Method and device for distinguishing voice originator

Abstract

The present disclosure relates to a method and device for distinguishing the originator of a speech signal, and relates to the technical field of speech signal processing. The method comprises: dividing a speech signal into multiple speech segments; extracting voiceprint features and spectral features of each speech segment, fusing them into feature vectors for each speech segment; and clustering each speech segment based on the feature vectors to distinguish the originator of each speech segment.

Description

Technical Field

[0001] The present disclosure relates to the technical field of voice signal processing, and in particular to a method for distinguishing a voice originator, a device for distinguishing a voice originator, and a non-volatile computer-readable storage medium. Background Art

[0002] With the development and progress of speech recognition technology, it plays an increasingly important role in the fields of artificial intelligence and big data analysis. For example, speech recognition technology plays a key role in offline customer service quality inspection speech analysis.

[0003] The offline customer service quality inspection system allows for customized inspection targets, violation type determination, and automatic report generation. This significantly improves inspectors' work efficiency and further enhances service quality. Voice recognition technology can be used to distinguish the originators of individual voice calls in single-channel recordings (i.e., speaker separation).

[0004] In the related art, semantic understanding is used as a basis to distinguish the originators of each voice in a voice call.

[0005] The accuracy of speaker separation is not high, and the speaker separation effect for short-duration speech is poor and the performance is low. Summary of the Invention

[0006] The inventors of the present disclosure have discovered that the above-mentioned related technologies have the following problems: they are unable to deeply explore the characteristics of different voice initiators, resulting in low differentiation accuracy.

[0007] In view of this, the present disclosure proposes a technical solution for distinguishing the originator of a voice, which can improve the accuracy of distinguishing the originator of a voice.

[0008] According to some embodiments of the present disclosure, a method for distinguishing the voice initiator is provided, comprising: dividing a voice signal into multiple voice segments; extracting voiceprint features and spectral features of each voice segment, and fusing them into feature vectors of each voice segment; and clustering each voice segment based on the feature vectors to distinguish the voice initiator of each voice segment.

[0009] In some embodiments, extracting the voiceprint features and spectral features of each speech segment includes: determining the voiceprint features based on the fundamental frequency features of each speech segment, or at least one of the short-time energy features; determining the spectral features based on the frequency distribution features of each speech segment, or at least one of the peak features of the signal envelope.

[0010] In some embodiments, determining the spectral features based on the frequency distribution features of each speech segment or at least one of the peak features of the signal envelope includes: extracting the frequency cepstral coefficients of each speech segment, and performing first-order difference on the frequency cepstral coefficients to determine the frequency distribution features; performing linear prediction coefficient analysis on each speech segment, and performing line spectrum pair solution processing to determine the line spectrum pair features as the peak features of the signal envelope.

[0011] In some embodiments, dividing a speech signal into multiple speech segments includes: dividing the speech signal into multiple first speech segments according to each valid speech contained in the speech signal; performing speech recognition processing on each first speech segment and dividing each first speech segment into each word; and judging whether to divide the first speech segment into multiple second speech segments again according to the interval length of each word in any first speech segment.

[0012] In some embodiments, judging whether to divide any first speech segment into multiple second speech segments again based on the interval duration of each word in any first speech segment includes: when the interval duration of each word in the first speech segment is greater than a first threshold, dividing the first speech segment into multiple second speech segments again based on the interval duration of each word in the first speech segment.

[0013] In some embodiments, clustering processing is performed on each speech segment based on the feature vector, including: dividing long speech segments from each speech segment based on whether the duration is greater than a second threshold; hierarchical clustering is performed on each long speech segment to determine each cluster center; and clustering processing is performed on each speech segment based on each cluster center.

[0014] In some embodiments, hierarchical clustering is performed on each long speech segment to determine each cluster center, including: hierarchical clustering is performed on each long speech segment to determine each candidate cluster center; based on the state probability and transition probability of each long speech segment to each candidate cluster center, the attribution relationship of each long speech segment to each candidate cluster center is re-determined, and the state probability is determined based on the distance from the feature vector of the long speech segment to the candidate cluster center; and each cluster center is determined based on the feature vector of each long speech segment belonging to each candidate cluster center.

