An adaptive VAD parameter adjustment method and system based on voiceprint recognition

Abstract

The present application relates to the technical field of data processing, and discloses a kind of adaptive VAD parameter adjustment method and system based on voiceprint identification, its method includes: obtaining audio signal by audio acquisition device, pre-processing audio signal and extracting voiceprint feature, voiceprint feature is mapped to user voiceprint model, current user voiceprint model is compared with the storage model in voiceprint library, judge whether it is registered user;VAD parameter is determined according to whether it is registered user;When judging as non-registered user, the VAD parameter is adjusted in real time by the characteristic fluctuation detection of mute section and speech section;The adjusted VAD parameter is stored in voiceprint library.The VAD parameter adjustment method of the present application can adapt to different user characteristics and environmental changes, and shows stronger robustness and adaptability in complex scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to an adaptive VAD parameter adjustment method and system based on voiceprint recognition. BACKGROUND

[0002] VAD (Voice Activity Detection) is a technology used to distinguish between speech signals and non-speech signals (such as silence, noise). Traditional VAD algorithms usually use fixed energy thresholds or spectral features for speech detection. In complex scenarios, fixed parameters and lack of adaptability can lead to a decrease in speech detection accuracy, resulting in false positives and false negatives. Specifically, in a multi-user scenario, the speech characteristics of different users differ significantly, such as volume, speech rate, frequency distribution, etc. If the VAD parameters are fixed, they cannot be adjusted adaptively for individual characteristics, making it difficult to accurately detect the speech of some users, resulting in missed detection or unstable detection.

[0003] Existing VAD parameter adjustment methods are often based on pre-set models or static rules, lacking dynamic adaptability. When the environment or user changes, the parameters cannot be automatically adjusted, affecting detection accuracy. In addition, in speech communication or speech recognition applications, if the VAD parameters are not adapted to the current environment or user characteristics, it may cause loss of speech segments or introduction of unnecessary background noise, affecting system performance.

[0004] In summary, existing VAD algorithms lack adaptability when faced with complex environments and individual differences, leading to a decrease in speech detection accuracy and false positives and false negatives. SUMMARY

[0005] To solve the above problems, the present application proposes an adaptive VAD parameter adjustment method and system based on voiceprint recognition.

[0006] In one aspect, the present application proposes an adaptive VAD parameter adjustment method based on voiceprint recognition, comprising:

[0007] An audio signal is obtained by an audio acquisition device, the audio signal is preprocessed and voiceprint features are extracted, the voiceprint features are mapped to a user voiceprint model, and the current user voiceprint model is compared with the stored models in the voiceprint library to determine whether it is a registered user;

[0008] The VAD parameters are determined according to whether it is a registered user. When it is determined to be a non-registered user, the VAD parameters are adjusted in real time by detecting the feature fluctuations of the silence and speech segments;

[0009] The adjusted VAD parameters are stored in the voiceprint library.

[0010] Further, when the audio signal is acquired by the audio acquisition device, comprising:

[0011] The audio signal is sampled at the same sampling rate, converted into a single channel, and formatted as a PCM format.

[0012] Further, when the audio signal is preprocessed, comprising:

[0013] S1: Calculate the average value of the audio signal, and subtract the average value of the audio signal from each sampling point;

[0014] S2: Pre-emphasis processing is performed according to the following relationship:

[0015] ;

[0016] S3: The audio signal is divided into frames with a frame shift of 10 ms;

[0017] S4: The audio signal is windowed based on a Hamming window, and each sampling point is multiplied by a Hamming window function;

[0018] Wherein, And are the audio signals of the nth sampling point before and after pre-emphasis processing, is a pre-emphasis coefficient, 0.9 < a < 1, is the audio signal of the n-1th sampling point before pre-emphasis processing.

[0019] Further, when the audio signal is extracted, the voiceprint feature includes MFCC values, LPCC values, and fundamental frequencies.

