CN102034472A - Speaker recognition method based on Gaussian mixture model embedded with time delay neural network - Google Patents

Abstract

The invention discloses a speaker recognition method based on a Gaussian mixture model (GMM) embedded with a time delay neural network (TDNN). In the speaker recognition method, the advantages of the TDNN and the GMM are fully considered, the TDNN is embedded into the GMM, and solves a residual of input and output vectors of the TDNN by fully utilizing the time sequence of an input characteristic vector through the conversion of a time delay network, and the residual modifies the training of the GMM through an expectation maximization method; besides, a likelihood probability is acquired by a modified GMM model parameter and the residual, and a TDNN parameter is modified by an inertial backward inversion method so as to ensure that parameters of the GMM and the TDNN are alternately updated. An experiment shows that: a recognition rate of the method is improved to a certain extent compared with that of a baseline GMM under various signal to noise ratios.

Description

A kind of method for distinguishing speek person based on the gauss hybrid models that embeds time-delay neural network

Technical field

The present invention relates to a kind of method for distinguishing speek person, particularly a kind of method for distinguishing speek person based on the gauss hybrid models that embeds time-delay neural network.

Background technology

At aspects such as gate inhibition, credit card trade and court evidences, automatic Speaker Identification, particularly important with the Speaker Identification play more and more of text-independent effect, its target are voice to be identified correctly to be judged to be belong in the sound bank some among a plurality of reference men.

On the method for Speaker Identification, more and more come into one's own based on gauss hybrid models (GMM) method, because it has the discrimination height, training is simple, amount of training data requires advantages such as little, has become the recognition methods of present main flow.Because gauss hybrid models (GMM) has the ability of the distribution of good expression data, as long as abundant item is arranged, abundant training data, GMM just can approach any distributed model.But there are several problems in reality when using GMM.At first, GMM does not utilize the temporal information of speaker's voice, and the result of training and identification and the input sequence of proper vector are irrelevant; Secondly, when GMM trained, we always supposed that proper vector is mutually independently, and this is obviously unreasonable; In addition, because we are when selecting the GMM model, the choosing of mixed term number not have the governing principle of getting well yet, and the result that obtain just requires the Gaussian Mixture item abundant.

Neural network is also occupied important position aspect Speaker Identification, multilayer perceptron, ray base net network and auto-associative neural network etc. have been successfully applied to Speaker Identification, especially time-delay neural network (TDNN) is used widely in signal Processing, speech recognition and Speaker Identification, it has made full use of the time sequence information of characteristic vector sequence, proper vector is learnt and conversion, made proper vector after the conversion (be generally minimum least square method) in some way and approach object vector.But GMM and TDNN just are used for Speaker Identification separately at present, also do not occur in conjunction with the two advantage separately, thereby the method that improves the Speaker Identification effect better occurs.

Summary of the invention

Purpose of the present invention just is to address the deficiencies of the prior art, and has proposed a kind of method for distinguishing speek person based on the gauss hybrid models that embeds time-delay neural network.Technical scheme of the present invention is:

A kind of method for distinguishing speek person based on the gauss hybrid models that embeds time-delay neural network, it may further comprise the steps:

(1) pre-service and feature extraction;

At first, used based on the method for energy and zero-crossing rate and carried out silence detection, and removed noise with spectrum-subtraction, and voice signal carried out pre-emphasis, divide frame, the line linearity of going forward side by side prediction (LPC) is analyzed, and obtains the proper vector of cepstrum coefficient as Speaker Identification then from the LPC coefficient that obtains.

(2) training;

During training, the proper vector process that extracts is postponed the input of back as TDNN, the structure of TDNN learning characteristic vector, the temporal information of extraction characteristic vector sequence.Then learning outcome is offered GMM with the form of residual error proper vector, adopt greatest hope (EM) criterion to carry out the GMM model training, and utilize the inversion method backward of band inertia to upgrade the weight coefficient of TDNN network.Concrete training process is as follows:

(2-1) determine GMM model and TDNN structure:

The probability density function of a M rank GMM is obtained by M Gaussian probability-density function weighted sum, can represent with following form:

$p\left(x_t\right|\mathrm{\λ})=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{p}_i{b}_i\left(x_t\right)$

X in the following formula _tBe D dimensional feature vector, D=13 here; b _i(x _t) be member's density function, it is u for mean value vector _i, covariance matrix is a ∑ _iGaussian function;

$b_i\left(x_t\right)=\frac{1}{{\left(2\mathrm{\π}\right)}^{D/2}{\left|{\mathrm{\Σ}}_i\right|}^{1/2}}\exp \{-\frac{1}{2}{(x_i-u_i)}^T{\mathrm{\Σ}}_i^{-1}(x_t-u_i)\}$

p _iBe that the mixed weight-value mixed weight-value satisfies condition:

Complete GMM model parameter is as follows:

λ＝{(p _i，u _i，∑ _i)，i＝1，2，...，M}

Here, that utilization is the TDNN that is not with feedback.After the delay of proper vector x (n) through the linear delay piece, as the input of TDNN, TDNN carries out nonlinear transformation to input, and linear weighted function obtains output vector then, compares with proper vector again, and normally used criterion is lowest mean square criterion (MMSE).Specifically, the ratio of the neuron number of the hidden layer of TDNN and the neuron number of input layer is 3: 2, and non-linear activation S function (" ∫ " mark among Fig. 2) is

Y is through the input after the weighted sum.During training, inertial coefficient γ=0.8 of neural network.

