Voiceprint processing method and system and storage medium

Abstract

The invention provides a voiceprint processing method and system and a storage medium, and relates to the technical field of voice processing, and the voiceprint processing method comprises the steps: obtaining audio sample data; performing noise separation on the audio sample data to obtain a voice signal; performing voiceprint extraction on the voice signal; inputting the extracted target voiceprint features into a selected clustering algorithm; comparing the voiceprint feature points of the voiceprint cluster with voiceprint features in a pre-constructed voiceprint library, and calculating the similarity; and selecting the voiceprint cluster of the corresponding speaker to carry out gain processing. According to the method, voiceprints with similar characteristics are clustered through the clustering algorithm, the subsequent comparison process is simplified, the recognition accuracy is improved, the identity of the speaker is determined based on the voiceprint clustering cluster, gain processing is carried out on the voiceprint clustering cluster of a specific speaker, the system can more accurately extract and recognize the voiceprint characteristics of the speaker, and the recognition accuracy is improved. And high identification accuracy can be maintained even in a noisy environment.

Description

Technical Field

[0001] The present invention relates to the field of speech processing technology, and in particular to a voiceprint processing method, system and storage medium. Background Art

[0002] With the development of society and the popularization of electronic products, people have higher and higher requirements for voice quality. How to improve the mobile communication quality of electronic products in noisy environments has become the most popular research direction. Speech enhancement can improve the quality and intelligibility of speech in noisy environments. Speech enhancement has important applications not only in the fields of hearing aids and cochlear implants.

[0003] Speaker recognition, like fingerprint, iris, and face recognition, is a type of biometric recognition and is considered the most natural way of biometric identification, also known as "voiceprint" recognition. In voiceprint recognition systems, speech enhancement technology also plays an important role. By enhancing speech signals, the accuracy of voiceprint feature extraction can be improved, thereby improving the accuracy of voiceprint recognition. In addition, speech enhancement technology can also help reduce the impact of noise on voiceprint features and improve the performance of voiceprint recognition systems in noisy environments.

[0004] At present, traditional voiceprint feature gain often adopts a global processing method, that is, uniformly gaining or optimizing all voiceprint features. Although this method can improve the effect of voiceprint recognition to a certain extent, it does not fully consider the differences and importance of different voiceprint features in the recognition process. Therefore, a voiceprint processing method that can gain the voiceprint features of specific people separately is needed, so as to further improve the voiceprint recognition performance in noisy environments. Summary of the invention

[0005] The main purpose of the present invention is to propose a voiceprint processing method, system and storage medium, aiming to solve the problem that traditional voiceprint feature processing methods cannot individually gain the voiceprint features of a specific person.

[0006] In view of the above problems, the present invention proposes a voiceprint processing method, comprising: Get audio sample data; Perform noise separation on the audio sample data to obtain a speech signal; Extract voiceprint from speech signals to obtain target voiceprint features; Input the extracted target voiceprint features into the selected clustering algorithm to obtain voiceprint clusters; Compare the voiceprint feature points of the voiceprint cluster with the voiceprint features in the pre-built voiceprint library and calculate the similarity to determine the identity of the corresponding speaker; Based on the determined speaker identity, in response to the received instruction, a voiceprint cluster corresponding to the speaker is selected for gain processing.

[0007] Optionally, the specific step of performing noise separation on the audio sample data to obtain the speech signal includes: Preprocessing audio sample data; Perform feature extraction on the preprocessed audio sample data to obtain human voice audio features; Based on the audio characteristics of human voice, a pure noise sample is obtained from the preprocessed audio sample data; Based on the pure noise samples, the audio sample data is noise separated to obtain the speech signal.

[0008] Optionally, the specific step of obtaining a pure noise sample from the preprocessed audio sample data based on the human voice audio feature includes: According to the human voice audio features, the preprocessed audio sample data is divided into sound segments with human voice audio features and sound segments without human voice audio features; Extracting the audio segments without human voice audio features from the preprocessed audio sample data; The sound segment without human voice audio features is a pure noise sample.

[0009] Optionally, the specific step of inputting the extracted target voiceprint features into a selected clustering algorithm to obtain voiceprint clusters includes: The extracted target voiceprint features are input into a selected clustering algorithm, and the clustering algorithm clusters similar voiceprint feature points in the target voiceprint features into voiceprint clusters.

[0010] Optionally, the voiceprint feature points of the same voiceprint cluster are voiceprint feature points of the same speaker.

