Voiceprint recognition method and system

Abstract

The invention discloses a voiceprint recognition method and system, and particularly relates to the field of voiceprint recognized.The voiceprint recognition method comprises the steps that a microphone is used for capturing a sound signal of a user, and recording is conducted through equipment; then, noise reduction is carried out on the collected sound signal data; then, the processed sound signals are converted into characteristic parameters, and the characteristic parameters comprise a first sound frequency, a spectral bandwidth and a second sound characteristic coefficient; thirdly, constructing a hidden Markov model by utilizing the characteristic parameters and training the hidden Markov model; finally, the trained model is used for voiceprint matching, a similarity score is calculated, and whether identity verification succeeds or not is judged according to a threshold value; and if the similarity score exceeds a set threshold value, determining that the voiceprint matching is successful, otherwise, determining that the voiceprint matching fails.

Description

Technical Field

[0001] The present invention relates to the technical field of voiceprint recognition, and more specifically, to a voiceprint recognition method and system. Background Art

[0002] As a more convenient and natural means of identity authentication, voiceprint recognition is widely used in identity authentication, telephone customer service, financial security, education and medical care, public security and criminal investigation, smart home, smart building and other fields; voiceprint recognition is an artificial intelligence technology with broad application prospects. From national security and military communications to smartphone voiceprint unlocking and smart device wake-up services, voiceprint recognition is playing a huge role; it brings better security, user experience and cost-effectiveness to many fields, and has broad social and economic significance.

[0003] Compared with traditional security authentication methods such as passwords, signatures, fingerprints, etc., voiceprint recognition has many advantages such as being difficult to be stolen, forgotten, lost or forged, and having higher security and reliability.

[0004] However, it still has some shortcomings in actual use. For example, the environmental noise of traditional voiceprint data will interfere with the extraction and matching of voiceprints, reducing the accuracy of recognition. In addition, the speaker's mental state, emotions or health status may cause changes in pronunciation, thus affecting the recognition effect. Summary of the invention

[0005] In order to overcome the above defects of the prior art, an embodiment of the present invention provides a voiceprint recognition method to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions: Step A1: Use a microphone acquisition device to capture the user's voice signal data; Step A2: using spectral subtraction to perform denoising on the collected sound signal data; Step A3: converting the preprocessed sound signal data into feature parameters; Step A4: constructing a hidden Markov model with feature parameters and training it; Step A5: Used to calculate the voiceprint value obtained by the data trained by the hidden Markov model and match it with the voiceprint model in the database to calculate the similarity score; Step A6: Determine whether the match is successful based on the similarity score and perform identity verification.

[0007] Preferably, in step A1, select a microphone and ensure that the microphone is compatible with the recording device; correctly connect the microphone to the recording device; for a USB microphone, simply plug it into the USB port of the device; for a microphone that requires a sound card or audio interface, connect the microphone to the sound card or audio interface, and then connect the sound card or audio interface to a computer or other recording device.

[0008] Set the correct input source as the microphone on the recording device; adjust the volume control to ensure that the microphone volume is moderate to avoid distortion or unclear recording caused by excessive or low volume; turn on the recording function of the recording software or device; press the record button to start recording; at this time, the sound will be captured through the microphone and converted into a digital signal stored in the device; during the recording process, pay attention to monitor the recording quality; if you find noise, hum or uneven volume, you can try to adjust the microphone position, reduce environmental noise or adjust the recording device settings.

[0009] Preferably, in step A2, the collected sound signal data is denoised and processed using spectral subtraction.

[0010] Preferably, in step A3, the characteristic parameters include a first sound frequency, a spectrum bandwidth, and a second sound characteristic coefficient; The calculation method of the first sound frequency is specifically as follows: , where I represents the first sound frequency and G represents the first fundamental frequency period; The calculation method of spectrum bandwidth is as follows: , where J represents the spectrum bandwidth, Expressed as the variance of the power spectral density function; The calculation method of the second sound characteristic coefficient is specifically as follows: ,in, It is represented as the second sound characteristic coefficient of the u-th dimension, F is represented as the length of the speech signal, and r is represented as the dimension of the frequency domain information. Represented as the FBank feature of the qth frame, It is expressed as pi, q is expressed as the qth frame; V is expressed as a constant, V is related to the filter, if M triangular filters are used, then M is the number of filters.