(2-2) set the condition of convergence and maximum iteration time; Particularly, the condition of convergence be the Euclidean distance of adjacent twice GMM coefficient and TDNN weight coefficient less than 0.0001, maximum iteration time is not more than 100 usually.

(2-3) determine the TDNN and the GMM model parameter of primary iteration at random; The initial coefficients of TDNN is set at the pseudo random number that is produced by computing machine, the initial mixing coefficient of GMM can be taken as 1/M, M is the mixing item number of GMM, initial average of GMM and variance are by the residual vector process LBG (Linde of TDNN, Buzo, Gray) method produces M polymeric type, and average and the variance of calculating this M polymeric type respectively obtain.

(2-4) proper vector x (n) input TDNN network, will subtract each other by proper vector x (n) before the TDNN and the output characteristic vector o (n) of TDNN, obtain all residual vectors;

(2-5) parameter of employing EM method correction GMM model;

If residual vector is r _t, at first calculate the classification posterior probability:

$p\left(i\right|r_t,\mathrm{\λ})=\frac{p_i{b}_i\left(r_t\right)}{{\mathrm{\Σ}}_{k=1}^M{p}_k{b}_k\left(r_t\right)}$

Upgrade mixed weight-value then Mean value vector

And covariance matrix

$\stackrel{\‾}{p_i}=\frac{1}{N}\underset{t=1}{\overset{N}{\mathrm{\Σ}}}p({i|r}_t,\mathrm{\λ})$

$\stackrel{\‾}{u_i}=\frac{{\mathrm{\Σ}}_{t=1}^{N}p({i|r}_t,\mathrm{\λ})x_t}{{\mathrm{\Σ}}_{t=1}^{N}p({i|r}_t,\mathrm{\λ})}$

${\stackrel{\‾}{\mathrm{\Σ}}}_i^2=\frac{{\mathrm{\Σ}}_{t=1}^{N}p({i|r}_t,\mathrm{\λ})x_t^2}{{\mathrm{\Σ}}_{t=1}^{N}p({i|r}_t,\mathrm{\λ})}-{\stackrel{\‾}{u_i}}^2$

(2-6) utilize the weight coefficient of revised each Gaussian distribution of GMM model, mean vector and variance are brought residual error into, obtain a likelihood probability, utilize the correction of the inversion method backward TDNN parameter of band inertia;

The TDNN network parameter obtains by making the function maximization in the following formula:

$L\left(X\right)=\underset{{\mathrm{\ω}}_{\mathrm{ij}}}{\arg \max }\underset{t=1}{\overset{N}{\mathrm{\Π}}}p\left((x_t-o_t)\right|\mathrm{\λ})$

O wherein _tBe neural network output, x _tEigenvector for input.

Get negatively after following formula taken the logarithm again, obtain:

$G\left(X\right)=\underset{{\mathrm{\ω}}_{\mathrm{ij}}}{\arg \min }(-\underset{t=1}{\overset{N}{\mathrm{\Σ}}}\ln p\left((x_t-o_t)\right|\mathrm{\λ}))$

Adopt the inversion method backward of band inertia to ask its iterative formula of G (X) as follows:

${\mathrm{\Δ\ω}}_{\mathrm{ij}}^k(m+1)={\mathrm{\γ\Δ\ω}}_{\mathrm{ij}}^k\left(m\right)-(1-\mathrm{\γ})\mathrm{\α}\frac{\∂F\left(x\right)}{\∂{\mathrm{\ω}}_{\mathrm{ij}}^k}{|}_{{\mathrm{\ω}}_{\mathrm{ij}}^k={\mathrm{\ω}}_{\mathrm{ij}}^k\left(m\right)}$

Wherein,

Be in the m time iteration, connect input x _iWith output y _jWeight coefficient, k is the layer sequence number of neural network, α is an iteration step length, F (x)=-lnp ((x _t-o _t) | λ), γ is an inertial coefficient.

(2-7) judge whether to satisfy the condition of convergence of setting in the step (2-2) or whether reach maximum iteration time, if, then stop training, otherwise, skip to step (2-4).

(3) Speaker Identification

During identification, characteristic vector sequence X is through postponing back input TDNN.Then the output sequence O of X and TDNN is subtracted each other resulting residual sequence R and offer the GMM model, for the sequence R=R of T residual error vector ₁, R ₂..., R _T, its GMM probability can be written as:

$P\left(R\right|\mathrm{\λ})=\underset{t=1}{\overset{T}{\mathrm{\Π}}}p\left(R_t\right|\mathrm{\λ})$

Be expressed as at log-domain:

$L\left(R\right|\mathrm{\λ})=\log P\left(R\right|\mathrm{\λ})=\underset{t=1}{\overset{T}{\mathrm{\Σ}}}\log p\left(R_t\right|\mathrm{\λ})$

Utilization Bayes' theorem during identification, in N unknown words person's model, the words person of the model correspondence of likelihood probability maximum is the target speaker:

$i^{*}=\underset{1\≤i\≤N}{\arg \max }L\left({R|\mathrm{\λ}}_i\right)$

In described a kind of method for distinguishing speek person, described based on the gauss hybrid models that embeds time-delay neural network

Computation process as follows:

$\frac{\∂F\left(x\right)}{\∂{\mathrm{\ω}}_{\mathrm{ij}}^k}=\frac{\∂F\left(x\right)}{{\∂y}_i^k}\frac{{\∂y}_i^k}{{\∂\mathrm{\ω}}_{\mathrm{ij}}^k}$