[0020] Further, when the current user voiceprint model is compared with the stored models in the voiceprint library to determine whether it is a registered user, comprising:

[0021] Obtain the MFCC values, LPCC values, and fundamental frequencies of the current user voiceprint model and each stored model in the voiceprint library and compare them, calculate the similarity, select the maximum similarity, and compare the maximum similarity with the similarity threshold; When the maximum similarity is greater than or equal to the similarity threshold, the current user model is determined to be a registered user, and when the maximum similarity is less than the similarity threshold, the current user model is determined to be a registered user; The similarity satisfies the following relationship:

[0022] ;

[0023] Wherein, is the similarity, is the jth voiceprint feature of the current user voiceprint model,​ is the jth feature of the stored model; is the number of features of the voiceprint, = 3, when j = 1, corresponding to MFCC value; when j = 2, corresponding to LPCC value; when j = 3, corresponding to fundamental frequency.

[0024] Further, when judging as a registered user, the VAD parameter of the stored model corresponding to the maximum similarity in the voiceprint library is called;

[0025] When judging as a non-registered user, the pre-processed audio signal is divided into silent segment and speech segment according to energy value, and the VAD parameter is adjusted in real time according to the feature fluctuation detection of the silent segment and the speech segment.

[0026] Further, when judging as a non-registered user, when divided into silent segment and speech segment according to energy value, it includes:

[0027] The energy value of each frame of audio signal is calculated, and the division into silent segment or speech segment is judged by comparing the energy value with the energy threshold value, and the energy value is obtained by the following relationship:

[0028] ;

[0029] ;

[0030] Wherein, is the energy value of the mth frame, is the kth sampling point of the mth frame, and N is the number of sampling points per frame; N≥2, is the energy threshold value;

[0031] When is greater than or equal to judging as a speech segment, otherwise as a silent segment.

[0032] Further, the VAD parameter is adjusted in real time according to the feature fluctuation detection of the silent segment, including:

[0033] The VAD parameter includes zero-crossing rate detection threshold, signal-to-noise ratio threshold, hang-up time and duration; the zero-crossing rate detection threshold, signal-to-noise ratio threshold, hang-up time and duration are adjusted by the following relationship:

[0034] ;

[0035] ;

[0036] ;

[0037] ;

[0038] ;

[0039] ;

[0040] ;

[0041] wherein, is an adjusted zero-crossing rate detection threshold, is an initial detection threshold, is a spectral fluctuation rate, is an hth dimension MFCC value of the mth frame, is an hth dimension MFCC value of the (m-1)th frame, K represents a dimension of the MFCC value, and K is 13 or 20, is a constant, = 10 -6 , is a signal-to-noise ratio, is an average energy value of all speech segments, is an average energy value of all silence segments, is an adjusted signal-to-noise ratio threshold, is an initial signal-to-noise ratio threshold, is an adjusted hangover time, is an initial hangover time, is an adjusted duration, is an initial duration, is an energy short-term energy fluctuation rate of the mth frame, is an energy short-term energy fluctuation rate of the (m-1)th frame, is a spectral fluctuation rate adjustment coefficient, is a spectral smoothness adjustment coefficient, is a first signal-to-noise ratio adjustment coefficient, is a second signal-to-noise ratio adjustment coefficient, is an energy fluctuation influence coefficient, is a spectral fluctuation influence coefficient, , , and all have a value range of [0.05, 0.2].

[0042] Further, the method further comprises: recording a time length of the stored model in the voiceprint library, and deleting the stored model when the time length is greater than or equal to a preset maximum time length.

[0043] In another aspect, the present application further provides an adaptive VAD parameter adjustment system based on voiceprint recognition, comprising:

[0044] The acquisition and judgment module is configured to acquire audio signals through an audio acquisition device, preprocess the audio signals and extract voiceprint features, map the voiceprint features to a user voiceprint model, compare the current user voiceprint model with the stored models in the voiceprint database, and determine whether the user is a registered user.

[0045] The adjustment processing module is configured to determine the VAD parameter based on whether the user is a registered user; when the user is determined to be an unregistered user, the VAD parameter is adjusted in real time through feature fluctuation detection of the silent segment and the voice segment.