[0011] Preferably, in step A4, the first sound frequency, spectrum bandwidth and linear prediction coefficient are used to construct a hidden Markov model, and the method for constructing the hidden Markov model is specifically as follows: Step 1: Data preprocessing: The calculated first sound frequency, spectral bandwidth, and linear prediction coefficient are standardized to ensure that different features are on the same scale to facilitate model training.

[0012] Step 2: Initialize model parameters: set the initial probability of each hidden state; Initialize the state transition probability matrix N; the calculation method of the state transition probability matrix N is specifically as follows: ,in, Expressed as a state transition probability matrix, It is expressed as the count of transitions from state j to state k; Initialize the observation probability matrix R, and for each hidden state, define the probability of observing a specific eigenvalue; the calculation method of the observation probability matrix is ​​as follows: ,in, Expressed as an observation probability matrix, the probability of observing the observation value p in state m, It is expressed as the count of observation value p observed in state m, Step 3: Model training: Use the expectation maximization algorithm to adjust the model parameters; Step 4: Model evaluation: Use the data as a test set to evaluate the performance of the model and adjust the model parameters based on the evaluation results; Through steps 1 to 4, a hidden Markov model is constructed using the first sound frequency, spectrum bandwidth and linear prediction coefficient for processing and analyzing acoustic signals.

[0013] Preferably, in step A5, the voiceprint value is calculated as follows: , where Z represents the voiceprint value, J represents the spectrum bandwidth, and I represents the first sound frequency. is represented as a time variable, Expressed as the second sound characteristic coefficient, , Expressed as weight coefficient.

[0014] The similarity score is calculated as follows: , where Z represents the voiceprint value, It represents the voiceprint value of the i-th voiceprint model in the database, and i represents the i-th voiceprint model.

[0015] Preferably, in step A6, the calculated similarity score is compared with the threshold value. If the calculated similarity score is greater than the threshold value, the voiceprint match is output as successful. At this time, the system prompts: user verification is successful; if the calculated similarity score is less than the threshold value, the voiceprint match fails. At this time, the system prompts: user verification failed, please re-verify.

[0016] Technical effects and advantages of the present invention: The system collects user voice signals through a microphone, performs noise reduction and preprocessing, and then extracts characteristic parameters such as the first sound frequency, spectral bandwidth and the second sound characteristic coefficient; then, these features are used to construct and train a hidden Markov model (HMM); the trained model is used to calculate the voiceprint value, and match it with the voiceprint model in the database to calculate the similarity score; based on the comparison result of the similarity score with a preset threshold, the system will output an instruction on whether the identity authentication is successful or failed; the present invention removes the environmental noise during voiceprint extraction through spectral subtraction noise reduction, thereby improving the accuracy of recognition, and the hidden Markov model is trained to remove the influence of changes in psychological state, emotion or health status on pronunciation, thereby reducing the impact on the recognition effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic diagram of the method flow of the present invention.

[0018] Figure 2 It is a schematic diagram of module connection of the present invention. DETAILED DESCRIPTION

[0019] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0020] See also Figure 1 As shown, the present invention provides a voiceprint recognition method, which is: Step A1: Use a microphone acquisition device to capture the user's voice signal data; In step A1, select a microphone and ensure that the microphone is compatible with the recording device; correctly connect the microphone to the recording device; for a USB microphone, just plug it into the USB port of the device; for a microphone that requires a sound card or audio interface, connect the microphone to the sound card or audio interface, and then connect the sound card or audio interface to the computer or other recording device.