In the TDNN network,

Output when being i neuron input of k layer sample x,

Input when being i neuron input of k layer sample x,

Be activation function.So:

$\frac{{\∂y}_i^k}{{\∂\mathrm{\ω}}_{\mathrm{ij}}^k}=o_j^{k-1}$

In described a kind of method for distinguishing speek person, described based on the gauss hybrid models that embeds time-delay neural network Computation process is divided into output layer and two kinds of situations of hidden layer of TDNN;

For output layer:

$\frac{\∂F\left(x\right)}{{\∂y}_i^k}=-\frac{1}{p\left((x-o)\right|\mathrm{\λ})}\frac{\∂p\left((x-o)\right|\mathrm{\λ})}{{\∂o}_i^k}\frac{{\∂o}_i^k}{y_i^k}$

$=-\frac{f^{\′}\left(y_i^k\right)}{p\left((x-o)\right|\mathrm{\λ})}\∂\left(\underset{n=1}{\overset{M}{\mathrm{\Σ}}}{p}_n{c}_n{e}^{-\frac{1}{2}{\left({x-o-u}_n\right)}^T{\mathrm{\Σ}}_n^{-1}\left({x-o-u}_n\right)}\right)/{\∂o}_i^k$

$=-\frac{f^{\′}\left(y_i^k\right)}{p\left((x-o)\right|\mathrm{\λ})}\underset{n=1}{\overset{M}{\mathrm{\Σ}}}{p}_n{c}_n\left(\frac{a_n\left({x-o-u}_n\right)}{{\mathrm{\σ}}_{n,i}^2}(x_i-o_i-u_{n,i})\right)---\left(15\right)$

Wherein: $a_n\left({x-o-u}_n\right)=e^{-\frac{1}{2}{\left({x-o-u}_n\right)}^T{\mathrm{\Σ}}_n^{-1}\left({x-o-u}_n\right)},$ $c_n=\frac{1}{{\left(2\mathrm{\π}\right)}^{D/2}{\left|{\mathrm{\Σ}}_n\right|}^{1/2}}$

For hidden layer:

$\frac{\∂F\left(x\right)}{{\∂y}_i^k}=\underset{j}{\mathrm{\Σ}}\frac{\∂F\left(x\right)}{\∂y_j^{k+1}}\frac{{\∂y}_i^{k+1}}{\∂y_i^k}=\underset{j}{\mathrm{\Σ}}\frac{\∂F\left(x\right)}{\∂y_j^{k+1}}\frac{\∂\left(\underset{n}{\mathrm{\Σ}}{\mathrm{\ω}}_{\mathrm{jn}}^{k+1}{o}_n^k\right)}{\∂y_i^k}$

$=\underset{j}{\mathrm{\Σ}}\frac{\∂F\left(x\right)}{\∂y_i^{k+1}}\frac{{\∂o}_i^k}{{\∂y}_i^k}{\mathrm{\ω}}_{\mathrm{ji}}^{k+1}=f^{\′}\left(y_i^k\right)\underset{j}{\mathrm{\Σ}}\frac{\∂F\left(x\right)}{\∂y_i^{k+1}}{\mathrm{\ω}}_{\mathrm{ji}}^{k+1}.$

In described a kind of method for distinguishing speek person based on the gauss hybrid models that embeds time-delay neural network, described pre-emphasis adopts f (Z)=1-0.97Z ^-1Wave filter, divide frame to adopt length 20ms and window to move the Hamming window of 10ms, the exponent number of linear prediction analysis is 20, proper vector is the cepstrum coefficient on 13 rank.

Advantage of the present invention and effect are:

1. made full use of TDNN and GMM advantage separately, make the temporal information that TDNN can the learning characteristic vector like this, set of eigenvectors is mapped to the subspace that can increase likelihood probability, and can reduce the independently influence of this unreasonable hypothesis of proper vector, likelihood probability to strengthening object module reduces the effect of the likelihood probability of non-object module.And GMM has the discrimination height, training is simple and the little advantage of amount of training data requirement.Institute is so that whole Speaker Recognition System discrimination improves greatly.

2. the method that proposes of the present invention Speaker Identification effect under voice and the noise circumstance voice when noiseless all increases than independent employing GMM.

Other advantages of the present invention and effect will continue to describe below.

Description of drawings

Fig. 1---speaker's training and model of cognition.

Fig. 2---time-delay neural network model.

Fig. 3---comparing data during 1conv4w-1conv4w during noiseless.

Comparing data during noise in Fig. 4---the automobile.

Fig. 5---the comparing data when showing the noise in the compartment.

Embodiment

Below in conjunction with drawings and Examples, technical solutions according to the invention are further elaborated.

Fig. 1 is training and the model of cognition that embeds the both speaker. identification of TDNN network, and it follows baseline GMM model (only adopting the GMM model as Speaker Identification) all different aspect training and identification.

1. pre-service and feature extraction

At first, used based on the method for energy and zero-crossing rate and carried out silence detection, and removed noise, again by f (Z)=1-0.97Z with spectrum-subtraction ^-1Wave filter carry out pre-emphasis, the Hamming window that moves 10ms with length 20ms and window divides frame, carries out 20 rank linear predictions (LPC) and analyzes, and obtains the proper vector of the cepstrum coefficient on 13 rank as Speaker Identification then from 20 rank LPC coefficients.

