CN111785285A - Voiceprint recognition method for home multi-feature parameter fusion - Google Patents

Abstract

The invention discloses a voiceprint recognition method for household multi-feature parameter fusion, comprising the following steps: respectively calculating MFCC feature parameters, GFCC feature parameters and LPCC feature parameters extracted from speech signals; Three Gaussian mixture models are trained with LPCC feature parameters; the results of the three Gaussian mixture models are weighted and fused, soft decision is made, the threshold is set, and the stochastic gradient descent method is used to obtain the optimal weight coefficient and output the final recognition result. The present invention fuses the MFCC feature parameters, the GFCC feature parameters and the LPCC feature parameters, which makes up for the defect that a single feature parameter cannot better express the speaker's features, thereby greatly improving the accuracy of voiceprint recognition.

Description

Voiceprint recognition method for home multi-feature parameter fusion

technical field

The invention belongs to the field of voiceprint recognition, in particular to a voiceprint recognition method for household multi-feature parameter fusion.

Background technique

Voiceprint recognition, also known as speaker recognition, includes speaker identification and speaker confirmation. Voiceprint recognition has a wide range of applications, including financial, military security, medical, and home security. Before the recognition of many voiceprint recognition systems, in addition to preprocessing operations, feature parameters and model matching are critical to the recognition accuracy.

The traditional single feature parameter cannot express the speaker's speech features well, which may cause over-fitting, and the MFCC feature parameters are easy to imitate. In addition to a single feature, many scholars directly connect GFCC and MFCC to form a new feature parameter vector, which will bring dimensional disaster and increase the computational complexity of the system. Therefore, the current home voiceprint recognition algorithm cannot meet the needs of expressing the characteristics of the speaker better, and its recognition accuracy needs to be improved.

SUMMARY OF THE INVENTION

Purpose of the invention: In order to overcome the deficiencies in the prior art, a voiceprint recognition method for household multi-feature parameter fusion is provided, which effectively solves the problem that a single feature parameter cannot fully express the speaker's voice features, and improves the voiceprint performance. recognition accuracy.

Technical solution: In order to achieve the above purpose, the present invention provides a voiceprint recognition method for household multi-feature parameter fusion, comprising the following steps:

S1: Calculate the MFCC feature parameters, GFCC feature parameters and LPCC feature parameters extracted from the speech signal respectively;

S2: Use MFCC feature parameters, GFCC feature parameters and LPCC feature parameters to train three mixture Gaussian models respectively;

S3: The results of the three mixed Gaussian models are weighted and fused, soft decision is made, the threshold is set, and the stochastic gradient descent method is used to obtain the optimal weight coefficient and output the final recognition result.

Further, in the step S1, the speech signal undergoes a preprocessing operation before the feature parameter extraction is performed.

Further, the preprocessing operations in the step S1 include sample quantization, pre-emphasis, frame-by-frame windowing, and endpoint detection.

Further, the extraction process of the MFCC feature parameters in the step S1 is:

A1) carry out preprocessing to the input speech signal, generate time domain signal, and obtain speech linear spectrum by fast Fourier transform or discrete Fourier transform processing for each frame of speech signal;

A2) filter the linear spectrum input Mel filter bank, generate Mel spectrum, get the logarithmic energy of Mel spectrum, generate corresponding logarithmic spectrum;

A3) Convert the logarithmic spectrum solution to MFCC characteristic parameters by using discrete cosine transform.

Further, the extraction process of the GFCC feature parameters in the step S1 is:

B1) preprocessing the speech signal to generate a time-domain signal, and obtain a discrete power spectrum through fast Fourier transform or discrete Fourier transform processing;

B2) square the discrete power spectrum to generate the speech energy spectrum, and use the Gammatone filter bank for filtering;

B3) exponentially compress the output of each Gammatone filter to obtain a set of exponential energy spectrum;

B4) Using discrete cosine transform to transform exponential energy spectrum into GFCC characteristic parameters.

Further, the extraction process of the LPCC feature parameters in the step S1 is:

C1) Set the system function of the channel model;

C2) Set the impulse response of the system function, and calculate the complex cepstrum of the impulse response;

C3) According to the relationship between the complex cepstrum and the cepstral coefficient, the LPCC characteristic parameters are obtained by calculation.

Further, the determination method of the recognition result in the step S3 is as follows: when the result of weighted fusion is greater than or equal to the threshold, the target speaker is recognized, otherwise, the non-target speaker is recognized.

Beneficial effects: Compared with the prior art, the present invention fuses MFCC feature parameters, GFCC feature parameters and LPCC feature parameters, which makes up for the defect that a single feature parameter cannot better express the speaker's features, thereby greatly improving voiceprint recognition. Accuracy.

Description of drawings

Fig. 1 is the overall structure block diagram of the method of the present invention;

Fig. 2 is a flowchart of MFCC feature parameter extraction;

Figure 3 is a flowchart of GFCC feature parameter extraction.