[0021] Set the correct input source as the microphone on the recording device; adjust the volume control to ensure that the microphone volume is moderate to avoid distortion or unclear recording caused by excessive or low volume; turn on the recording function of the recording software or device; press the record button to start recording; at this time, the sound will be captured through the microphone and converted into a digital signal stored in the device; during the recording process, pay attention to monitor the recording quality; if you find noise, hum or uneven volume, you can try to adjust the microphone position, reduce environmental noise or adjust the recording device settings.

[0022] Step A2: using spectral subtraction to perform denoising on the collected sound signal data; In step A2, the step of denoising the collected sound signal data using spectral subtraction is specifically as follows: Step B1: Obtaining sound signal data containing noise ; Step B2: Use short-time Fourier transform to transform the time domain signal into the frequency domain. The specific calculation method is: ,in, It is represented as the transformation of the frequency of b at time a; a is represented as the time frame index, indicating the time frame position of the current analysis; It is represented as a window function, b is the frequency bin index, and represents the frequency domain component after discrete Fourier transform; f is the length of the window; d is the imaginary unit; e is the base of the natural logarithm; It is expressed as pi; Step B3: Select a noise-dominated segment and calculate the noise power spectrum of the segment. The specific method for calculating the noise power spectrum is: ,in, Expressed as the noise power spectrum, represents the noise power spectrum estimate at the bth frequency bin, It is expressed as the transformation of the frequency of b at time a; Represented as a smoothing parameter, the value range is It controls the weight of the current noise estimate and the previous noise estimate; smaller values will make the impact of new data greater, and larger The value will make the historical data have a greater impact; Step B4: Calculate the power spectrum of the mixed signal. The specific calculation method is: ,in, Expressed as a power spectrum, It is expressed as the transformation of the frequency of b at time a; Step B5: For each frequency b, subtract the noise power spectrum from the signal power spectrum. The specific calculation method is: ,in, Expressed as the enhanced power spectrum, Expressed as a power spectrum, Expressed as a noise power spectrum; Step B6: Calculate the amplitude spectrum after denoising. The specific calculation method is: ,in, Expressed as an amplitude spectrum, Represented as the phase information of the sound signal data, It is represented as the enhanced power spectrum, d is represented as the imaginary unit, and e is represented as the base of the natural logarithm; Step B7: Use the inverse short-time Fourier transform to convert the denoised spectrum back to the time domain, and obtain the denoised sound signal data. The calculation method of the inverse short-time Fourier transform is specifically as follows: ,in, represents the denoised sound signal data obtained by inverse short-time Fourier transform, b represents the frequency bin index, which represents the frequency domain component after discrete Fourier transform; f represents the length of the window; d represents the imaginary unit; e represents the base of the natural logarithm; Expressed as pi, The complex value of the b-th frequency bin; After steps B1 to B7, the collected sound signal data can be denoised to obtain denoised sound signal data; Step A3: converting the preprocessed sound signal data into feature parameters; In step A3, the characteristic parameters include the first sound frequency, the spectrum bandwidth, and the second sound characteristic coefficient; The calculation method of the first sound frequency is specifically as follows: , where I represents the first sound frequency and G represents the first fundamental frequency period; The calculation method of the first fundamental frequency period is specifically as follows: , where G represents the first fundamental frequency period, Represents the independent variable that minimizes the function. In the formula, it is used to find the function The minimum value of c is reached, Expressed as an autocorrelation function; The calculation method of the autocorrelation function is as follows: ,in, Expressed as the autocorrelation function, It represents the hth sampling point of the speech signal; c represents the delay amount.