2. speaker model training

During training, the process of the process of training TDNN and training GMM model hockets.TDNN is a kind of multilayer perceptron network (MLP), as shown in Figure 2.Proper vector through the delay of linear delay piece after as the input of TDNN, the structure of TDNN learning characteristic vector is extracted the temporal information of characteristic vector sequence.Then learning outcome is offered GMM with the form of residual error proper vector (being output poor of input vector and TDNN), adopt greatest hope (EM) method to carry out the GMM model training, and utilize the weight coefficient of the renewal of the inversion method backward TDNN network of band inertia.Here the criterion of TDNN and GMM model learning and training all is the maximum likelihood probability.Like this, by study, residual error distributes and just might carry out towards the direction that strengthens likelihood probability.Concrete training process is described below:

(1) determine GMM model and TDNN structure:

The probability density function of a M rank GMM is obtained by M Gaussian probability-density function weighted sum, can represent with following form:

$p\left(x_t\right|\mathrm{\λ})=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{p}_i{b}_i\left(x_t\right)---\left(1\right)$

Here x _tBe a D dimension random vector, in Speaker Identification is used, x _tBe proper vector; b _i(x _t), i=1,2 ..., M is member's density; p _i, i=1,2 ..., M is a mixed weight-value.Each member's density function is a D dimension variable, and mean value vector is u _i, covariance matrix is a ∑ _iGaussian function, form is as follows:

$b_i\left(x_t\right)=\frac{1}{{\left(2\mathrm{\π}\right)}^{D/2}{\left|{\mathrm{\Σ}}_i\right|}^{1/2}}\exp \{-\frac{1}{2}{(x_i-u_i)}^T{\mathrm{\Σ}}_i^{-1}(x_t-u_i)\}---\left(2\right)$

Wherein mixed weight-value satisfies condition: $\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{p}_i=1.$

Complete GMM model is by mean value vector, covariance matrix and the mixed weight-value parameter of all member's density.These parameters gatherings are expressed as together:

λ＝{(p _i，u _i，∑ _i)，i＝1，2，...，M} (3)

Because Speaker Identification general training and recognition data are less, be diagonal matrix so set the covariance matrix of each Gaussian Mixture Model Probability Density in actual the use usually.

Here, utilization be the TDNN that is not with feedback, as shown in Figure 2.After the delay of proper vector x (n) through the linear delay piece, as the input of TDNN, TDNN carries out nonlinear transformation to input, and linear weighted function obtains output vector then, compares with proper vector again, and normally used criterion is lowest mean square criterion (MMSE).Specifically, the ratio of the neuron number of the hidden layer of TDNN and the neuron number of input layer is 3: 2, and non-linear activation S function (" ∫ " mark among Fig. 2) is

Y is through the input after the weighted sum.

When training, inertial coefficient γ=0.8 of neural network.

(2) set the condition of convergence and maximum iteration time; Particularly, the condition of convergence be the Euclidean distance of adjacent twice GMM coefficient and TDNN weight coefficient less than 0.0001, maximum iteration time is not more than 100 usually.

(3) determine the TDNN and the GMM model parameter of primary iteration at random; The initial coefficients of TDNN is set at the pseudo random number that is produced by computing machine, the initial mixing coefficient of GMM can be taken as 1/M, M is the mixing item number of GMM, initial average of GMM and variance are by the residual vector process LBG (Linde of TDNN, Buzo, Gray) method produces M polymeric type, and average and the variance of calculating this M polymeric type respectively obtain.

(4) proper vector x (n) input TDNN network, will subtract each other by proper vector x (n) before the TDNN and the output characteristic vector o (n) of TDNN, obtain all residual vectors;

(5) parameter of employing EM method correction GMM model;

If residual vector is r _t, at first use formula (4) to calculate the classification posterior probability:

$p\left(i\right|r_t,\mathrm{\λ})=\frac{p_i{b}_i\left(r_t\right)}{{\mathrm{\Σ}}_{k=1}^M{p}_k{b}_k\left(r_t\right)}---\left(4\right)$

Use formula (5) (6) (7) to obtain the mixed weight-value that upgrades then

Mean value vector

And covariance matrix

$\stackrel{\‾}{p_i}=\frac{1}{N}\underset{t=1}{\overset{N}{\mathrm{\Σ}}}p({i|r}_t,\mathrm{\λ})---\left(5\right)$

$\stackrel{\‾}{u_i}=\frac{{\mathrm{\Σ}}_{t=1}^{N}p({i|r}_t,\mathrm{\λ})x_t}{{\mathrm{\Σ}}_{t=1}^{N}p({i|r}_t,\mathrm{\λ})}---\left(6\right)$

${\stackrel{\‾}{\mathrm{\Σ}}}_i^2=\frac{{\mathrm{\Σ}}_{t=1}^{N}p({i|r}_t,\mathrm{\λ})x_t^2}{{\mathrm{\Σ}}_{t=1}^{N}p({i|r}_t,\mathrm{\λ})}-{\stackrel{\‾}{u_i}}^2---\left(7\right)$

(6) utilize the weight coefficient of revised each Gaussian distribution of GMM model, mean vector and variance are brought residual error into, obtain a likelihood probability, utilize the parameter lambda of the correction of the inversion method backward TDNN of band inertia;

Specifically, by making function maximization in the following formula obtain the parameter lambda of TDNN:

$L\left(X\right)=\underset{{\mathrm{\ω}}_{\mathrm{ij}}}{\arg \max }\underset{t=1}{\overset{N}{\mathrm{\Π}}}p\left((x_t-o_t)\right|\mathrm{\λ})---\left(8\right)$

Wherein p (x| λ) sees formula (1), o _tBe the output vector of TDNN, x _tEigenvector for input TDNN.

