-crossing rate:

$\mathrm{Zcr}=\frac{1}{2N-1}\underset{n=1}{\overset{N-1}{\mathrm{\Σ}}}\left|\right[\mathrm{sgn}\left(s(n+1)\right)-\mathrm{sgn}\left(s\left(n\right)\right)\left]\right|$

3), Mel cepstrum coefficient, the i.e. extraction of MFCC:

A), the exponent number p of Mel cepstrum coefficient is set;

B), be fast fourier transform FFT, time-domain signal s (n) is become frequency domain signal X (k);

C), calculate Mel territory scale:

$M_i=\frac{i}{p}\×2595\log (1+\frac{8000/2.0}{700.0}),(i=\mathrm{0,1,2},...,p)$

D), calculate corresponding frequency domain scale:

$f_i=700\×e^{\frac{M_i}{2595}\ln 10}-1,$ (i＝0，1，2，...，p)

E), calculate each Mel territory passage φ _jOn the logarithm energy spectrum:

$E_j=\underset{k=0}{\overset{\frac{K}{2}-1}{\mathrm{\Σ}}}{\mathrm{\φ}}_j\left(k\right){\left|X\left(k\right)\right|}^2$

Wherein $\underset{k=0}{\overset{\frac{K}{2}-1}{\mathrm{\Σ}}}{\mathrm{\φ}}_j\left(k\right)=1.$

F), be discrete cosine transform DCT.

5, according to claim 2ly change the audio frequency splitting method of detection based on decision tree and speaker, it is characterized in that: described silence detection concrete steps are:

A, calculating adaptive energy threshold value:

Threshold(E)＝min(E)+0.3×[mean(E)-min(E)]

Wherein, Threshold (E) is the adaptive energy threshold value, and min (E) is the minimum value of each frame energy, and mean (E) is the mean value of each frame energy;

B, quiet section detection: the energy and the energy threshold T of each audio frame are compared, and the frame that is lower than threshold value is quiet frame, and continuous quiet frame is formed one quiet section;

The calculating of C, zero-crossing rate threshold value:

Threshold(Zcr)＝0.5×mean(Zcr _i)，i∈{i|E _i＜Threshold(E)}

Wherein, Threshold (Zcr) is a self-adaptation zero-crossing rate threshold value, mean (Zcr _i) be the zero-crossing rate mean value of quiet frame;

D, quiet section zero-crossing rate correction: check each frame zero-crossing rate successively from each two ends of quiet section,, then be considered as the initial or voiceless sound when finishing of syllable, shift out quiet section if be higher than threshold value;

E, smoothing processing: be lower than 10 frames, promptly 0.1 second quiet section is regarded as the pause in short-term between continuous speech and casts out.

6, according to claim 2ly change the audio frequency splitting method of detection based on decision tree and speaker, it is characterized in that: described voice are cut apart concrete steps and are:

The parameter estimation that A, energy and zero-crossing rate distribute: in non-quiet section of need further cut apart, be that window is long with 50 frames, 10 frames are the window step-length, calculate the energy and the zero-crossing rate x of 50 frames in each window ²The parameter that distributes:

$p\left(x\right)=\frac{x^a{e}^{-\mathrm{bx}}}{b^{a+1}\mathrm{\Γ}(a+1)},(x\≥0)$

Wherein $a=\frac{{\mathrm{\μ}}^2}{{\mathrm{\σ}}^2}-1,b=\frac{{\mathrm{\σ}}^2}{\mathrm{\μ}},$ μ is a mean value, and σ is a variance;

B, distance calculation:

The distance definition of each window is as follows:

$D\left(i\right)=1-\frac{\mathrm{\Γ}(\frac{a_{i-1}+a_{i+1}}{2}+1)}{\sqrt{\mathrm{\Γ}(a_{i-1}+1)\mathrm{\Γ}(a_{i+1}+1)}}\frac{2^{\frac{a_{i-1}+a_{i+1}}{2}+1}{b}_{i-1}^{\frac{a_{i+1}+1}{2}}{b}_{i+1}^{\frac{a_{i-1}+1}{2}}}{{(b_{i-1}+b_{i+1})}^{\frac{a_{i-1}+a_{i+1}}{2}+1}}$

a _I-1, a _I+1, b _I-1, b _I+1The parameter a of window before and after being respectively, b;

C, sudden change detect: in each has the window of maximum value distance D (i), once more each frame is calculated same distance, getting the maximum frame of distance is cut-point.

7, according to claim 2ly change the audio frequency splitting method of detection based on decision tree and speaker, it is characterized in that: section feature selected in step 5) is as follows:

A), high zero-crossing rate ratio HZCRR:

$\mathrm{HZCRR}=\frac{1}{2N}\underset{n=0}{\overset{N-1}{\mathrm{\Σ}}}[\mathrm{sgn}(\mathrm{Zcr}\left(n\right)-1.5\×\stackrel{\‾}{\mathrm{Zcr}})+1]$

The relative noise of HZCRR distribution center of voice segments and music etc. are higher, and HZCRR is higher than f _hSection be regarded as voice segments;

B), low-yield ratio LRMSR:

$\mathrm{LRMSR}=\frac{1}{2N}\underset{n=0}{\overset{N-1}{\mathrm{\Σ}}}[\mathrm{sgn}(0.5\×\stackrel{\‾}{E}-E\left(n\right)+1]$

The relative noise of LRMSR distribution center of voice segments and music etc. are lower, and LRMSR is lower than f _lSection be regarded as voice segments;

C), fundamental frequency Mean F:

The fundamental frequency of audio section is estimated with zero-crossing rate:

MeanF＝max[Zcr(n)]×F/N，

Wherein F is a sample frequency 8000, and N is frame length 32ms, and fundamental frequency is higher than f _fSection be regarded as non-speech segment;

D), no zero-crossing rate space-number NZCRR:

The number of times that zero-crossing rate is zero frame appears in the definition section of being of NZCRR the inside, and continuous no zero passage frame is only calculated once, and NZCRR is lower than f _nSection be regarded as voice segments;

E), energy variance Var RMS

The variance of the definition section of the being self-energy of Var RMS, variance is less than f _vSection be regarded as voice segments;

F), f _h, f _l, f _f, f _n, f _vValue all obtain by decision tree training.

8, the audio frequency splitting method that changes detection based on decision tree and speaker according to claim 2 is characterized in that: use T in step 6) ²Distance is calculated the MFCC characteristic distance between each voice segments;

1), T ²Distance definition is as follows:

$T^2=\frac{\mathrm{ab}}{a+b}{({\mathrm{\μ}}_1-{\mathrm{\μ}}_2)}^T{\mathrm{\Σ}}^{-1}({\mathrm{\μ}}_1-{\mathrm{\μ}}_2)$

A wherein, the length of the b section of being, μ ₁, μ ₂Be the mean value of MFCC in each section, ∑ is common covariance matrix;

2), adaptive threshold calculates

By comparing T ²Whether distance and threshold value can detect and exist the speaker to change, and the computing formula of adaptive threshold is as follows:

T＝μ+λσ

Wherein μ is a mean distance, and σ is a distance variance, and λ is a penalty coefficient;

3), merge:

If the distance between two voice segments is less than threshold value, these two voice segments are regarded as belonging to same speaker so, these two voice segments can be merged into one; If exist quietly between these two voice segments, this section is quiet so also will merge; If have non-voice between two voice segments, then nonjoinder.

Images