[0046] The storage management module includes a voiceprint library, and is configured to store the adjusted VAD parameters into the voiceprint library.

[0047] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0048] This invention provides an adaptive VAD parameter adjustment method and system based on voiceprint recognition. By introducing voiceprint recognition technology, it achieves personalized and real-time dynamic adjustment of VAD parameters, effectively overcoming the problems of fixed and unadaptive VAD algorithm parameters in existing technologies. When detecting audio signals, this invention first extracts and compares user voiceprint features to determine if the user is a registered user. In the case of registered users, historically optimized VAD parameters are loaded to ensure the stability and accuracy of detection. In the case of unregistered users, VAD parameters are adjusted in real time through feature fluctuation detection of silence and speech segments, including zero-crossing rate threshold, signal-to-noise ratio threshold, hang time, and duration, significantly improving the sensitivity and accuracy of VAD detection in complex environments.

[0049] This invention introduces an adaptive signal-to-noise ratio (SNR) adjustment mechanism, enabling the system to automatically optimize VAD parameters under different SNR environments, effectively reducing the risk of false positives and false negatives caused by noise interference. Simultaneously, by employing a comprehensive detection method combining short-time energy fluctuation rate and spectral fluctuation rate, the VAD parameters maintain a dynamic and smooth transition between silence and speech segments, reducing the loss or truncation of speech segments and improving the completeness and coherence of speech detection. Furthermore, this invention includes a model maintenance mechanism that records the duration of model storage and periodically cleans up expired models, preventing redundant expansion of the speaker database and maintaining efficient system operation and data validity. Attached Figure Description

[0050] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0051] Figure 1 This is a flowchart of an adaptive VAD parameter adjustment method based on voiceprint recognition, provided as an embodiment of the present invention.

[0052] Figure 2 This is a functional framework diagram of an adaptive VAD parameter adjustment system based on voiceprint recognition, provided in an embodiment of the present invention. Detailed Implementation

[0053] Exemplary embodiments of the present application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0054] See Figure 1 As shown, this embodiment of the invention provides an adaptive VAD parameter adjustment method based on voiceprint recognition, including:

[0055] Audio signals are acquired through audio acquisition devices, preprocessed and voiceprint features are extracted, the voiceprint features are mapped to the user's voiceprint model, and the current user's voiceprint model is compared with the stored model in the voiceprint database to determine whether the user is a registered user.

[0056] The VAD parameter is determined based on whether the user is a registered user; when the user is determined to be an unregistered user, the VAD parameter is adjusted in real time through feature fluctuation detection of silent and speech segments.

[0057] The adjusted VAD parameters are stored in the voiceprint library.

[0058] It should be noted that the aforementioned adaptive VAD parameter adjustment method based on voiceprint recognition combines voiceprint recognition with VAD parameter adjustment to achieve adaptive parameter optimization for different user individual characteristics, effectively overcoming the problem of fixed parameters and decreased detection accuracy in existing VAD algorithms under complex environments. Firstly, this method compares the user model with the stored models in the voiceprint database using voiceprint recognition technology. This automatically identifies registered users and loads historically optimized VAD parameters, thereby maintaining the stability and accuracy of detection parameters in scenarios with registered users and reducing false positives or false negatives caused by individual differences.

[0059] Secondly, in scenarios involving unregistered users, this method employs a feature fluctuation detection mechanism for silence and speech segments, enabling real-time dynamic adjustment of VAD parameters. By detecting changes in short-term energy fluctuation rate and spectral fluctuation rate during silence and speech segments, it automatically optimizes parameters such as the zero-crossing rate detection threshold, signal-to-noise ratio threshold, suspension time, and duration, effectively adapting to changes in user characteristics and the environment. This mechanism improves the sensitivity and adaptability of speech detection in complex scenarios, significantly reduces the interference of background noise on VAD detection, minimizes false positives and false negatives, and ensures the accuracy and completeness of speech detection.