Because general minimizing during the neural network iteration, and and formula more more convenient than product, so get negatively after following formula taken the logarithm again, obtain:

$G\left(X\right)=\underset{{\mathrm{\ω}}_{\mathrm{ij}}}{\arg \min }(-\underset{t=1}{\overset{N}{\mathrm{\Σ}}}\ln p\left((x_t-o_t)\right|\mathrm{\λ}))---\left(9\right)$

Here, parameter lambda is embodied in

Be in the m time iteration, connect input x _iWith output y _jWeight coefficient, k is the layer sequence number of neural network.Adopt the inversion method backward of band inertia to ask G (X), because this method can be quickened repeatedly convergence process, and can better handle the local minimum problem, the iterative formula of the inversion method backward of band inertia is as follows:

${\mathrm{\Δ\ω}}_{\mathrm{ij}}^k(m+1)={\mathrm{\γ\Δ\ω}}_{\mathrm{ij}}^k\left(m\right)-(1-\mathrm{\γ})\mathrm{\α}\frac{\∂F\left(x\right)}{\∂{\mathrm{\ω}}_{\mathrm{ij}}^k}{|}_{{\mathrm{\ω}}_{\mathrm{ij}}^k={\mathrm{\ω}}_{\mathrm{ij}}^k\left(m\right)}---\left(10\right)$

Wherein,

α is an iteration step length, F (x)=-lnp ((x _t-o _t) | λ), γ is an inertial coefficient, and γ gets 0.8 herein.

In the formula (10),

Computation process as follows:

$\frac{\∂F\left(x\right)}{\∂{\mathrm{\ω}}_{\mathrm{ij}}^k}=\frac{\∂F\left(x\right)}{{\∂y}_i^k}\frac{{\∂y}_i^k}{{\∂\mathrm{\ω}}_{\mathrm{ij}}^k}---\left(11\right)$

Seek the computing formula of two product terms in the formula (11) below respectively, because in neural network:

$y_i^k=\underset{j}{\mathrm{\Σ}}{\mathrm{\ω}}_{\mathrm{ij}}^k{o}_j^{k-1}---\left(12\right)$

$o_i^k=f\left(y_i^k\right)---\left(13\right)$

In last two formulas,

Output when being i neuron input of k layer sample x,

Input when being i neuron input of k layer sample x,

Be activation function.So:

$\frac{\∂y_i^k}{{\∂\mathrm{\ω}}_{\mathrm{ij}}^k}=o_j^{k-1}---\left(14\right)$

For

Find the solution, be divided into two kinds of situations of output layer and hidden layer:

(a) output layer

$\frac{\∂F\left(x\right)}{{\∂y}_i^k}=-\frac{1}{p\left((x-o)\right|\mathrm{\λ})}\frac{\∂p\left((x-o)\right|\mathrm{\λ})}{{\∂o}_i^k}\frac{{\∂o}_i^k}{y_i^k}$

$=-\frac{f^{\′}\left(y_i^k\right)}{p\left((x-o)\right|\mathrm{\λ})}\∂\left(\underset{n=1}{\overset{M}{\mathrm{\Σ}}}{p}_n{c}_n{e}^{-\frac{1}{2}{\left({x-o-u}_n\right)}^T{\mathrm{\Σ}}_n^{-1}\left({x-o-u}_n\right)}\right)/{\∂o}_i^k$

$=-\frac{f^{\′}\left(y_i^k\right)}{p\left((x-o)\right|\mathrm{\λ})}\underset{n=1}{\overset{M}{\mathrm{\Σ}}}{p}_n{c}_n\left(\frac{a_n\left({x-o-u}_n\right)}{{\mathrm{\σ}}_{n,i}^2}(x_i-o_i-u_{n,i})\right)---\left(15\right)$

Wherein:

$a_n\left({x-o-u}_n\right)=e^{-\frac{1}{2}{\left({x-o-u}_n\right)}^T{\mathrm{\Σ}}_n^{-1}\left({x-o-u}_n\right)}$

$c_n=\frac{1}{{\left(2\mathrm{\π}\right)}^{D/2}{\left|{\mathrm{\Σ}}_n\right|}^{1/2}}$

(b) hidden layer

$\frac{\∂F\left(x\right)}{{\∂y}_i^k}=\underset{j}{\mathrm{\Σ}}\frac{\∂F\left(x\right)}{\∂y_j^{k+1}}\frac{{\∂y}_j^{k+1}}{\∂y_i^k}$

$=\underset{j}{\mathrm{\Σ}}\frac{\∂F\left(x\right)}{\∂y_j^{k+1}}\frac{\∂\left(\underset{n}{\mathrm{\Σ}}{\mathrm{\ω}}_{\mathrm{jn}}^{k+1}{o}_n^k\right)}{\∂y_i^k}$

$=\underset{j}{\mathrm{\Σ}}\frac{\∂F\left(x\right)}{\∂y_j^{k+1}}\frac{{\∂o}_i^k}{{\∂y}_i^k}{\mathrm{\ω}}_{\mathrm{ji}}^{k+1}$

${=f}^{\′}\left(y_i^k\right)\underset{j}{\mathrm{\Σ}}\frac{\∂F\left(x\right)}{\∂y_i^{k+1}}{\mathrm{\ω}}_{\mathrm{ji}}^{k+1}---\left(16\right)$

Because what adopt is inverting backward, so calculating

The time

Known, substitution formula (16) can be obtained

(7) judge whether to satisfy the condition of convergence of setting in the step (2) or whether reach maximum iteration time, if, then stop training, otherwise, skip to step (4).