[0015] In some embodiments, before clustering each speech segment, the differentiation method also includes at least one of the following steps: filtering the noise segments in each speech segment based on whether the voiceprint features of each speech segment belong to the noise voiceprint range; or matching each speech segment with a preset voice template and filtering the system broadcast voice segments in each speech segment.

[0016] According to other embodiments of the present disclosure, a device for distinguishing the voice initiator is provided, including: a division unit for dividing a voice signal into multiple voice segments; a fusion unit for extracting voiceprint features and spectral features of each voice segment and fusing them into feature vectors of each voice segment; and a clustering unit for clustering each voice segment according to the feature vectors to distinguish the voice initiator of each voice segment.

[0017] In some embodiments, the fusion unit determines the voiceprint features based on the fundamental frequency features of each speech segment or at least one of the short-time energy features, and determines the spectral features based on the frequency distribution features of each speech segment or at least one of the peak features of the signal envelope.

[0018] In some embodiments, the fusion unit extracts the frequency cepstral coefficients of each speech segment, performs first-order difference on the frequency cepstral coefficients, determines the frequency distribution characteristics, performs linear prediction coefficient analysis on each speech segment, and performs line spectrum pair solution processing to determine the line spectrum pair characteristics as the peak characteristics of the signal envelope.

[0019] In some embodiments, the division unit divides the speech signal into multiple first speech segments according to the valid speech contained in the speech signal, performs speech recognition processing on each first speech segment, divides each first speech segment into each word, and determines whether to divide the first speech segment into multiple second speech segments again according to the interval length of each word in any first speech segment.

[0020] In some embodiments, when the interval duration of each word in the first speech segment is greater than a first threshold, the division unit divides the first speech segment into multiple second speech segments according to the interval duration of each word in the first speech segment.

[0021] In some embodiments, the division unit divides long speech segments from each speech segment based on whether the duration is greater than a second threshold, the clustering unit performs hierarchical clustering on each long speech segment, determines each cluster center, and clusters each speech segment based on each cluster center.

[0022] In some embodiments, the clustering unit performs hierarchical clustering on each long speech segment to determine each candidate cluster center, and re-determines the belonging relationship of each long speech segment to each candidate cluster center based on the state probability and transition probability of each long speech segment to each candidate cluster center. The state probability is determined based on the distance from the feature vector of the long speech segment to the candidate cluster center, and each cluster center is determined based on the feature vector of each long speech segment belonging to each candidate cluster center.

[0023] In some embodiments, the distinguishing device also includes: a filtering unit, which is used to perform at least one of the following steps: filtering the noise segments in each voice segment based on whether the voiceprint features of each voice segment belong to the noise voiceprint range; or using a preset voice template to match each voice segment and filter the system broadcast voice segments in each voice segment.

[0024] According to some further embodiments of the present disclosure, a device for distinguishing a voice originator is provided, comprising: a memory; and a processor coupled to the memory, wherein the processor is configured to execute the method for distinguishing a voice originator in any of the above embodiments based on instructions stored in the memory device.

[0025] According to further embodiments of the present disclosure, a non-volatile computer-readable storage medium is provided, on which a computer program is stored. When the program is executed by a processor, the method for distinguishing a voice originator in any of the above embodiments is implemented.

[0026] In the above embodiment, voiceprint features and spectral features are combined as clustering basis to distinguish the originator of the voice. In this way, the voice features of different originators can be deeply mined as the basis for differentiation, thereby improving the accuracy of originator differentiation. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments of the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0028] The present disclosure can be more clearly understood from the following detailed description with reference to the accompanying drawings:

[0029] Figure 1 A flowchart illustrating some embodiments of the method for distinguishing a voice originator of the present disclosure;

[0030] Figure 2 Show Figure 1 Schematic diagrams of some embodiments of step 110;

[0031] Figure 3 Show Figure 1 Schematic diagrams of some embodiments of step 120;

[0032] Figure 4 Show Figure 1 Flowcharts of some embodiments of step 130;

[0033] Figure 5 Show Figure 1 Schematic diagrams of other embodiments of step 130;

[0034] Figure 6 A block diagram illustrating some embodiments of a device for distinguishing a voice originator according to the present disclosure;

[0035] Figure 7 A block diagram showing some other embodiments of the apparatus for distinguishing a voice originator according to the present disclosure;

[0036] Figure 8 The block diagram shows some further embodiments of the apparatus for distinguishing a voice originator of the present disclosure. DETAILED DESCRIPTION

[0037] Various exemplary embodiments of the present disclosure will now be described in detail with reference to the accompanying drawings. It should be noted that unless otherwise specifically stated, the relative arrangement of components and steps, numerical expressions and numerical values set forth in these embodiments do not limit the scope of the present disclosure.