[0023] The calculation method of spectrum bandwidth is as follows: , where J represents the spectrum bandwidth, Expressed as the variance of the power spectral density function; The calculation method of the variance of the power spectral density function is as follows: ,in, It is expressed as the variance of the power spectral density function, K is the total number of frequency components, It is expressed as the mean of the power spectral density function, and L is expressed as the power spectral density function; The calculation method of the mean of the power spectral density function is as follows: ,in, It is represented as the mean of the power spectral density function, K is represented as the total number of frequency components, and L is represented as the power spectral density function; The calculation method of the power spectral density function is as follows: , where L is the power spectral density function, It is represented by pi, and M is the spectrum of the sound signal after Fourier transformation; The calculation method of the second sound characteristic coefficient is specifically as follows: ,in, It is represented as the second sound characteristic coefficient of the u-th dimension, F is represented as the length of the speech signal, and r is represented as the dimension of the frequency domain information. Represented as the FBank feature of the qth frame, It is represented by pi, q represents the qth frame; V represents a constant, V is related to the filter, if M triangular filters are used, then M is the number of filters; ,in, It is represented as the FBank feature of the qth frame, F represents the length of the speech signal, and r represents the dimension of the frequency domain information. It is expressed as the signal amplitude of the rth frequency in the qth frame, It represents the weight coefficient of the rth filter in the Mel filter bank for the rth frequency of the qth frame, and t is represented by an arbitrary base.

[0024] Step A4: constructing a hidden Markov model with feature parameters and training it; In step A4, the first sound frequency, spectrum bandwidth and linear prediction coefficient are used to construct a hidden Markov model. The method for constructing the hidden Markov model is specifically as follows: Step 1: Data preprocessing: The calculated first sound frequency, spectral bandwidth, and linear prediction coefficient are standardized to ensure that different features are on the same scale to facilitate model training.

[0025] Step 2: Initialize model parameters: set the initial probability of each hidden state; Initialize the state transition probability matrix N; the calculation method of the state transition probability matrix N is specifically as follows: ,in, Expressed as a state transition probability matrix, It is expressed as the count of transitions from state j to state k; Initialize the observation probability matrix R, and for each hidden state, define the probability of observing a specific eigenvalue; the calculation method of the observation probability matrix is ​​as follows: ,in, Expressed as an observation probability matrix, the probability of observing the observation value p in state m, It is expressed as the count of observation value p observed in state m, Step 3: Model training: Use the expectation maximization algorithm to adjust the model parameters; Step 4: Model evaluation: Use the data as a test set to evaluate the performance of the model and adjust the model parameters based on the evaluation results; Through steps 1 to 4, a hidden Markov model is constructed using the first sound frequency, spectrum bandwidth and linear prediction coefficient for processing and analyzing acoustic signals.

[0026] Step A5: Calculate the voiceprint value using the data trained by the hidden Markov model, match the calculated voiceprint value with the voiceprint model in the database, and calculate the similarity score; In step A5, the voiceprint value is calculated as follows: , where Z represents the voiceprint value, J represents the spectrum bandwidth, and I represents the first sound frequency. is represented as a time variable, Expressed as the second sound characteristic coefficient, , Expressed as weight coefficient.

[0027] The similarity score is calculated as follows: , where Z represents the voiceprint value, It represents the voiceprint value of the i-th voiceprint model in the database, and i represents the i-th voiceprint model; The database stores the voiceprint models of v users and their corresponding voiceprint thresholds; these thresholds are obtained through the statistics of the voiceprint values ​​of the training samples and are used to distinguish between matches with high confidence and matches with low confidence.

[0028] Step A6: Determine whether the match is successful based on the similarity score and perform identity authentication; In step A6, the calculated similarity score is compared with the threshold. If the calculated similarity score is greater than the threshold, the voiceprint match is output as successful. At this time, the system prompts: user verification is successful; if the calculated similarity score is less than the threshold, the voiceprint match fails. At this time, the system prompts: user verification failed, please re-verify.

[0029] See also Figure 2 As shown, in this embodiment, it should be specifically explained that the present invention provides a voiceprint recognition system, including the following modules: Sound signal data acquisition module: used to capture the user's sound signal data using a microphone acquisition device; Sound signal data processing module: used to perform denoising on the collected sound signal data using spectral subtraction; Sound signal data feature extraction module: used to convert the pre-processed sound signal data into feature parameters. Voiceprint model construction and training module: used to construct and train the hidden Markov model with feature parameters, and transmit the trained feature parameters to the voiceprint matching module; Voiceprint matching module: used to calculate the voiceprint value obtained by the data trained by the hidden Markov model and match it with the voiceprint model in the database, calculate the similarity score, and transmit the similarity score to the identity authentication module; Authentication module: used to determine whether a match is successful based on the similarity score, and to perform identity authentication or identification.