3. Speaker Identification

During identification, proper vector enters TDNN after postponing.Because TDNN has learnt the structure and the time sequence information of feature space, so import characteristic vector sequence X to be identified, TDNN can do corresponding conversion to proper vector.Because passed through training,, reduce the effect of the likelihood probability of non-object module so the likelihood probability that strengthens object module has been played in the TDNN conversion.Then the sequence O that is exported after X and the process TDNN conversion is subtracted each other resulting residual sequence R and offer the GMM model, for the sequence R=R of T residual error vector ₁, R ₂..., R _T, its GMM probability can be written as:

$P\left(R\right|\mathrm{\λ})=\underset{t=1}{\overset{T}{\mathrm{\Π}}}p\left(R_t\right|\mathrm{\λ})---\left(17\right)$

Be expressed as at log-domain:

$L\left(R\right|\mathrm{\λ})=\log P\left(R\right|\mathrm{\λ})=\underset{t=1}{\overset{T}{\mathrm{\Σ}}}\log p\left(R_t\right|\mathrm{\λ})---\left(18\right)$

Utilization Bayes' theorem during identification, in N unknown words person's model, the words person of the model correspondence of likelihood probability maximum is the target speaker:

$i^{*}=\underset{1\&le;i\&le;N}{\mathrm{<figure-callout\; id="1"\; label="arg\; max"\; filenames="H2009100354240E0000011.png,H2009100354240E0000012.png"\; state="\{\{state\}\}">arg</figure-callout>}\max }L\left({R|\mathrm{\&lambda;}}_i\right)---\left(19\right)$

Here the 1conv4w-1conv4w that adopts NIST test in 2006 is as experiment, and we have selected 107 target speakers therein, and wherein the male sex is 63,44 of women.Everyone chooses about about 2 minutes voice as training utterance, and all the other voice form about 23000 tests like this as tested speech.

Improvement effect for method of the present invention under the test noise environment, the noise data of choosing is noise (stationary noise) and the interior noise (nonstationary noise) of the displaying compartment in the exhibition in the automobile (2000cc group, Ordinary Rd) in the travelling in the NEC association criterion noise data storehouse.These noises are superposeed by certain signal to noise ratio (snr) in the 1conv4w-1conv4w voice, generate the voice that contain noise.

Adopt correct recognition rata as the standard of passing judgment on the Speaker Identification effect, correct_ratio=N _v/ N _tWherein, correct_ratio is a correct recognition rata, N _vBe the testing time of correct identification, N _tTotal testing time.

Here with method of the present invention (representing) with only adopt speaker training and the recognition methods (GMM represents with baseline) of GMM with TDNN-GMM.Experimental result is seen Fig. 3-Fig. 5.1conv4w-1conv4w changed the recognition effect comparing result of the mixing item number M of the Gaussian probability-density function among the GMM down when Fig. 3 was noiseless; Find to embed TDNN from Fig. 3 after, the recognition effect of GMM has improvement really, and it is few more to mix item number M, and improvement effect is obvious more, this be since in the class subclass more after a little while, the results of learning of neural network are better.

Fig. 4 and Fig. 5 have compared the comparing result under different noises, different signal to noise ratio (snr) condition, M=80.As can be seen from Figure 4 and Figure 5, the method that the present invention proposes is relative baseline GMM under different signal to noise ratio (S/N ratio)s, and the Speaker Identification effect all has greatly improved.

Claims (4)

1. method for distinguishing speek person based on the gauss hybrid models that embeds time-delay neural network is characterized in that may further comprise the steps:

(1) pre-service and feature extraction;

At first, used based on the method for energy and zero-crossing rate and carried out silence detection, and removed noise with spectrum-subtraction, and voice signal carried out pre-emphasis, divide frame, the line linearity of going forward side by side prediction (LPC) is analyzed, and obtains the proper vector of cepstrum coefficient as Speaker Identification then from the LPC coefficient that obtains;

(2) training;

During training, through postponing the laggard time-delay neural network (TDNN) of going into, the structure of TDNN learning characteristic vector is extracted the temporal information of characteristic vector sequence with the proper vector that extracts; Then learning outcome is offered gauss hybrid models (GMM) with the form of residual error proper vector, adopt the greatest hope method to carry out the GMM model training, and utilize the inversion method backward of band inertia to upgrade the weight coefficient of TDNN; Concrete training process is as follows:

(2-1) determine GMM model and TDNN structure:

The probability density function of a M rank GMM is obtained by M Gaussian probability-density function weighted sum, can represent with following form:

$p\left(x_t\right|\mathrm{\&lambda;})=\underset{i=1}{\overset{M}{\mathrm{\&Sigma;}}}{p}_i{b}_i\left(x_t\right)$

X in the following formula _tBe D dimensional feature vector, D=13 here; b _i(x _t) be member's density function, it is u for mean value vector _i, covariance matrix is a ∑ _iGaussian function;