[0038] At the same time, it should be understood that for the convenience of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship.

[0039] The following description of at least one exemplary embodiment is merely illustrative in nature and is in no way intended to limit the present disclosure, its application, or uses.

[0040] Technologies, methods and equipment known to ordinary technicians in the relevant field may not be discussed in detail, but where appropriate, technologies, methods and equipment should be considered part of the authorization specification.

[0041] In all examples shown and discussed herein, any specific values should be interpreted as merely exemplary and not limiting. Therefore, other examples of the exemplary embodiments may have different values.

[0042] It should be noted that like reference numerals and letters refer to like items in the following figures, and therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures.

[0043] Figure 1 A flow chart illustrating some embodiments of the method for distinguishing a voice originator of the present disclosure.

[0044] like Figure 1 As shown, the distinguishing method includes: step 110, dividing the speech into segments; step 120, fusing feature vectors; and step 130, distinguishing the originator of the speech.

[0045] In step 110, the speech signal is divided into a plurality of speech segments.

[0046] In some embodiments, before dividing the speech signal, the speech signal may be subjected to front-end processing, for example, the front-end processing may include at least one of pre-filtering, pre-emphasis, framing, and windowing.

[0047] In some embodiments, the speech signal is divided into a plurality of first speech segments based on valid speech contained in the speech signal. Speech recognition processing is performed on each first speech segment to divide each first speech segment into words. Based on the interval length between words in any first speech segment, it is determined whether to further divide the first speech segment into a plurality of second speech segments.

[0048] For example, when the interval durations of each word in the first speech segment are greater than a first threshold, the first speech segment is further divided into multiple second speech segments according to the interval durations of each word in the first speech segment.

[0049] In some embodiments, the Figure 2 The embodiment in implements step 110.

[0050] Figure 2 Show Figure 1 Schematic diagram of some embodiments of step 110 in FIG.

[0051] like Figure 2 As shown, a VAD (Voice Activate Detect) method may be firstly used to detect effective speech and perform a first segmentation on the speech signal to determine a plurality of first speech segments.

[0052] In this way, the influence of invalid sounds such as noise, silence, ringback tone, and ringing in the voice signal can be removed, thereby improving the efficiency and accuracy of voice recognition transcription, and further improving the accuracy of distinguishing the voice initiator.

[0053] In some embodiments, since the VAD method cannot distinguish between voices with short intervals and originating from different originators, a secondary segmentation process may be performed.

[0054] For example, a first speech segment is segmented using speech recognition. If the interval between each segment is greater than a first threshold (e.g., 0.01s), the first speech segment is segmented twice to obtain a second speech segment. The second speech segment and the first speech segment that has not been segmented twice are taken together as the final speech segment.

[0055] For example, in the first speech segment, it is determined whether the interval between each segment is greater than a third threshold; at two segmentations whose interval is greater than the third threshold, the first speech segment is divided into two second speech segments.

[0056] In this way, the segmented speech of different speakers can be separated to improve the clustering accuracy.

[0057] After segmentation, you can use Figure 1 The remaining steps in continue to distinguish the initiator.

[0058] In step 120, the voiceprint features and spectrum features of each speech segment are extracted and fused into a feature vector of each speech segment.

[0059] In some embodiments, the voiceprint feature is determined based on at least one of the fundamental frequency feature or the short-time energy feature of each speech segment, and the spectral feature is determined based on at least one of the frequency distribution feature or the peak feature of the signal envelope of each speech segment.

[0060] For example, the MFCC (Mel-Frequency Cepstral Coefficients) of each speech segment is extracted and first-order differencing is performed on the MFCC to determine the frequency distribution characteristics. LPC (Linear Prediction Coefficient) analysis is performed on each speech segment, and LSP (Line Spectrum Pair) solution processing is performed to determine the LSP pair characteristics as the peak characteristics of the signal envelope.

[0061] In some embodiments, the Figure 3 The embodiment in implements step 120.