[0030] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A voiceprint recognition method, characterized in that: include: Step A1: Use a microphone acquisition device to capture the user's voice signal data; Step A2: using spectral subtraction to perform denoising on the collected sound signal data; Step A3: converting the preprocessed sound signal data into feature parameters; Step A4: constructing a hidden Markov model with feature parameters and training it; Step A5: Used to calculate the voiceprint value obtained by the data trained by the hidden Markov model and match it with the voiceprint model in the database to calculate the similarity score; Step A6: Determine whether the match is successful based on the similarity score and perform identity verification.

2. A voiceprint recognition method according to claim 1, characterized in that: In step A1, select a microphone and ensure that the microphone is compatible with the recording device; correctly connect the microphone to the recording device; for a USB microphone, just plug it into the USB port of the device; for a microphone that requires a sound card or audio interface, connect the microphone to the sound card or audio interface, and then connect the sound card or audio interface to the computer or other recording device.

3. A voiceprint recognition method according to claim 1, characterized in that: In step A2, the step of denoising the collected sound signal data using spectral subtraction is specifically as follows: Step B1: Obtaining sound signal data containing noise ; Step B2: Use short-time Fourier transform to transform the time domain signal into the frequency domain. The specific calculation method is: ,in, It is represented as the transformation of the frequency of b at time a; a is represented as the time frame index, indicating the time frame position of the current analysis; It is represented as a window function, b is the frequency bin index, and represents the frequency domain component after discrete Fourier transform; f is the length of the window; d is the imaginary unit; e is the base of the natural logarithm; It is expressed as pi; Step B3: Select a noise-dominated segment and calculate the noise power spectrum of the segment. The specific method for calculating the noise power spectrum is: ,in, Expressed as the noise power spectrum, represents the noise power spectrum estimate at the bth frequency bin, It is expressed as the transformation of the frequency of b at time a; It is represented as a smoothing parameter, and its value range is It controls the weight of the current noise estimate and the previous noise estimate; smaller values will make the impact of new data greater, and larger The value will make the historical data have a greater impact; Step B4: Calculate the power spectrum of the mixed signal. The specific calculation method is: ,in, Expressed as a power spectrum, It is expressed as the transformation of the frequency of b at time a; Step B5: For each frequency b, subtract the noise power spectrum from the signal power spectrum. The specific calculation method is: ,in, Expressed as the enhanced power spectrum, Expressed as a power spectrum, Expressed as a noise power spectrum; Step B6: Calculate the amplitude spectrum after denoising. The specific calculation method is: ,in, Expressed as an amplitude spectrum, Represented as the phase information of the sound signal data, It is represented as the enhanced power spectrum, d is represented as the imaginary unit, and e is represented as the base of the natural logarithm; Step B7: Use the inverse short-time Fourier transform to convert the denoised spectrum back to the time domain, and obtain the denoised sound signal data. The calculation method of the inverse short-time Fourier transform is specifically as follows: ,in, represents the denoised sound signal data obtained by inverse short-time Fourier transform, b represents the frequency bin index, which represents the frequency domain component after discrete Fourier transform; f represents the length of the window; d represents the imaginary unit; e represents the base of the natural logarithm; Expressed as pi, The complex value of the bth frequency bin; Through step B1 to step B7, the collected sound signal data can be denoised to obtain denoised sound signal data.

4. A voiceprint recognition method according to claim 1, characterized in that: In step A3, the characteristic parameters include the first sound frequency, the spectrum bandwidth, and the second sound characteristic coefficient; The calculation method of the first sound frequency is specifically as follows: , where I represents the first sound frequency and G represents the first fundamental frequency period; The calculation method of the first fundamental frequency period is specifically as follows: , where G represents the first fundamental frequency period, Represents the independent variable that minimizes the function. In the formula, it is used to find the function The c value that reaches the minimum value; It is expressed as an autocorrelation function.