[0011] Optionally, the specific steps of comparing the voiceprint feature points of the voiceprint cluster with the voiceprint features in the pre-built voiceprint library and calculating the similarity to determine the identity of the corresponding speaker include: Match the voiceprint feature points of each voiceprint cluster with the voiceprint features in the pre-built voiceprint library one by one; For each voiceprint feature point of the voiceprint cluster, calculate its similarity with the voiceprint feature of each speaker in the pre-built voiceprint library; When the similarity exceeds a preset threshold, the voiceprint feature points of the voiceprint cluster are successfully matched with the current voiceprint feature, and the speaker of the voiceprint cluster is the speaker of the current voiceprint feature; When the similarity does not exceed the preset threshold, the voiceprint feature points of the voiceprint cluster fail to match the current voiceprint feature, and the speaker of the voiceprint cluster is not the speaker of the current voiceprint feature.

[0012] Optionally, the specific step of selecting a voiceprint cluster corresponding to the speaker for gain processing based on the determined speaker identity and in response to a received instruction includes: Responding to the received instruction based on the determined speaker identity, and parsing the received instruction to extract speaker information; According to the parsed speaker information, searching and selecting the voiceprint cluster corresponding to the speaker from the determined speaker identities; Perform gain processing on the selected voiceprint cluster.

[0013] Optionally, the specific step of performing gain processing on the selected voiceprint cluster includes: The voiceprint feature points of the selected voiceprint cluster are input into the gain processing algorithm to obtain the gained voiceprint feature points and the gained voiceprint cluster.

[0014] The present invention also proposes a voiceprint processing system, comprising a processor and a memory, wherein the memory stores machine-readable instructions executable by the processor, and when the machine-readable instructions are executed by the processor, the steps of the voiceprint processing method as described above are performed.

[0015] The present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the voiceprint processing method as described above are executed.

[0016] The present invention clusters voiceprints with similar features into clusters through a clustering algorithm, simplifies the subsequent comparison process, improves recognition accuracy, and determines the identity of the speaker based on the voiceprint clustering clusters. By performing gain processing on the voiceprint clustering clusters of a specific speaker, the system can more accurately extract and identify the voiceprint features of the speaker, and maintain a high recognition accuracy even in a noisy environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying creative work.

[0018] Figure 1 A flow chart of the steps of a voiceprint processing method of the present invention; Figure 2 This is a flow chart of step S200 of a voiceprint processing method of the present invention; Figure 3 This is a flow chart of the steps of S230 of a voiceprint processing method of the present invention; Figure 4 This is a flow chart of step S500 of a voiceprint processing method of the present invention; Figure 5 This is a flow chart of step S600 of a voiceprint processing method of the present invention.

[0019] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0020] 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.

[0021] It should be noted that all directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of the present invention are only used to explain the relative position relationship, movement status, etc. between the components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly.

[0022] In the present invention, unless otherwise clearly specified and limited, the terms "connection", "fixation", etc. should be understood in a broad sense. For example, "fixation" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements, unless otherwise clearly defined. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0023] In addition, in the present invention, descriptions such as "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In addition, the technical solutions between the various embodiments can be combined with each other, but they must be based on the ability of ordinary technicians in the field to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0024] The present invention proposes a voiceprint processing method, such as Figure 2 As shown, including: S100: Acquire audio sample data; S200: performing noise separation on the audio sample data to obtain a speech signal; S300: extracting voiceprint from the speech signal to obtain target voiceprint features; S400: inputting the extracted target voiceprint features into the selected clustering algorithm to obtain voiceprint clusters; S500: comparing the voiceprint feature points of the voiceprint cluster with the voiceprint features in the pre-built voiceprint library and calculating the similarity to determine the corresponding speaker identity; S600: Based on the determined speaker identity, in response to the received instruction, select a voiceprint cluster corresponding to the speaker for gain processing.

[0025] Voiceprint recognition is a technology used to verify personal identity. It uses language characteristics to identify the speaker's voice. From the perspective of application scenarios, it can be divided into two categories: speaker confirmation and speaker identification. The former usually selects and confirms among known speakers. It is a binary classification problem, that is, to determine whether a piece of speech belongs to a known speaker. The latter is a multi-classification problem, that is, to determine which known speaker the speaker in a piece of speech belongs to.