[0062] Figure 3 Show Figure 1 Schematic diagram of some embodiments of step 120 in FIG.

[0063] like Figure 3 As shown, the feature extraction in step 120 may include voiceprint feature extraction, spectrum feature extraction and feature fusion. For example, before extracting features, the voice signal may be processed at the front end.

[0064] In some embodiments, MFCC focuses on the auditory characteristics of the human ear, and the scale of the Mel frequency corresponds to the logarithmic distribution of the actual frequency f. For example, the Mel frequency can be obtained by performing FFT (Fast Fourier Transform) on the speech segment and then performing Mel filtering:

[0065]

[0066] For example, when the number of filter banks is 12, DCT (Discrete Cosine Transform) is performed on the filtering result to obtain 12-dimensional MFCC. The first-order difference of MFCC is performed to obtain 24-dimensional frequency distribution features.

[0067] In some embodiments, LSP features are obtained by LPC analysis and LSP solution processing, thereby solving frequency domain parameters related to the peak value of the speech signal spectrum envelope. For example, the LSP feature can be a 14-dimensional vector.

[0068] In some embodiments, the fundamental frequency (Pitch), the inverse of the pitch period, is one of the most important parameters of a speech signal. The changing pattern of the fundamental frequency is the pitch, carrying very important information that can be used to discern meaning. For example, a first-order difference can be taken on the fundamental frequency to obtain a two-dimensional fundamental frequency feature.

[0069] In some embodiments, short-term energy (Energy) is the energy characteristic of each frame of speech signal. The energy of a speech signal varies over time, and the energy difference between unvoiced and voiced sounds is quite significant. Analysis of short-term energy can describe the changes in speech characteristics. Because speakers (speech initiators) have different energy levels, short-term energy characteristics can be used as a basis for clustering speech segments to distinguish between speakers.

[0070] For example, for the sample value x(m) at the speech signal sampling point m, the short-time energy can be calculated according to the following formula:

[0071] E(m)=10log(x(m) 2 )

[0072] In some embodiments, the 24-dimensional MFCC, 14-dimensional LSP, 2-dimensional fundamental frequency features, and 1-dimensional short-time energy features extracted above are spliced together and fused into a 41-dimensional feature vector as a basis for distinguishing (clustering) different speakers.

[0073] In some embodiments, background noise may be saved as speech segments during the speech segmentation process and included in speaker clustering. This can affect the accuracy of clustering by the initiator. Noise segments can be filtered out of each speech segment based on whether their voiceprint features fall within the noise voiceprint range.

[0074] For example, the voiceprint feature range (fundamental frequency and energy range) of the initiator's voice is obtained to determine the noise voiceprint range (voiceprint threshold). Speech segments whose voiceprint features are not within the noise voiceprint range are considered noise segments. Filtering out noise segments prevents them from participating in hierarchical clustering, thereby improving the accuracy of the cluster center and, in turn, the performance of the initiator's clustering.

[0075] In some embodiments, each voice segment is matched using a preset voice template, and the system broadcast voice (such as machine broadcast voice) segment in each voice segment is filtered.

[0076] For example, in the telephone customer service field, there are some system announcements, such as "This is operator number x, here to assist you." These machine announcements are not part of the conversation between the customer and the agent and can interfere with the initiator clustering. Template matching can be used to filter out these system announcements.

[0077] After feature extraction, we can use Figure 1 The remaining steps in continue to distinguish the initiator.

[0078] In step 130, clustering is performed on each speech segment based on the feature vector to distinguish the speech originator of each speech segment.

[0079] In some embodiments, the speech segments include short segments whose duration is less than or equal to a second threshold. These short segments are prone to misjudgment during the clustering process, which in turn can have an erroneous impact on the next step of clustering. This impact can lead to cumulative errors, resulting in poor clustering performance. Therefore, the impact of these short segments on clustering performance should be minimized before clustering.

[0080] For example, based on whether the duration is greater than a second threshold, long speech segments are divided from each speech segment. Hierarchical clustering is performed on each long speech segment to determine each cluster center. Clustering processing is then performed on each speech segment based on each cluster center.

[0081] In some embodiments, the Figure 4 The embodiment in implements step 130.

[0082] Figure 4 Show Figure 1 Flowchart of some embodiments of step 130 in FIG.