[0060] Furthermore, by storing the adjusted VAD parameters in a speaker database, dynamic parameter updates and model self-optimization are achieved. This allows the system to continuously optimize VAD parameters through repeated use, improving long-term stability and adaptability. Overall, this method effectively improves the accuracy, robustness, and real-time performance of speech detection in complex environments and multi-user scenarios, and is widely applicable to fields such as speech recognition, voice calls, and intelligent voice interaction.

[0061] In some embodiments of this application, when acquiring audio signals via an audio acquisition device, the following is included:

[0062] The audio signal is sampled at the same sampling rate and converted to mono in PCM format.

[0063] In this embodiment, the sampling rate is uniformly set to 16kHz or 48kHz.

[0064] It should be noted that by unifying the sampling rate and audio format, the consistency and high-quality acquisition of the audio signal are ensured, thereby improving the accuracy and robustness of subsequent voiceprint recognition and VAD parameter adjustment. In this embodiment, unifying the sampling rate to 16kHz or 48kHz effectively balances the fidelity of the audio signal and processing efficiency. The 16kHz sampling rate is commonly used in the field of speech processing for speech recognition and voice calls, which can meet the feature extraction requirements of most speech signals, while reducing the amount of data and improving real-time processing performance; while the 48kHz sampling rate is suitable for high-fidelity speech or audio processing, can capture richer spectral information, and is suitable for high-quality speech or audio analysis.

[0065] Converting audio signals to mono simplifies the audio data structure and avoids redundant calculations or feature inconsistencies caused by adjusting VAD parameters in stereo or multi-channel audio, thus improving processing efficiency and model adaptability. Simultaneously, storing audio signals using PCM (Pulse Code Modulation) format ensures lossless compression, maintaining the accuracy and integrity of the original signal and avoiding the loss or distortion of speech features due to lossy compression, thereby improving the accuracy of subsequent speaker feature extraction and model comparison.

[0066] In some embodiments of this application, preprocessing of the audio signal includes:

[0067] S1: Calculate the average value of the audio signal and subtract the average value of the audio signal from each sample point;

[0068] S2: Pre-emphasis processing is performed based on the following relationship:

[0069] ;

[0070] S3: Divide the audio signal into frames every 25ms and set the frame shift to 10ms;

[0071] S4: Window the audio signal based on the Hamming window, and multiply the sample point by the Hamming window function;

[0072] in, and These are the audio signals at the nth sampling point before and after pre-emphasis processing, respectively. The pre-emphasis coefficient is 0.9 < ≤1, This is the audio signal of the (n-1)th sampling point before pre-emphasis processing.

[0073] It should be noted that by calculating the average value of the audio signal during the preprocessing stage and subtracting the average value from each sampling point, DC components and signal offsets can be effectively removed. This ensures that the audio signal is in a zero-mean state for subsequent feature extraction and analysis, thereby improving the stability and consistency of the features. The pre-emphasis processing step uses a first-order high-pass filter to enhance high-frequency components and suppress low-frequency components, thereby improving the clarity and resolution of the speech signal. By setting the range of the pre-emphasis coefficients to 0.9 ≤ α ≤ 1, noise caused by over-filtering can be suppressed while ensuring high-frequency enhancement, thus maintaining signal quality.

[0074] Dividing the audio signal into 25ms frames and setting the frame shift to 10ms helps to balance time resolution and frequency resolution, ensuring that the speech signal can be fully captured in the time and frequency domains, while reducing the signal overlap between adjacent frames and improving the accuracy of speech activity detection.

[0075] Applying a Hamming window to audio signals effectively suppresses spectral leakage and reduces spectral distortion caused by truncation. The Hamming window function has a smaller amplitude at spectral edges, thus smoothing the spectrum in the frequency domain and improving frequency resolution and the accuracy of spectral feature extraction. This preprocessing method significantly enhances the accuracy and robustness of subsequent speaker feature extraction and VAD parameter adjustment.

[0076] In some embodiments of this application, when extracting voiceprint features from an audio signal, the voiceprint features include MFCC value, LPCC value, and fundamental frequency.