[0026] At present, traditional voiceprint feature gain often adopts a global processing method, that is, uniformly gaining or optimizing all voiceprint features. Whether it is speaker confirmation or speaker recognition, the traditional global voiceprint feature gain method has its limitations. This method fails to fully consider the uniqueness and importance of different voiceprint features in the recognition process, which may lead to poor recognition effect in some complex or noisy environments. Therefore, the present invention proposes a voiceprint processing method that can individually gain the voiceprint features of specific people, thereby further improving the voiceprint recognition performance in noisy environments.

[0027] The voiceprint processing method of the present invention is described more specifically: In step S100, audio sample data is obtained, and the audio sample data should include human voice features and background noise.

[0028] In step S200, the audio sample data is subjected to noise separation to obtain a speech signal; a speech signal is a periodic signal that varies with time, which contains the speaker's identity information and the transmitted text information. Based on this, by extracting features from the speech signal, the speaker's identity feature information can be extracted to realize voiceprint recognition. However, due to the influence of the speaker's real environment, the components contained in the speech are very complex. In addition to the pauses and silent segments caused by the speaker during the expression process, it is also interfered by background noise. The actual collected speech often contains many blank parts and noise parts, which are all irrelevant to voiceprint recognition. Therefore, the speech data should be preprocessed first. The speech signal preprocessing method generally includes multiple processes such as pre-emphasis, framing, and windowing.

[0029] After preprocessing, the background noise is estimated before the speech signal starts or in the silent segment, which can be done by calculating the average energy or power spectral density in these segments. The estimated noise information is used to suppress or reduce the noise in the actual speech signal. Common methods include spectral subtraction, Wiener filtering, Kalman filtering, etc. At the same time, the speech signal after noise suppression is post-processed, such as de-jittering, smoothing, etc., to improve the speech quality.

[0030] In step S300, signal processing technology is used to extract features from the speech signal, such as Mel-frequency cepstral coefficients (MFCC), linear prediction coefficients (LPC), spectrum envelope, etc.

[0031] As a commonly used feature extraction method in the field of voiceprint recognition and speech recognition, Mel-frequency cepstral coefficients (MFCC) are designed based on the human ear's perception of sound pitch and are considered to be an efficient feature extraction method. Based on the human ear's ability to perceive speech signals, a filter group that simulates the human ear function has emerged, called the Mel frequency filter. The Mel frequency filter group is usually composed of a series of triangular filters, which converts the frequency axis of the audio signal into the Mel frequency axis, so that the division on the frequency axis is more in line with the characteristics of human ear perception.

[0032] LPC, or the coefficient in Linear Predictive Coding, is a key parameter in a method for speech signal compression and parameter extraction. LPC is based on the self-predictability of the signal, that is, the current signal sample can be approximated by a linear combination of several previous samples. By solving the autoregressive model (AR model), the linear combination coefficient that minimizes the prediction error of the current sample is obtained, which is the LPC coefficient.

[0033] Select the most useful features for voiceprint recognition from the extracted features. This can be done through feature dimension reduction techniques (such as principal component analysis PCA) or feature selection algorithms. Normalize the selected features to ensure that they are comparable across different speakers and under different conditions. Feature normalization is the process of scaling feature values ​​to a uniform range. Since different features may have different dimensions and value ranges, if used directly for recognition, it may cause some features to have too much weight in the recognition process. Therefore, the features need to be normalized to ensure that they have equal weight in the recognition process.

[0034] In step S400, the extracted target voiceprint features are input into the selected clustering algorithm to obtain voiceprint clusters. Clustering algorithm is a method in unsupervised learning, which is used to divide objects in a data set into several groups (or "clusters"), so that the similarity between objects in the same group is high, while the similarity between objects in different groups is low. Commonly used clustering algorithms include K-means, hierarchical clustering, DBSCAN, etc. These algorithms can classify similar feature vectors into one category to form voiceprint clusters.

[0035] In step S500, the voiceprint feature points of the voiceprint cluster are compared with the voiceprint features in the pre-built voiceprint library and the similarity is calculated to determine the corresponding speaker identity. The extracted target voiceprint features are grouped by a clustering algorithm to form voiceprint clusters. Each voiceprint cluster represents a group of speaker voiceprint features with similar acoustic characteristics.

[0036] The pre-built voiceprint library can be extracted and stored from the speaker's voice sample in the previous voiceprint processing, or an existing pre-built voiceprint library can be obtained through sharing, etc. Furthermore, the voiceprint library can be built by storing the identity information of relevant personnel in advance.