[0083] like Figure 4 As shown in the figure, short speech segments (e.g., duration of 1s to 2s) are not included in AHC (Agglomerative Hierarchical Clustering) to avoid adversely affecting the cluster center. When performing K-means (K-means clustering algorithm) clustering, short speech segments are clustered together with long speech segments.

[0084] In some embodiments, to prevent a decrease in clustering performance due to a lack of remaining speech segments after filtering out short speech segments, the second threshold can be dynamically adjusted. For example, the ratio of the number of remaining long speech segments to the number of original speech segments can be determined. If the ratio is less than a threshold, the second threshold is lowered.

[0085] In this way, the interference of short speech segments on the clustering center can be prevented, and the clustering accuracy of short speech segments can be improved.

[0086] In some embodiments, hierarchical clustering is performed on each long speech segment to determine candidate cluster centers. The attribution of each long speech segment to each candidate cluster center is re-determined based on the state probability and transition probability of each long speech segment with respect to each candidate cluster center. The state probability is determined based on the distance between the feature vector of the long speech segment and the candidate cluster center. Each cluster center is determined based on the feature vector of each long speech segment belonging to each candidate cluster center.

[0087] In some embodiments, each cluster center may be optimized using the Viterbi algorithm.

[0088] For example, after AHC, candidate cluster centers and the distances from each speech segment to each cluster center are obtained. Different cluster centers are used as different states. Based on the distances from each speech segment to different cluster centers, the state probabilities of the speech segments relative to the cluster centers are determined, and the transition probabilities between the states are set.

[0089] On this basis, the Viterbi algorithm is used to optimize the relationship between each speech segment and each cluster center, and then the cluster centers are recalculated as the basis for K-Means clustering. Figure 5 The embodiment in implements step 130.

[0090] Figure 5 Show Figure 1 Schematic diagrams of other embodiments of step 130 in FIG.

[0091] like Figure 5 As shown, cluster centers 1 and 2 are two states of the Viterbi algorithm. For example, after AHC, speech segments 1 to N (all long speech segments) are divided into two state categories, so that cluster centers 1 and 2 can be calculated.

[0092] Figure 5 The arrows in the figure represent the state transition paths for each speech segment. Based on the decoding process in the Viterbi algorithm, the state probability and transition probability for each state of each speech segment are calculated. The weight of each state transition path for each speech segment is calculated based on the product of the state probability and the transition probability.

[0093] The optimal decoding path is determined by backtracking the Viterbi algorithm. For example, the state transition path with the highest weight for each speech segment can be determined as the optimal decoding path (dashed arrow in the figure). Each speech segment is then re-divided into different states. Based on the re-divided state, the new cluster center of each state is calculated, thereby optimizing the cluster center.

[0094] In the above embodiment, the method of fusion of voiceprint and spectrum features is used as the clustering basis; secondary segmentation is performed according to the time point of speech recognition word segmentation; the speech segments are filtered; and after hierarchical clustering, the Viterbi algorithm is used to optimize the hierarchical clustering results.

[0095] This improves the accuracy of speaker separation and the robustness of the recognition algorithm, with the performance improvement being particularly significant for short speech sequences.

[0096] Figure 6 A block diagram illustrating some embodiments of a device for distinguishing a voice originator according to the present disclosure.

[0097] like Figure 6 As shown, the device 6 for distinguishing the voice originator includes a dividing unit 61 , a fusion unit 62 and a clustering unit 63 .

[0098] The division unit 61 divides the speech signal into a plurality of speech segments.

[0099] The fusion unit 62 extracts the voiceprint features and spectrum features of each speech segment and fuses them into a feature vector of each speech segment.

[0100] The clustering unit 63 performs clustering processing on each speech segment according to the feature vector, and distinguishes the speech originator of each speech segment.

[0101] In some embodiments, the fusion unit 62 determines the voiceprint feature based on at least one of the fundamental frequency feature or the short-time energy feature of each speech segment. The fusion unit 62 determines the spectral feature based on at least one of the frequency distribution feature or the peak feature of the signal envelope of each speech segment.

[0102] In some embodiments, the fusion unit 62 extracts the frequency cepstral coefficients of each speech segment and performs first-order difference on the frequency cepstral coefficients to determine the frequency distribution characteristics. The fusion unit 62 performs linear prediction coefficient analysis on each speech segment and performs line spectrum pair processing to determine the line spectrum pair characteristics as the peak characteristics of the signal envelope.