[0077] In some embodiments of this application, when comparing the current user's voiceprint model with the stored models in the voiceprint database to determine whether the user is a registered user, the following steps are included:

[0078] The system retrieves and compares the MFCC, LPCC, and fundamental frequency of the current user's voiceprint model with those of each stored model in the voiceprint database. It calculates the similarity, selects the maximum similarity, and compares it with a similarity threshold. If the maximum similarity is greater than or equal to the threshold, the current user model is determined to be a registered user; if the maximum similarity is less than the threshold, the current user model is determined not to be a registered user. The similarity satisfies the following relationship:

[0079] ;

[0080] in, For similarity, Let j be the j-th voiceprint feature of the current user's voiceprint model. To store the j-th voiceprint feature of the model; The number of voiceprint features. =3, when j=1, it corresponds to the MFCC value; when j=2, it corresponds to the LPCC value; when j=3, it corresponds to the baseband.

[0081] It should be noted that by comparing the current user's voiceprint model with the stored models in the voiceprint database, the identity of registered users can be effectively identified. Using MFCC, LPCC, and fundamental frequency as three types of voiceprint features for comparison helps to comprehensively represent the user's speech characteristics. MFCC captures the short-time spectral characteristics of the speech signal, LPCC reflects the linear prediction characteristics of the speech signal, and fundamental frequency describes the pitch characteristics of the speech. The joint comparison of these three features enhances the system's ability to distinguish individual speech characteristics, thereby improving the accuracy and robustness of the comparison.

[0082] Similarity calculation employs either the Euclidean distance formula or cosine similarity, quantifying the differences in feature space between the current user's voiceprint model and the stored model. By selecting the maximum similarity and comparing it with a preset similarity threshold, the system ensures that it identifies a user as a registered user when the similarity between the user model and the stored model is high, and identifies them as an unregistered user when the similarity is low. This threshold-based discrimination method effectively reduces the false recognition rate and the missed recognition rate, guaranteeing the system's recognition accuracy and stability.

[0083] Furthermore, multi-dimensional feature comparison enhances the system's robustness against interference in complex environments. Even with noise or speech variations, the model comparison can still make robust judgments based on comprehensive features, thereby improving the system's adaptability and reliability across various scenarios.

[0084] The current user's voiceprint model is constructed as follows: Using the extracted voiceprint features, a voiceprint model for the current user is built. This model can be a multi-dimensional feature vector or a trained machine learning model (such as a Gaussian Mixture Model (GMM), Support Vector Machine (SVM), or Deep Neural Network (DNN). The voiceprint models of registered users are retrieved from the voiceprint database. The voiceprint database stores the voiceprint models of all registered users, and each user's model contains representative information about their speech features. The model construction is existing technology and will not be elaborated upon here.

[0085] In some embodiments of this application, when it is determined that the user is a registered user, the VAD parameter of the stored model corresponding to the maximum similarity in the voiceprint database is retrieved;

[0086] When it is determined that the user is not a registered user, the preprocessed audio signal is divided into silence and speech segments according to the energy value, and the VAD parameters are adjusted in real time based on the characteristic fluctuation detection of the silence and speech segments.

[0087] In some embodiments of this application, when determining that the user is not a registered user, when dividing the user into a silent segment and a voice segment based on the energy value, the following is included:

[0088] The energy value of each frame of audio signal is calculated, and the segment is divided into a silent segment or a speech segment by comparing the energy value with an energy threshold. The energy value is obtained through the following relationship:

[0089] ;

[0090] ;

[0091] in, The energy value of the m-th frame. Let N be the k-th sampling point of the m-th frame, and N be the number of sampling points per frame; N≥2. Energy threshold;

[0092] when Greater than or equal to If it is determined to be a speech segment, then it is a silent segment.