[0037] It is worth noting that the speaker's voiceprint characteristics may change over time. Therefore, the pre-built voiceprint library needs to be updated regularly to ensure the accuracy and reliability of the recognition system.

[0038] In step S600, based on the determined speaker identity, in response to the received instruction, the voiceprint cluster corresponding to the speaker is selected for gain processing. This process improves the performance of the system in subsequent processing or application by enhancing or optimizing the voiceprint features.

[0039] For example, in the field of hearing aids, users can issue commands through user buttons, voice or mobile phone programs. The system receives the command information through the corresponding wired or wireless module and parses and executes it. The purpose of the command is to select a determined speaker through the user and perform gain processing on the corresponding voiceprint cluster. This helps to highlight the speaker's voice in a noisy environment and reduce interference from other noises, thereby improving the user's clarity and recognition of the speaker's voice.

[0040] In the field of voiceprint recognition, gain-processed voiceprint features can be used for more accurate voiceprint recognition; in applications such as speech recognition and speech synthesis, gain processing can improve the quality and clarity of speech; in the security field, gain processing can enhance the reliability of voiceprint features in the identity authentication process.

[0041] The present invention clusters voiceprints with similar features into clusters through a clustering algorithm, simplifies the subsequent comparison process, improves recognition accuracy, and determines the identity of the speaker based on the voiceprint clustering clusters. By performing gain processing on the voiceprint clustering clusters of a specific speaker, the system can more accurately extract and identify the voiceprint features of the speaker, and maintain a high recognition accuracy even in a noisy environment.

[0042] In one embodiment, if Figure 2 As shown, the specific steps of performing noise separation on the audio sample data to obtain the speech signal include: S210: Preprocessing the audio sample data; Speech signal preprocessing methods generally include multiple processes such as pre-emphasis, framing, and windowing. Pre-emphasis, as a signal processing technology, is used to enhance the energy of the high-frequency part of the speech signal, making the high-frequency part more prominent and thus improving the quality and recognizability of the signal.

[0043] Since there is a correlation between frequency and time in speech signals, the frequency characteristics of the signal will change over time. Therefore, in order to convert the speech signal into a signal that is short in time and stable in frequency for subsequent processing and analysis, the speech signal needs to be framed after pre-emphasis. First, select the appropriate frame length according to the specific application needs, which is generally 10-30 milliseconds. Then select the frame shift so that there is a certain overlap between adjacent frames to ensure the continuity of the analysis sequence.

[0044] Framing speech signals often leads to fluctuations in the frequency domain, while windowing can make the frequency characteristics smoother, reduce frequency leakage caused by framing, and improve the problem of discontinuity of speech signals caused by framing.

[0045] S220: extracting features from the preprocessed audio sample data to obtain human voice audio features; After preprocessing, the voice signal needs to be extracted for feature parameters. The quality of the feature parameters is directly related to the performance of the voiceprint recognition system. It is worth noting that the human voice audio features in this step are different from the subsequent voiceprint feature extraction.

[0046] Human voice audio features refer to acoustic characteristics directly extracted from the original speech signal without further processing or conversion. These features usually come from the original speech waveform data and are initially extracted through signal processing techniques (such as Fourier transform, filtering, etc.). Human voice audio features may contain rich information, such as frequency, amplitude, phase, etc., but are usually relatively primitive and complex. These features may be affected by factors such as noise and environmental interference, and require further preprocessing and feature extraction before they can be used for voiceprint recognition.

[0047] The human voice audio feature is a direct reflection of the original speech signal, while the voiceprint feature is a feature vector that has been further processed and extracted to characterize the speaker's unique acoustic characteristics. The human voice audio feature may contain a lot of redundant information and noise, while the voiceprint feature is more refined and accurate.

[0048] S230: Based on the human voice audio feature, obtain a pure noise sample from the preprocessed audio sample data; The preprocessed audio sample data is analyzed using human voice audio features. By comparing these features with preset thresholds (thresholds are set based on the statistical characteristics of human voice audio features, and these thresholds are used to determine whether the features in the audio sample belong to human voice audio features), the algorithm can determine which segments contain human voice audio features and which segments do not. Segments that do not contain human voice audio features are silent segments (pure noise samples).