[0103] In some embodiments, the division unit 61 divides the speech signal into a plurality of first speech segments based on each valid speech signal contained in the speech signal. The division unit 61 performs speech recognition processing on each first speech segment and divides each first speech segment into each word segment. The division unit 61 determines whether to further divide any first speech segment into a plurality of second speech segments based on the interval length between each word segment in the first speech segment.

[0104] In some embodiments, when the interval duration of each word in the first speech segment is greater than a first threshold, the division unit 61 divides the first speech segment into multiple second speech segments according to the interval duration of each word in the first speech segment.

[0105] In some embodiments, the segmentation unit 61 divides the speech segments into long speech segments based on whether the duration is greater than a second threshold. The clustering unit 63 performs hierarchical clustering on the long speech segments to determine cluster centers. The clustering unit 63 clusters the speech segments based on the cluster centers.

[0106] In some embodiments, clustering unit 63 performs hierarchical clustering on each long speech segment to determine candidate cluster centers. Clustering unit 63 re-determines the attribution relationship of each long speech segment to each candidate cluster center based on the state probability and transition probability of each long speech segment to each candidate cluster center. The state probability is determined based on the distance between the feature vector of the long speech segment and the candidate cluster center. Clustering unit 63 determines each cluster center based on the feature vector of each long speech segment that belongs to each candidate cluster center.

[0107] In some embodiments, the distinguishing device 6 also includes: a filtering unit 64, which is used to perform at least one of the following steps: filtering the noise segments in each speech segment according to whether the voiceprint features of each speech segment belong to the noise voiceprint range; or using a preset voice template to match each speech segment and filter the system broadcast voice segments in each speech segment.

[0108] Figure 7 A block diagram showing some other embodiments of the apparatus for distinguishing a voice originator of the present disclosure.

[0109] like Figure 7 As shown, the voice originator distinguishing device 7 of this embodiment includes: a memory 71 and a processor 72 coupled to the memory 71, and the processor 72 is configured to execute the voice originator distinguishing method in any embodiment of the present disclosure based on the instructions stored in the memory 71.

[0110] The memory 71 may include, for example, a system memory, a fixed non-volatile storage medium, etc. The system memory may store, for example, an operating system, application programs, a boot loader, a database, and other programs.

[0111] Figure 8 The block diagram shows some further embodiments of the apparatus for distinguishing a voice originator of the present disclosure.

[0112] like Figure 8As shown, the voice originator distinguishing device 8 of this embodiment includes: a memory 810 and a processor 820 coupled to the memory 810, and the processor 820 is configured to execute the voice originator distinguishing method in any of the aforementioned embodiments based on the instructions stored in the memory 810.

[0113] The memory 810 may include, for example, a system memory, a fixed non-volatile storage medium, etc. The system memory may store, for example, an operating system, application programs, a boot loader, and other programs.

[0114] The device for distinguishing the originator of a voice call may further include an input / output interface 830, a network interface 840, a storage interface 850, and the like. These interfaces 830, 840, 850, as well as the memory 810 and the processor 820, may be connected, for example, via a bus 860. The input / output interface 830 provides a connection interface for input / output devices such as a display, mouse, keyboard, touch screen, microphone, and speakers. The network interface 840 provides a connection interface for various networked devices. The storage interface 850 provides a connection interface for external storage devices such as SD cards and USB flash drives.

[0115] Those skilled in the art will appreciate that embodiments of the present disclosure may be provided as methods, systems, or computer program products. Thus, the present disclosure may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present disclosure may take the form of a computer program product implemented on one or more computer-usable non-transitory storage media, including but not limited to magnetic disk storage, CD-ROMs, optical storage, and the like, containing computer-usable program code.

[0116] The method for distinguishing a voice originator, the device for distinguishing a voice originator, and the non-volatile computer-readable storage medium according to the present disclosure have been described in detail. To avoid obscuring the concepts of the present disclosure, some details known in the art have been omitted. Based on the above description, those skilled in the art will fully understand how to implement the technical solutions disclosed herein.