[0093] It's important to note that separating silence segments from speech segments allows the speech recognition model to focus solely on the speech component, ignoring useless silence or noise data. This helps reduce background noise interference and improves speech recognition accuracy. By preemptively removing silence segments, computational resource consumption for speech-free signals is reduced, thus improving system efficiency. Especially in real-time speech recognition systems, reducing silence segments significantly lowers latency. Dynamic judgment based on real-time energy values ​​and energy thresholds adapts to varying speech intensity in different environments, such as speaker speed and volume, preventing incorrect classification as silence due to weak speech signals. The energy threshold can be adjusted based on different application scenarios (e.g., noisy environments, speaker groups), making the method highly flexible and adaptable to various sound environments. If the energy value calculation method and threshold settings are appropriate, misclassifying silence segments as speech segments can be avoided, further improving the model's ability to recognize unregistered or abnormal users. This method effectively handles various speech signal variations, enhancing the system's ability to distinguish between silence and speech signals, maintaining relatively stable recognition performance even in various noise environments.

[0094] In some embodiments of this application, the VAD parameter is adjusted in real time based on the characteristic fluctuation detection of the silent segment, including:

[0095] VAD parameters include zero-crossing detection threshold, signal-to-noise ratio threshold, hang-up time, and duration; these parameters are adjusted using the following relationships:

[0096] ;

[0097] ;

[0098] ;

[0099] ;

[0100] ;

[0101] ;

[0102] ;

[0103] in, The adjusted zero-crossing rate detection threshold. As the initial detection threshold, For spectral volatility, The h-th dimension MFCC value of the m-th frame. Let K be the h-th dimension MFCC value of the (m-1)-th frame, where K represents the dimension of the MFCC value, and K can be 13 or 20. It is a constant. =10 -6 , For signal-to-noise ratio, The average energy value across all speech segments. The average energy value across all silent segments. This is the adjusted signal-to-noise ratio threshold. This is the initial signal-to-noise ratio threshold. The adjusted suspension time. This is the initial suspension time. The adjusted duration, For the initial duration, Let m be the short-time energy fluctuation rate of the m-th frame. The energy short-time energy fluctuation rate of the (m-1)th frame. This is the spectral volatility adjustment coefficient. This is the spectral stability adjustment coefficient. This is the first signal-to-noise ratio adjustment factor. This is the second signal-to-noise ratio adjustment factor. The coefficient representing the influence of energy fluctuations. This is the spectral fluctuation influence coefficient. , , and The value range is [0.05, 0.2].

[0104] It's important to note the definition of the zero-crossing rate detection threshold: The zero-crossing rate is the frequency at which a signal waveform crosses the zero axis, typically used to measure the "variability" of a signal. In speech signals, the level of the zero-crossing rate reflects the activity level of the speech. Adjustment method: By combining the MFCC features of the current frame with those of the previous frame, the change in MFCC features is calculated, which helps to adjust the zero-crossing rate detection threshold in real time. This allows for dynamic responses to changes in speech and different noise environments.

[0105] Signal-to-Noise Ratio (SNR) Threshold Definition: The SNR reflects the ratio of useful components to noise components in a signal. A higher SNR generally indicates better speech signal quality and lower noise. Adjustment Method: By calculating the average energy of speech segments and silence segments, the SNR threshold is adjusted based on this information, thereby improving signal recognition accuracy in noisy environments. The SNR threshold needs to be dynamically adjusted under different noise conditions.

[0106] Suspension time definition: Suspension time refers to the time the system needs to maintain a "voice activity state" after detecting a speech segment to avoid misjudgment due to brief silences. Adjustment method: The suspension time is dynamically adjusted based on energy fluctuation rate and spectral fluctuation. If the speech signal fluctuates significantly, the suspension time can be appropriately extended to ensure the continuity of speech detection.

[0107] Duration refers to the length of time during which speech activity is considered valid, and is typically used to prevent brief, non-speech disturbances from being misinterpreted as speech. Adjustment methods include using characteristics such as short-term energy fluctuation rate, spectral fluctuation rate, and energy and spectral stability. In particular, the energy fluctuation influence coefficient and the spectral fluctuation influence coefficient can help determine the duration adjustment under different fluctuation conditions.

[0108] In some embodiments of this application, the method further includes: recording the duration for which the stored model is stored in the voiceprint database; and deleting the stored model when the duration is greater than or equal to a preset maximum duration. Additionally, if the current user's voiceprint model is compared with the stored models in the voiceprint database and it is determined to be a non-registered user, the current user's voiceprint model is stored in the voiceprint database.