[0049] Extracting silent segments (pure noise samples) is of great significance in audio processing. First, these samples can be used in subsequent noise suppression and speech enhancement tasks to help the algorithm more accurately identify and remove background noise, thereby improving the clarity and intelligibility of speech. Second, these samples can also be used to train noise classification models to improve the algorithm's ability to recognize different types of noise. In addition, silent segments can also be used as comparison samples to evaluate the performance differences of the algorithm under sound and silence conditions.

[0050] S240: Based on the pure noise sample, the audio sample data is subjected to noise separation to obtain a speech signal, which is generally referred to as speech enhancement or noise suppression. Noise features are extracted from the pure noise sample, and these features may be spectral features, time domain features, or statistical features.

[0051] The extracted features should be able to fully characterize the characteristics and distribution of the noise.

[0052] The extracted noise features are used to build a noise model. This model can be a statistical model, a machine learning model, or a deep learning model. The role of the noise model is to predict and estimate the noise part of the input audio sample.

[0053] The noise model is used to estimate the noise part of the input audio sample, and the estimated noise part is subtracted from the input audio sample to obtain a pure speech signal. This process involves complex signal processing algorithms, such as spectral subtraction, Wiener filtering, and minimum mean square error estimation.

[0054] In one embodiment, if Figure 3 As shown, based on the human voice audio features, the specific steps of obtaining a pure noise sample from the preprocessed audio sample data include: S231: dividing the preprocessed audio sample data into sound segments with human voice audio features and sound segments without human voice audio features according to human voice audio features; In this step, the system uses human voice audio features (such as fundamental frequency, formant, Mel-frequency cepstral coefficients, etc.) to analyze the preprocessed audio sample data. By comparing these features with the preset threshold or model, the algorithm can determine which segments contain human voice audio features and which segments do not. By comparison, the algorithm can determine which segments in the audio sample contain significant human voice audio features (i.e., voiced segments) and which segments do not (i.e., silent segments). It is worth noting that the purpose of this step is not to completely strip out the voiced segments, but to focus on extracting those segments that do not contain human voice audio features, i.e., pure noise or background noise samples.

[0055] S232: extracting a sound segment without a human voice audio feature from the preprocessed audio sample data; After determining which audio segments contain human voice audio features, the next step is to extract those audio segments that do not contain human voice audio features. These audio segments are usually considered to be pure noise or background noise because they do not contain significant human voice audio features.

[0056] S233: The sound segment without human voice audio features is a pure noise sample.

[0057] The last step is to mark the extracted sound segments that do not contain human voice audio features as pure noise samples. These samples can be used for subsequent analysis, such as noise suppression, speech enhancement and other tasks, and can also be used to train noise classification models or as a reference for background noise.

[0058] In one embodiment, the specific steps of inputting the extracted target voiceprint features into the selected clustering algorithm to obtain the voiceprint clusters include: The extracted target voiceprint features are input into the selected clustering algorithm, and the clustering algorithm clusters similar voiceprint feature points in the target voiceprint features into voiceprint clusters.

[0059] Common clustering algorithms include K-means, hierarchical clustering, DBSCAN, etc.

[0060] K-Means algorithm is one of the most popular clustering algorithms. It iteratively assigns data points to K clusters so that the data points within a cluster are as similar as possible and the clusters are as different as possible. Hierarchical Clustering achieves clustering by creating a tree structure (called a dendrogram) consisting of hierarchical nested clusters. It can be agglomerated (starting from a single data point and gradually merging into clusters) or divisive (starting from all data points in one cluster and gradually splitting into smaller clusters). DBSCAN (Density-Based Spatial Clustering of Applications with Noise) is a density-based clustering algorithm that can identify clusters of arbitrary shapes and can identify noise points. The core idea is to divide areas with sufficiently high density into clusters and allow gaps between clusters.

[0061] The clustering algorithm clusters similar voiceprint feature points in the target voiceprint features into voiceprint clustering algorithms, wherein the output clustering results include the number of clusters, the feature points contained in each cluster, and the like.

[0062] The application of clustering algorithms enables the voiceprint recognition system to process a large number of voice signals more efficiently. By clustering similar voiceprint feature points into clusters, the system reduces unnecessary one-by-one comparison processes, thereby significantly improving the processing speed. This not only enables the system to respond to user requests faster, but also reduces the consumption of computing resources and storage space.