Detailed ways

Below in conjunction with the accompanying drawings and specific embodiments, the present invention will be further clarified. It should be understood that these embodiments are only used to illustrate the present invention and not to limit the scope of the present invention. Modifications of equivalent forms all fall within the scope defined by the appended claims of this application.

As shown in Figure 1, the present invention provides a voiceprint recognition method for household CNN classification and feature matching, comprising the following steps:

1) Preprocess the input speaker's speech, including sample quantization, pre-emphasis, windowing and framing, endpoint detection, etc. The purpose of preprocessing is to eliminate the interference of vocal organs and voice acquisition equipment and improve the recognition rate of the system.

2) Calculate the MFCC feature parameters, GFCC feature parameters and LPCC feature parameters extracted from the speech signal respectively;

3) Using MFCC feature parameters, GFCC feature parameters and LPCC feature parameters to train three Gaussian mixture models, which are GMM model A, GMM model B and GMM model C respectively;

4) The results of GMM model A, GMM model B and GMM model C are weighted and fused, soft decision is made, threshold is set, and the stochastic gradient descent method is used to obtain the optimal weight coefficient and output the final recognition result.

As shown in Figure 2, the extraction process of the MFCC feature parameters in this embodiment is:

A1) Preprocess the input speech signal s(n) to generate a time-domain signal x(n) (the length of the signal sequence N=256), and then, perform fast Fourier transform or discrete Fourier transform on each frame of speech signal Lie transform processing obtains the speech linear spectrum X(k), which can be expressed as:

A2) Input the linear spectrum X(k) into the Mel filter bank for filtering to generate the Mel spectrum, and then take the logarithmic energy of the Mel spectrum to generate the corresponding logarithmic spectrum S(m).

Here, the Mel filter bank is a set of triangular-strip filters H _m (k), and needs to satisfy 0≤m≤M, where M represents the number of filters, usually 20-28. The transfer function of the bandpass filter can be expressed as:

In formula (2), f(m) is the center frequency.

Among them, the reason for taking the logarithm of the Mel energy spectrum is to promote the improvement of the performance of the voiceprint recognition system. The transfer function from the speech linear spectrum X(k) to the logarithmic spectrum S(m) is:

A3) The logarithmic spectrum S(m) is converted into MFCC feature parameters by using discrete cosine transform (DCT), and the expression of the nth dimension feature component C(n) of the MFCC feature parameters is:

The MFCC characteristic parameters obtained through the above steps only reflect the static characteristics of the speech signal, and the dynamic characteristic parameters can be obtained by calculating the first-order and second-order differences thereof.

In the present embodiment, the extraction process of the characteristic parameter of GFCC (Gammatone frequency cepstral coefficient) is applied to the Gammatone filter, and its design scheme is:

The Gammatone filter bank is used to simulate the auditory properties of the cochlear basilar membrane, and its time domain expression is as follows:

g(f, t)=t ^n-1 e ^-2πbt cos(2πf _i +φ _i )U(t), 1≤i≤N (5) In the formula, N——the number of filters;

n——Number of filter stages, generally 4;

i——filter ordinal;

f _i ——the center frequency of the filter;

U(t)——unit step function;

b _i ——the attenuation factor of the filter;

φ _i ——The phase of the filter whose sequence is i, generally 0.

The bandwidth of each filter is related to the auditory critical band of the human ear. According to the theory of psychology, the critical band can be expressed by the equivalent rectangular bandwidth:

The attenuation factor b _i of the filter is related to the bandwidth, and the attenuation rate of the impulse response is determined by the attenuation factor b _i . Its expression is:

b _i = 1.019EBR(f) (7)

The time-domain impulse function of the Gammatone filter is an analog function. In order to facilitate the calculation and processing, it needs to be discretized. The Laplace transform of formula (4) is as follows:

The input speech signal s(n) and g _i (n) can get the output of Gammatone filter through convolution operation.

The extraction process of GFCC feature parameters is similar to the extraction process of MFCC feature parameters. It only needs to use the Gammatone filter bank to replace the traditional Mel filter bank, which effectively utilizes the cochlear basilar membrane characteristics of the Gammatone filter The speech signal is processed non-linearly.

Based on the above Gammatone filter, as shown in Figure 3, the extraction process of GFCC (Gammatone frequency cepstral coefficient) characteristic parameters is as follows:

B1) First, the input speech signal s(n) is preprocessed to generate a time domain signal x(n), and the discrete power spectrum X(k) is obtained through fast Fourier transform or discrete Fourier transform processing, and its expression for:

B2) Square the discrete power spectrum X(k) to generate the speech energy spectrum, and then use the Gammatone filter bank for filtering.

B3) In order to better improve the performance of the voiceprint recognition system, exponentially compress the output of each filter to obtain a set of exponential energy spectra s ₁ , s ₂ ,...,s _M :

In the formula, e(f)——exponential compression value, M——number of filter channels.

B4) Finally, use discrete cosine transform (DCT) to transform the exponential energy spectrum into the GFCC characteristic parameters, whose expression is:

In the formula, L is the dimension of the feature parameter.