[0109] See Figure 2 As shown, this embodiment of the invention provides an adaptive VAD parameter adjustment system based on voiceprint recognition, comprising:

[0110] The acquisition and judgment module is configured to acquire audio signals through an audio acquisition device, preprocess the audio signals and extract voiceprint features, map the voiceprint features to the user's voiceprint model, compare the current user's voiceprint model with the stored models in the voiceprint database, and determine whether the user is a registered user.

[0111] The adjustment processing module is configured to determine the VAD parameter based on whether the user is a registered user; when the user is determined to be an unregistered user, the VAD parameter is adjusted in real time through feature fluctuation detection of the silent segment and the voice segment.

[0112] The storage management module has a voiceprint library and is configured to store the adjusted VAD parameters into the voiceprint library.

[0113] In summary, this invention provides an adaptive VAD parameter adjustment method and system based on voiceprint recognition. By introducing voiceprint recognition technology, it achieves personalized and real-time dynamic adjustment of VAD parameters, effectively overcoming the problems of fixed and unadaptive VAD algorithm parameters in existing technologies. When detecting audio signals, this invention first extracts and compares user voiceprint features to determine if the user is a registered user. In the case of registered users, historically optimized VAD parameters are loaded to ensure the stability and accuracy of detection. In the case of unregistered users, VAD parameters are adjusted in real time through feature fluctuation detection of silence and speech segments, including zero-crossing rate threshold, signal-to-noise ratio threshold, hang time, and duration, significantly improving the sensitivity and accuracy of VAD detection in complex environments.

[0114] This invention introduces an adaptive signal-to-noise ratio (SNR) adjustment mechanism, enabling the system to automatically optimize VAD parameters under different SNR environments, effectively reducing the risk of false positives and false negatives caused by noise interference. Simultaneously, by employing a comprehensive detection method combining short-time energy fluctuation rate and spectral fluctuation rate, the VAD parameters maintain a dynamic and smooth transition between silence and speech segments, reducing the loss or truncation of speech segments and improving the completeness and coherence of speech detection. Furthermore, this invention includes a model maintenance mechanism that records the duration of model storage and periodically cleans up expired models, preventing redundant expansion of the speaker database and maintaining efficient system operation and data validity.