[0063] In one embodiment, the voiceprint feature points of the same voiceprint cluster are the voiceprint feature points of the same speaker. If multiple voiceprint feature points are clustered into the same cluster, then the distance between these feature points in the feature space is relatively close, indicating that they have a high similarity. Ideally, these similar voiceprint feature points should come from the same speaker. It should be noted that the clustering algorithm is not always able to perfectly cluster the voiceprint feature points of the same speaker together. This may be due to the following reasons: Inaccuracy of feature extraction: If there are errors or noise in the voiceprint feature extraction process, the extracted feature points may not accurately reflect the real voiceprint characteristics of the speaker.

[0064] Limitations of clustering algorithms: Different clustering algorithms have different characteristics and limitations. For example, the K-means algorithm is sensitive to the selection of the initial cluster center, while the DBSCAN algorithm is sensitive to parameters such as the neighborhood radius and the minimum number of points.

[0065] Complexity of data distribution: The distribution of voiceprint features in the feature space may be complex, with overlap or ambiguity. This may make it difficult for the clustering algorithm to accurately cluster the voiceprint feature points of the same speaker together.

[0066] To address the above issues, the following measures can be taken: Optimize feature extraction algorithm: Improve the accuracy and robustness of voiceprint feature extraction to reduce the impact of errors and noise.

[0067] Select a suitable clustering algorithm: Select a suitable clustering algorithm based on the actual application scenario and data characteristics, and adjust the algorithm parameters to optimize the clustering effect.

[0068] Introduce post-processing steps: Introduce post-processing steps based on the clustering results, such as merging similar clusters, removing outliers, etc., to further improve the accuracy of the clustering results.

[0069] Continuous learning and updating: As new voiceprint data is continuously added, the voiceprint library and clustering model are continuously learned and updated to adapt to changes.

[0070] In one embodiment, if Figure 4 As shown, the specific steps of comparing the voiceprint feature points of the voiceprint cluster with the voiceprint features in the pre-built voiceprint library and calculating the similarity to determine the identity of the corresponding speaker include: S510: Matching the voiceprint feature points of each voiceprint cluster with the voiceprint features in the pre-built voiceprint library one by one; The pre-built voiceprint library is a key component of the voiceprint recognition system. It stores voiceprint feature templates of multiple known speakers, which are extracted and stored from the speaker's voice samples in the previous stage. They are used to match the voiceprint features to be recognized to determine the speaker's identity.

[0071] The larger the size of the voiceprint database, the more known speakers it contains, and the higher the accuracy of voiceprint recognition. However, as the size of the voiceprint database increases, the computational complexity and storage requirements of the matching process will also increase accordingly.

[0072] S520: For each voiceprint feature point of each voiceprint cluster, calculate its similarity with the voiceprint feature of each speaker in the pre-built voiceprint library; In this step, the system will use a similarity measurement method (such as Euclidean distance, cosine similarity, dynamic time warping, etc.) to calculate the similarity between the voiceprint feature points in the voiceprint cluster and each voiceprint feature template in the pre-built voiceprint library. The choice of similarity measurement method depends on the type of voiceprint features used and the performance requirements of the recognition system.

[0073] Among them, the Euclidean distance is suitable for feature spaces with clear geometric meanings, such as voiceprint recognition based on spectral features. It is simple to calculate, but may be affected by the "dimensionality curse" when processing high-dimensional data.

[0074] Cosine similarity is suitable for measuring the directional similarity between feature vectors, rather than just the distance. It performs well when processing text data or features with directional characteristics. In voiceprint recognition, when the feature vector represents a spectrum or cepstrum feature, cosine similarity is also a good choice.

[0075] Dynamic Time Warping (DTW) is suitable for processing speech signals with time variability. It can align speech signals of different lengths through nonlinear mapping to calculate the similarity between them. It is particularly useful when processing speakers with different speaking speeds or pronunciation habits.

[0076] Mel-frequency cepstral coefficient (MFCC) distance, a feature designed for speech signals, is usually used in conjunction with DTW. The similarity between speech signals is evaluated by calculating the distance between MFCC feature vectors.

[0077] There are other methods including machine learning methods such as support vector machines (SVM) and neural networks, which can learn and extract more complex feature representations to more accurately evaluate the similarity between voiceprints.

[0078] S530: When the similarity exceeds a preset threshold, the voiceprint feature points of the voiceprint cluster are successfully matched with the current voiceprint feature, and the speaker of the voiceprint cluster is the speaker of the current voiceprint feature; If the similarity between a certain voiceprint feature point in the voiceprint cluster and a certain voiceprint feature template in the pre-built voiceprint library exceeds the preset threshold, the system considers that the voiceprint feature point matches the voiceprint feature template successfully. At this time, the system will identify the speaker of the voiceprint cluster as the speaker corresponding to the voiceprint feature template.