5. A voiceprint recognition method according to claim 4, characterized in that: The calculation method of spectrum bandwidth is as follows: , where J represents the spectrum bandwidth, Expressed as the variance of the power spectral density function; The calculation method of the variance of the power spectral density function is as follows: ,in, It is expressed as the variance of the power spectral density function, K is the total number of frequency components, is represented as the mean of the power spectral density function, and L is represented as the power spectral density function.

6. A voiceprint recognition method according to claim 4, characterized in that: The calculation method of the second sound characteristic coefficient is specifically as follows: ,in, It is represented as the second sound characteristic coefficient of the u-th dimension, F is represented as the length of the speech signal, and r is represented as the dimension of the frequency domain information. Represented as the FBank feature of the qth frame, It is represented as pi, q represents the qth frame; V is represented as a constant, V is related to the filter, if M triangular filters are used, then M is the number of filters.

7. A voiceprint recognition method according to claim 1, characterized in that: In step A4, the first sound frequency, spectrum bandwidth and linear prediction coefficient are used to construct a hidden Markov model. The method for constructing the hidden Markov model is specifically as follows: Step 1: Data preprocessing: Standardize the calculated first sound frequency, spectral bandwidth, and linear prediction coefficient to ensure that different features are on the same scale to facilitate model training; Step 2: Initialize model parameters: set the initial probability of each hidden state; Initialize the state transition probability matrix N; the calculation method of the state transition probability matrix N is specifically as follows: ,in, Expressed as a state transition probability matrix, It is expressed as the count of transitions from state j to state k; Initialize the observation probability matrix R, and for each hidden state, define the probability of observing a specific eigenvalue; the calculation method of the observation probability matrix is ​​as follows: ,in, Expressed as an observation probability matrix, the probability of observing the observation value p in state m, It is expressed as the count of observation value p observed in state m, Step 3: Model training: Use the expectation maximization algorithm to adjust the model parameters; Step 4: Model evaluation: Use the data as a test set to evaluate the performance of the model and adjust the model parameters based on the evaluation results.

8. A voiceprint recognition method according to claim 1, characterized in that: In step A5, the voiceprint value is calculated as follows: , where Z represents the voiceprint value, J represents the spectrum bandwidth, and I represents the first sound frequency. is represented as a time variable, Expressed as the second sound characteristic coefficient, , Expressed as weight coefficient; The similarity score is calculated as follows: , where Z represents the voiceprint value, It represents the voiceprint value of the i-th voiceprint model in the database, and i represents the i-th voiceprint model.

9. A voiceprint recognition method according to claim 1, characterized in that: In step A6, the calculated similarity score is compared with the threshold. If the calculated similarity score is greater than the threshold, the voiceprint match is output as successful. At this time, the system prompts: user verification is successful; if the calculated similarity score is less than the threshold, the voiceprint match fails. At this time, the system prompts: user verification failed, please re-verify.

10. A voiceprint recognition system, using a voiceprint recognition method as claimed in any one of claims 1 to 9, characterized in that: Sound signal data acquisition module: used to capture the user's sound signal data using a microphone acquisition device; Sound signal data processing module: used to perform denoising on the collected sound signal data using spectral subtraction; Sound signal data feature extraction module: used to convert the pre-processed sound signal data into feature parameters. Voiceprint model construction and training module: used to construct and train the hidden Markov model with feature parameters, and transmit the trained feature parameters to the voiceprint matching module; Voiceprint matching module: used to calculate the voiceprint value obtained by the data trained by the hidden Markov model and match it with the voiceprint model in the database, calculate the similarity score, and transmit the similarity score to the identity authentication module; Authentication module: used to determine whether a match is successful based on the similarity score, and to perform identity authentication or identification.

Images