[0115] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. An adaptive VAD parameter adjustment method based on voiceprint recognition, characterized in that, include: Audio signals are acquired through audio acquisition devices, preprocessed and voiceprint features are extracted from the audio signals, the voiceprint features are mapped to the user's voiceprint model, and the current user's voiceprint model is compared with the stored model in the voiceprint database to determine whether the user is a registered user. The VAD parameter is determined based on whether the user is a registered user. When the user is determined to be an unregistered user, the VAD parameter is adjusted in real time by detecting the characteristic fluctuations of the silent segment and the voice segment. The adjusted VAD parameters are stored in the voiceprint library; When comparing the current user's voiceprint model with the stored models in the voiceprint database to determine whether the user is a registered user, the process includes: The MFCC value, LPCC value, and fundamental frequency of the current user's voiceprint model are obtained and compared with those of each stored model in the voiceprint database. A similarity score is calculated, and the maximum similarity score is selected. This maximum similarity score is then compared with a similarity threshold. When the maximum similarity score is greater than or equal to the similarity threshold, the current user model is determined to be a registered user; when the maximum similarity score is less than the similarity threshold, the current user model is determined not to be a registered user. The similarity score satisfies the following relationship: ； in, For similarity, Let j be the j-th voiceprint feature of the current user's voiceprint model. To store the j-th voiceprint feature of the model; The number of voiceprint features. =3, when j=1, it corresponds to the MFCC value; when j=2, it corresponds to the LPCC value; when j=3, it corresponds to the baseband. When the user is identified as a registered user, the VAD parameter of the stored model corresponding to the highest similarity in the voiceprint database is retrieved. When it is determined that the user is not a registered user, the preprocessed audio signal is divided into a silence segment and a speech segment according to the energy value, and the VAD parameter is adjusted in real time according to the characteristic fluctuation detection of the silence segment and the speech segment. When a user is identified as unregistered, and when the user is categorized into a silent segment and a voice segment based on their energy level, the following applies: The energy value of each frame of audio signal is calculated, and the segment is divided into a silent segment or a speech segment by comparing the energy value with an energy threshold. The energy value is obtained through the following relationship: ； ； in, The energy value of the m-th frame. Let N be the k-th sampling point of the m-th frame, and N be the number of sampling points per frame; N≥2. Energy threshold; when Greater than or equal to If it is determined to be a speech segment, then it is a silent segment; otherwise, it is a speech segment. When preprocessing audio signals, the following are included: S1: Calculate the average value of the audio signal and subtract the average value of the audio signal from each sample point; S2: Pre-emphasis processing is performed based on the following relationship: ； S3: Divide the audio signal into frames every 25ms and set the frame shift to 10ms; S4: Window the audio signal based on the Hamming window, and multiply the sample point by the Hamming window function; in, and These are the audio signals at the nth sampling point before and after pre-emphasis processing, respectively. The pre-emphasis coefficient is 0.9 < ≤1, This refers to the audio signal at the (n-1)th sampling point before pre-emphasis processing; The VAD parameters are adjusted in real time based on the characteristic fluctuation detection of the silent segment, including: The VAD parameters include zero-crossing rate detection threshold, signal-to-noise ratio threshold, suspension time, and duration; the zero-crossing rate detection threshold, signal-to-noise ratio threshold, suspension time, and duration are adjusted according to the following relationship: ； ； ； ； ； ； ； in, The adjusted zero-crossing rate detection threshold. As the initial detection threshold, For spectral volatility, The h-th dimension MFCC value of the m-th frame. Let K be the h-th dimension MFCC value of the (m-1)-th frame, where K represents the dimension of the MFCC value, and K can be 13 or 20. It is a constant. =10 -6 , For signal-to-noise ratio, The average energy value across all speech segments. The average energy value across all silent segments. This is the adjusted signal-to-noise ratio threshold. This is the initial signal-to-noise ratio threshold. The adjusted suspension time. This is the initial suspension time. The adjusted duration, For the initial duration, Let m be the short-time energy fluctuation rate of the m-th frame. The energy short-time energy fluctuation rate of the (m-1)th frame. This is the spectral volatility adjustment coefficient. This is the spectral stability adjustment coefficient. This is the first signal-to-noise ratio adjustment factor. This is the second signal-to-noise ratio adjustment factor. The coefficient representing the influence of energy fluctuations. This is the spectral fluctuation influence coefficient. , , and The values ​​range for all values ​​are [0.05, 0.2]. It also includes: recording the duration for which the storage model is stored in the voiceprint library, and deleting the storage model when the duration is greater than or equal to a preset maximum duration.

2. The adaptive VAD parameter adjustment method based on voiceprint recognition according to claim 1, characterized in that, When acquiring audio signals through an audio acquisition device, this includes: The audio signal is sampled at the same sampling rate and converted to mono in PCM format.

3. The adaptive VAD parameter adjustment method based on voiceprint recognition according to claim 1, characterized in that, When extracting voiceprint features from an audio signal, the voiceprint features include MFCC value, LPCC value, and fundamental frequency.

4. An adaptive VAD parameter adjustment system based on voiceprint recognition, applied to the adaptive VAD parameter adjustment method based on voiceprint recognition as described in any one of claims 1-3, characterized in that, include: The acquisition and judgment module is configured to acquire audio signals through an audio acquisition device, preprocess the audio signals and extract voiceprint features, map the voiceprint features to a user voiceprint model, compare the current user voiceprint model with the stored models in the voiceprint database, and determine whether the user is a registered user. The adjustment processing module is configured to determine the VAD parameter based on whether the user is a registered user; when the user is determined to be an unregistered user, the VAD parameter is adjusted in real time through feature fluctuation detection of the silent segment and the voice segment. The storage management module includes a voiceprint library, and is configured to store the adjusted VAD parameters into the voiceprint library.

Images