[0079] S540: When the similarity does not exceed the preset threshold, the voiceprint feature points of the voiceprint cluster fail to match the current voiceprint feature, and the speaker of the voiceprint cluster is not the speaker of the current voiceprint feature.

[0080] If the similarity between a certain voiceprint feature point in the voiceprint cluster and all voiceprint feature templates in the pre-built voiceprint library does not exceed the preset threshold, the system considers that this voiceprint feature point does not match any known voiceprint feature template. At this time, the system may identify the speaker of the voiceprint cluster as an unknown speaker, or take other measures according to the specific design of the system.

[0081] It should be noted that the preset similarity threshold is a key parameter that directly affects the accuracy and performance of voiceprint recognition. Setting the threshold too high may increase the false recognition rate (i.e., voiceprints of different speakers are mistakenly recognized as the same speaker), while setting the threshold too low may increase the rejection rate (i.e., voiceprints of the same speaker are mistakenly recognized as different speakers or unknown speakers). Therefore, in practical applications, it is necessary to reasonably set the similarity threshold according to the specific needs and performance requirements of the system.

[0082] In one embodiment, if Figure 5 As shown, based on the determined speaker identity, in response to the received instruction, the specific steps of selecting the voiceprint cluster corresponding to the speaker for gain processing include: S610: responding to a received instruction based on the determined speaker identity, and parsing the received instruction to extract speaker information; In this step, the system has already determined the identity of the speaker through the previous voiceprint recognition process. When the system receives a command related to the speaker, it will first parse the command to extract the speaker information contained in it. This information may include the speaker's ID, name, authority level, or other data related to the speaker's identity.

[0083] The system may need access to a command parser or natural language processing module.

[0084] S620: searching and selecting a voiceprint cluster corresponding to the speaker from the determined speaker identities according to the parsed speaker information; After determining the speaker information in the instruction, the system will search for the voiceprint cluster corresponding to the speaker identity in its internal data structure. This data structure may be a database, hash table or other form of data storage, which saves the mapping relationship between the speaker identity and the voiceprint cluster. By searching this data structure, the system can quickly locate the voiceprint cluster that matches the speaker information in the instruction.

[0085] When searching for voiceprint clusters, the system may need to access a database management system.

[0086] S630: Perform gain processing on the selected voiceprint cluster.

[0087] Once a voiceprint cluster matching the speaker information in the command is selected, the system may perform gain processing on this cluster. Gain processing is a signal processing technique that aims to enhance the strength and / or clarity of a specific signal (in this case, the voiceprint feature signal) so that it can be more easily detected and identified in subsequent processing or recognition stages. In a voiceprint recognition system, gain processing may include increasing the amplitude of the voiceprint feature, adjusting the frequency response, reducing noise interference, etc.

[0088] In one embodiment, the specific steps of performing gain processing on the selected voiceprint cluster include: The voiceprint feature points of the selected voiceprint cluster are input into the gain processing algorithm to obtain the gained voiceprint feature points and the gained voiceprint cluster.

[0089] First, the system extracts all the voiceprint feature points from the selected voiceprint cluster. Next, the system inputs these voiceprint feature points into the gain processing algorithm. The gain processing algorithm is a signal processing technique that aims to enhance the strength and / or clarity of a specific signal. In the voiceprint recognition system, the gain processing algorithm may include a variety of techniques, such as amplitude gain, frequency enhancement, noise suppression, etc.

[0090] After being processed by the gain processing algorithm, the system will obtain a set of gain voiceprint feature points. These feature points are enhanced in signal strength or clarity, making them easier to detect and identify in subsequent processing or recognition stages. At the same time, since these feature points still belong to the same voiceprint cluster, they can also be combined into a gain voiceprint cluster.

[0091] The present invention clusters voiceprints with similar features into clusters through a clustering algorithm, simplifies the subsequent comparison process, improves recognition accuracy, and determines the identity of the speaker based on the voiceprint clustering clusters. By performing gain processing on the voiceprint clustering clusters of a specific speaker, the system can more accurately extract and identify the voiceprint features of the speaker, and maintain a high recognition accuracy even in a noisy environment.