$b_i\left(x_t\right)=\frac{1}{{\left(2\mathrm{\&pi;}\right)}^{D/2}{\left|{\mathrm{\&Sigma;}}_i\right|}^{1/2}}\exp \{-\frac{1}{2}{(x_i-u_i)}^T{\mathrm{\&Sigma;}}_i^{-1}(x_t-u_i)\}$

p _iBe that the mixed weight-value mixed weight-value satisfies condition:

Complete GMM model parameter is as follows:

λ＝{(p _i，u _i，∑ _i)，i＝1，2，...，M}

Here, what utilize is after not being with the delay of TDNN proper vector x (n) through the linear delay piece of feedback, input as TDNN, TDNN carries out nonlinear transformation to input, linear weighted function then, obtain output vector, compare with proper vector, normally used criterion is lowest mean square criterion (MMSE); The ratio of the neuron number of the hidden layer of TDNN and the neuron number of input layer is 3: 2, and non-linear activation S function is Y is through the input after the weighted sum; When training, inertial coefficient γ=0.8 of neural network;

(2-2) set the condition of convergence and maximum iteration time; Particularly, the condition of convergence be the Euclidean distance of adjacent twice GMM coefficient and TDNN weight coefficient less than 0.0001, maximum iteration time is not more than 100 usually;

(2-3) determine the TDNN and the GMM model parameter of primary iteration at random; The initial coefficients of TDNN is set at the pseudo random number that is produced by computing machine, the initial mixing coefficient of GMM can be taken as 1/M, M is the mixing item number of GMM, initial average of GMM and variance are by the residual vector process LBG (Linde of TDNN, Buzo, Gray) method produces M polymeric type, and average and the variance of calculating this M polymeric type respectively obtain;

(2-4) proper vector x (n) input TDNN network, will subtract each other by proper vector x (n) before the TDNN and the output characteristic vector o (n) of TDNN, obtain all residual vectors;

(2-5) parameter of employing greatest hope method correction GMM model;

If residual vector is r _t, at first calculate the classification posterior probability:

$p\left(i\right|r_t,\mathrm{\&lambda;})=\frac{p_i{b}_i\left(r_t\right)}{{\mathrm{\&Sigma;}}_{k=1}^M{p}_k{b}_k\left(r_t\right)}$

Upgrade mixed weight-value then

Mean value vector

And covariance matrix

$\stackrel{\&OverBar;}{p_i}=\frac{1}{N}\underset{t=1}{\overset{N}{\mathrm{\&Sigma;}}}p({i|r}_t,\mathrm{\&lambda;})$

$\stackrel{\&OverBar;}{u_i}=\frac{{\mathrm{\&Sigma;}}_{t=1}^{N}p({i|r}_t,\mathrm{\&lambda;})x_t}{{\mathrm{\&Sigma;}}_{t=1}^{N}p({i|r}_t,\mathrm{\&lambda;})}$

${\stackrel{\&OverBar;}{\mathrm{\&Sigma;}}}_i^2=\frac{{\mathrm{\&Sigma;}}_{t=1}^{N}p({i|r}_t,\mathrm{\&lambda;})x_t^2}{{\mathrm{\&Sigma;}}_{t=1}^{N}p({i|r}_t,\mathrm{\&lambda;})}-{\stackrel{\&OverBar;}{u_i}}^2$

(2-6) utilize the weight coefficient of revised each Gaussian distribution of GMM model, mean vector and variance are brought residual error into, obtain a likelihood probability, utilize the correction of the inversion method backward TDNN parameter of band inertia;

The TDNN parameter obtains by making the function maximization in the following formula:

$L\left(X\right)=\underset{{\mathrm{\&omega;}}_{\mathrm{ij}}}{\arg \max }\underset{t=1}{\overset{N}{\mathrm{\&Pi;}}}p\left((x_t-o_t)\right|\mathrm{\&lambda;})$

O wherein _tBe neural network output, x _tEigenvector for input;

Get negatively after following formula taken the logarithm again, obtain:

$G\left(X\right)=\underset{{\mathrm{\&omega;}}_{\mathrm{ij}}}{\arg \min }(-\underset{t=1}{\overset{N}{\mathrm{\&Sigma;}}}\ln p\left((x_t-o_t)\right|\mathrm{\&lambda;}))$

Adopt the inversion method backward of band inertia to ask its iterative formula of G (X) as follows:

${\mathrm{\&Delta;\&omega;}}_{\mathrm{ij}}^k(m+1)={\mathrm{\&gamma;\&Delta;\&omega;}}_{\mathrm{ij}}^k\left(m\right)-(1-\mathrm{\&gamma;})\mathrm{\&alpha;}\frac{\&PartialD;F\left(x\right)}{\&PartialD;{\mathrm{\&omega;}}_{\mathrm{ij}}^k}{|}_{{\mathrm{\&omega;}}_{\mathrm{ij}}^k={\mathrm{\&omega;}}_{\mathrm{ij}}^k\left(m\right)}$

Wherein,

Be in the m time iteration, connect input x _iWith output y _jWeight coefficient, k is the layer sequence number of neural network, α is an iteration step length, F (x)=-lnp ((x _t-o _t) | λ), γ is an inertial coefficient;

(2-7) judge whether to satisfy the condition of convergence of setting in the step (2-2) or whether reach maximum iteration time, if, then stop training, otherwise, skip to step (2-4);

(3) identification

During identification, characteristic vector sequence X is through postponing back input TDNN; Then the output sequence O of X and TDNN is subtracted each other resulting residual sequence R and offer the GMM model, for the sequence R=R of T residual error vector ₁, R ₂..., R _T, its GMM probability can be written as:

$P\left(R\right|\mathrm{\&lambda;})=\underset{t=1}{\overset{T}{\mathrm{\&Pi;}}}p\left(R_t\right|\mathrm{\&lambda;})$

Be expressed as at log-domain:

$L\left(R\right|\mathrm{\&lambda;})=\log P\left(R\right|\mathrm{\&lambda;})=\underset{t=1}{\overset{T}{\mathrm{\&Sigma;}}}\log p\left(R_t\right|\mathrm{\&lambda;})$

Utilization Bayes' theorem during identification, in N unknown words person's model, the words person of the model correspondence of likelihood probability maximum is the target speaker:

$i^{*}=\underset{1\&le;i\&le;N}{\arg \max }L\left({R|\mathrm{\&lambda;}}_i\right).$

2. a kind of method for distinguishing speek person based on the gauss hybrid models that embeds time-delay neural network according to claim 1 is characterized in that, and is described

Computation process is as follows:

$\frac{\&PartialD;F\left(x\right)}{\&PartialD;{\mathrm{\&omega;}}_{\mathrm{ij}}^k}=\frac{\&PartialD;F\left(x\right)}{{\&PartialD;y}_i^k}\frac{{\&PartialD;y}_i^k}{{\&PartialD;\mathrm{\&omega;}}_{\mathrm{ij}}^k}$

In the TDNN network,

Output when being i neuron input of k layer sample x,

Input when being i neuron input of k layer sample x, Be activation function, so:

$\frac{{\&PartialD;y}_i^k}{{\&PartialD;\mathrm{\&omega;}}_{\mathrm{ij}}^k}=o_j^{k-1}.$

3. a kind of method for distinguishing speek person based on the gauss hybrid models that embeds time-delay neural network according to claim 2 is characterized in that, and is described

Computation process is divided into output layer and two kinds of situations of hidden layer of TDNN;

For output layer:

$\frac{\&PartialD;F\left(x\right)}{{\&PartialD;y}_i^k}=-\frac{1}{p\left((x-o)\right|\mathrm{\&lambda;})}\frac{\&PartialD;p\left((x-o)\right|\mathrm{\&lambda;})}{{\&PartialD;o}_i^k}\frac{{\&PartialD;o}_i^k}{y_i^k}$

$=-\frac{f^{\&prime;}\left(y_i^k\right)}{p\left((x-o)\right|\mathrm{\&lambda;})}\&PartialD;\left(\underset{n=1}{\overset{M}{\mathrm{\&Sigma;}}}{p}_n{c}_n{e}^{-\frac{1}{2}{\left({x-o-u}_n\right)}^T{\mathrm{\&Sigma;}}_n^{-1}\left({x-o-u}_n\right)}\right)/{\&PartialD;o}_i^k$

$=-\frac{f^{\&prime;}\left(y_i^k\right)}{p\left((x-o)\right|\mathrm{\&lambda;})}\underset{n=1}{\overset{M}{\mathrm{\&Sigma;}}}{p}_n{c}_n\left(\frac{a_n\left({x-o-u}_n\right)}{{\mathrm{\&sigma;}}_{n,i}^2}(x_i-o_i-u_{n,i})\right)---\left(15\right)$

Wherein: $a_n\left({x-o-u}_n\right)=e^{-\frac{1}{2}{\left({x-o-u}_n\right)}^T{\mathrm{\&Sigma;}}_n^{-1}\left({x-o-u}_n\right)},$ $c_n=\frac{1}{{\left(2\mathrm{\&pi;}\right)}^{D/2}{\left|{\mathrm{\&Sigma;}}_n\right|}^{1/2}}$

For hidden layer:

$\frac{\&PartialD;F\left(x\right)}{{\&PartialD;y}_i^k}=\underset{j}{\mathrm{\&Sigma;}}\frac{\&PartialD;F\left(x\right)}{\&PartialD;y_j^{k+1}}\frac{{\&PartialD;y}_i^{k+1}}{\&PartialD;y_i^k}=\underset{j}{\mathrm{\&Sigma;}}\frac{\&PartialD;F\left(x\right)}{\&PartialD;y_j^{k+1}}\frac{\&PartialD;\left(\underset{n}{\mathrm{\&Sigma;}}{\mathrm{\&omega;}}_{\mathrm{jn}}^{k+1}{o}_n^k\right)}{\&PartialD;y_i^k}$

$=\underset{j}{\mathrm{\&Sigma;}}\frac{\&PartialD;F\left(x\right)}{\&PartialD;y_j^{k+1}}\frac{{\&PartialD;o}_i^k}{{\&PartialD;y}_i^k}{\mathrm{\&omega;}}_{\mathrm{ji}}^{k+1}=f^{\&prime;}\left(y_i^k\right)\underset{j}{\mathrm{\&Sigma;}}\frac{\&PartialD;F\left(x\right)}{\&PartialD;y_i^{k+1}}{\mathrm{\&omega;}}_{\mathrm{ji}}^{k+1}.$

4. a kind of method for distinguishing speek person based on the gauss hybrid models that embeds time-delay neural network according to claim 1 is characterized in that described pre-emphasis adopts f (Z)=1-0.97Z ^-1Wave filter, divide frame to adopt length 20ms and window to move the Hamming window of 10ms, the exponent number of linear prediction analysis is 20, proper vector is the cepstrum coefficient on 13 rank.

Images