In the present embodiment, the extraction process of LPCC (Linear Prediction Cepstral Coefficient) characteristic parameter is:

Suppose the system function of the vocal tract model is as follows:

In equation (12), p is the order of the predictor.

是h(n)的复倒谱，则Let h(n) be the impulse response of H(z),

is the complex cepstrum of h(n), then

Synthesizing formula (12) and formula (13), and taking the derivative of z ^-1 , after simplification, we can get:

The complex cepstrum can be obtained by adding the coefficients of the powers of z ^-1 on both sides of the equal sign of formula (14), as follows:

According to the relationship between complex cepstrum and cepstral coefficient:

The linear prediction cepstral coefficients can be calculated:

where c( _n ) is the linear prediction cepstral coefficient LPCC, and an is the linear prediction coefficient.

最终结果D＝ω_1a+ω_2b+ω_3c，设定阈值γ，当D大于等于阈值γ时，识别为目标说话人，否则识别为非目标说话人。In step 4 in this embodiment, the mixing degree of GMM model A, GMM model B, and GMM model C are all set to 1024. The output results of the three models are a, b, and c, respectively, and the three results are weighted and fused, and the weight coefficient is ω _i and

The final result is D=ω _1a +ω _2b +ω _3c , and a threshold γ is set. When D is greater than or equal to the threshold γ, it is recognized as a target speaker, otherwise it is recognized as a non-target speaker.

Claims (8)

1. The voiceprint recognition method for home multi-feature parameter fusion is characterized by comprising the following steps: the method comprises the following steps:

s1: respectively calculating and extracting MFCC characteristic parameters, GFCC characteristic parameters and LPCC characteristic parameters of the voice signals;

s2: training three Gaussian mixture models by respectively using MFCC characteristic parameters, GFCC characteristic parameters and LPCC characteristic parameters;

s3: and weighting and fusing the results of the three Gaussian mixture models, performing soft decision, setting a threshold value, obtaining an optimal weight coefficient by using a random gradient descent method, and outputting a final recognition result.

2. The home furnishing multi-feature parameter fusion-oriented voiceprint recognition method according to claim 1, wherein: the speech signal undergoes a preprocessing operation in the step S1 before feature parameter extraction.

3. The home furnishing multi-feature parameter fusion-oriented voiceprint recognition method according to claim 2, wherein: the preprocessing operation in step S1 includes sampling quantization, pre-emphasis, frame windowing, and endpoint detection.

4. The home furnishing multi-feature parameter fusion-oriented voiceprint recognition method according to claim 1, wherein: the extraction process of the MFCC characteristic parameters in step S1 is as follows:

A1) preprocessing an input voice signal to generate a time domain signal, and processing each frame of voice signal through fast Fourier transform or discrete Fourier transform to obtain a voice linear frequency spectrum;

A2) inputting the linear frequency spectrum into a Mel filter bank for filtering to generate a Mel frequency spectrum, and taking the logarithmic energy of the Mel frequency spectrum to generate a corresponding logarithmic frequency spectrum;

A3) the log spectrum solution is converted to MFCC feature parameters by using a discrete cosine transform.

5. The home furnishing multi-feature parameter fusion-oriented voiceprint recognition method according to claim 1, wherein: the extraction process of the GFCC characteristic parameters in step S1 is as follows:

B1) preprocessing a voice signal to generate a time domain signal, and obtaining a discrete power spectrum through fast Fourier transform or discrete Fourier transform processing;

B2) squaring the discrete power spectrum to generate a voice energy spectrum, and performing filtering processing by using a Gamma atom filter bank;

B3) performing exponential compression on the output of each Gamma filter to obtain a group of exponential energy spectrums;

B4) the exponential energy spectrum is converted into GFCC characteristic parameters using a discrete cosine transform.

6. The home furnishing multi-feature parameter fusion-oriented voiceprint recognition method according to claim 1, wherein: the extraction process of the LPCC characteristic parameters in step S1 is as follows:

C1) setting a system function of the vocal tract model;

C2) setting the impulse response of a system function, and calculating the complex cepstrum of the impulse response;

C3) and calculating to obtain the LPCC characteristic parameters according to the relation between the complex cepstrum and the cepstrum coefficient.

7. The home furnishing multi-feature parameter fusion-oriented voiceprint recognition method according to claim 1, wherein: the determination method of the recognition result in step S3 is: and when the result of the weighted fusion is greater than or equal to the threshold value, the target speaker is identified, otherwise, the non-target speaker is identified.

8. The home furnishing multi-feature parameter fusion-oriented voiceprint recognition method according to claim 1, wherein: the Gammatone filter bank in the step B2 is used for simulating the auditory characteristics of the cochlear basilar membrane, and the time domain expression thereof is as follows:

g(f，t)＝t^n-1e^-2πbtcos(2πf_i+φ_i)U(t)，1≤i≤N

wherein N is the number of filters, N is the number of filter stages, i is the number of filter stages, f_iFor the center frequency of the filter, U (t) is the unit step function, b_iIs the attenuation factor of the filter, phi_iIs the phase of the filter of sequence i.

Images