[0092] The present invention also provides a voiceprint processing system, including a processor and a memory, wherein the memory stores machine-readable instructions executable by the processor, and when the machine-readable instructions are executed by the processor, the steps of the voiceprint processing method as described above are performed.

[0093] The present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of the voiceprint processing method as described above are executed.

[0094] The above descriptions are only optional embodiments of the present invention, and are not intended to limit the patent scope of the present invention. All equivalent structural changes made using the contents of the present invention's specification and drawings, or directly / indirectly applied in other related technical fields, are included in the patent protection scope of the present invention.

Claims

1. A voiceprint processing method, characterized in that: include: Get audio sample data; Perform noise separation on the audio sample data to obtain a speech signal; Extract voiceprint from speech signals to obtain target voiceprint features; Input the extracted target voiceprint features into the selected clustering algorithm to obtain voiceprint clusters; Compare the voiceprint feature points of the voiceprint cluster with the voiceprint features in the pre-built voiceprint library and calculate the similarity to determine the identity of the corresponding speaker; Based on the determined speaker identity, in response to the received instruction, a voiceprint cluster corresponding to the speaker is selected for gain processing.

2. The voiceprint processing method according to claim 1, characterized in that: The specific steps of performing noise separation on the audio sample data to obtain the speech signal include: Preprocessing audio sample data; Perform feature extraction on the preprocessed audio sample data to obtain human voice audio features; Based on the audio characteristics of human voice, a pure noise sample is obtained from the preprocessed audio sample data; Based on the pure noise samples, the audio sample data is noise separated to obtain the speech signal.

3. The voiceprint processing method according to claim 2, characterized in that: The specific steps of obtaining a pure noise sample from the preprocessed audio sample data based on the human voice audio features include: According to the human voice audio features, the preprocessed audio sample data is divided into sound segments with human voice audio features and sound segments without human voice audio features; Extracting the audio segments without human voice audio features from the preprocessed audio sample data; The sound segment without human voice audio features is a pure noise sample.

4. The voiceprint processing method according to any one of claims 1 to 3, characterized in that: The specific steps of inputting the extracted target voiceprint features into the selected clustering algorithm to obtain the voiceprint clusters include: The extracted target voiceprint features are input into a selected clustering algorithm, and the clustering algorithm clusters similar voiceprint feature points in the target voiceprint features into voiceprint clusters.

5. The voiceprint processing method according to claim 4, characterized in that: The voiceprint feature points of the same voiceprint cluster are the voiceprint feature points of the same speaker.

6. The voiceprint processing method according to any one of claims 1 to 3, characterized in that: The specific steps of comparing the voiceprint feature points of the voiceprint cluster with the voiceprint features in the pre-built voiceprint library and calculating the similarity to determine the identity of the corresponding speaker include: Match the voiceprint feature points of each voiceprint cluster with the voiceprint features in the pre-built voiceprint library one by one; For each voiceprint feature point of the voiceprint cluster, calculate its similarity with the voiceprint feature of each speaker in the pre-built voiceprint library; When the similarity exceeds a preset threshold, the voiceprint feature points of the voiceprint cluster are successfully matched with the current voiceprint feature, and the speaker of the voiceprint cluster is the speaker of the current voiceprint feature; When the similarity does not exceed the preset threshold, the voiceprint feature points of the voiceprint cluster fail to match the current voiceprint feature, and the speaker of the voiceprint cluster is not the speaker of the current voiceprint feature.

7. The voiceprint processing method according to claim 1, characterized in that: The specific steps of selecting the voiceprint cluster corresponding to the speaker for gain processing based on the determined speaker identity and in response to the received instruction include: Responding to the received instruction based on the determined speaker identity, and parsing the received instruction to extract speaker information; According to the parsed speaker information, searching and selecting the voiceprint cluster corresponding to the speaker from the determined speaker identities; Perform gain processing on the selected voiceprint cluster.

8. The voiceprint processing method according to claim 7, characterized in that: The specific steps of performing gain processing on the selected voiceprint cluster include: The voiceprint feature points of the selected voiceprint cluster are input into the gain processing algorithm to obtain the gained voiceprint feature points and the gained voiceprint cluster.

9. A voiceprint processing system, characterized in that: It comprises a processor and a memory, wherein the memory stores machine-readable instructions executable by the processor, and when the machine-readable instructions are executed by the processor, the steps of the voiceprint processing method as claimed in any one of claims 1 to 8 are performed.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the voiceprint processing method according to any one of claims 1 to 8 are executed.

Images