[0117] The methods and systems of the present disclosure may be implemented in many ways. For example, the methods and systems of the present disclosure may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above order of steps for the method is for illustration only, and the steps of the method of the present disclosure are not limited to the order specifically described above unless otherwise specified. In addition, in some embodiments, the present disclosure may also be implemented as programs recorded in a recording medium, which include machine-readable instructions for implementing the methods according to the present disclosure. Therefore, the present disclosure also covers recording media that store programs for executing the methods according to the present disclosure.

[0118] Although some specific embodiments of the present disclosure have been described in detail by way of examples, those skilled in the art will appreciate that the above examples are for illustrative purposes only and are not intended to limit the scope of the present disclosure. Those skilled in the art will appreciate that modifications may be made to the above embodiments without departing from the scope and spirit of the present disclosure. The scope of the present disclosure is defined by the appended claims.

Claims

1. A method for distinguishing a voice originator, comprising: Dividing the speech signal into multiple speech segments; Extracting voiceprint features and spectral features of each speech segment and fusing them into feature vectors of each speech segment; According to the feature vector, clustering processing is performed on the speech segments to distinguish the speech originators of the speech segments. The step of dividing the speech signal into a plurality of speech segments includes: According to each valid speech contained in the speech signal, the speech signal is divided into a plurality of first speech segments, Perform speech recognition processing on each first speech segment, dividing the first speech segment into words, Determining whether to divide any first speech segment into multiple second speech segments based on the interval lengths between each segment in the first speech segment; The clustering of the speech segments according to the feature vectors comprises: According to whether the duration is greater than a second threshold, the speech segments are divided into long speech segments and short speech segments, the duration of the long speech segments is greater than the second threshold, and the duration of the short speech segments is less than or equal to the second threshold, Perform hierarchical clustering on each long speech segment to determine each cluster center. The short speech segment does not participate in the hierarchical clustering. Clustering processing is performed on each speech segment according to each cluster center, and the clustering processing includes K-means clustering.

2. The method of claim 1, wherein: The extraction of voiceprint features and spectral features of each speech segment includes: Determining the voiceprint feature according to at least one of the fundamental frequency feature or the short-time energy feature of each speech segment; The spectral feature is determined according to at least one of the frequency distribution feature of each speech segment or the peak feature of the signal envelope.

3. The method of claim 2, wherein: Determining the spectral feature according to at least one of the frequency distribution feature of each speech segment or the peak feature of the signal envelope includes: Extracting frequency cepstral coefficients of each speech segment, and performing first-order difference on the frequency cepstral coefficients to determine the frequency distribution characteristics; A linear prediction coefficient analysis is performed on each speech segment, and a line spectrum pair solution is performed to determine the line spectrum pair feature as the peak feature of the signal envelope.

4. The method of claim 1, wherein: The determining whether to divide any first speech segment into a plurality of second speech segments again according to the interval duration of each segment in any first speech segment comprises: When the interval durations of each word in the first speech segment are greater than a first threshold, the first speech segment is further divided into a plurality of second speech segments according to the interval durations of each word in the first speech segment.

5. The method of claim 1, wherein: The hierarchical clustering of each long speech segment to determine each cluster center includes: Perform hierarchical clustering on each long speech segment to determine each candidate cluster center; Re-determining the attribution relationship between each long speech segment and each candidate cluster center according to the state probability and transition probability of each long speech segment to each candidate cluster center, wherein the state probability is determined according to the distance between the feature vector of the long speech segment and the candidate cluster center; The cluster centers are determined according to the feature vectors of the long speech segments belonging to the candidate cluster centers.

6. The method for distinguishing according to any one of claims 1 to 5, further comprising at least one of the following steps before clustering the speech segments: filtering out noise segments in each of the speech segments according to whether the voiceprint features of the speech segments belong to the noise voiceprint range; or The voice segments are matched using a preset voice template, and the system announcement voice segments in the voice segments are filtered.

7. A device for distinguishing a voice originator, comprising: A division unit, configured to divide a speech signal into a plurality of speech segments; A fusion unit, configured to extract voiceprint features and spectral features of each speech segment and fuse them into feature vectors of each speech segment; A clustering unit is used to perform clustering processing on each of the speech segments according to the feature vector to distinguish the speech originators of each of the speech segments, The division unit divides the speech signal into a plurality of first speech segments according to each valid speech contained in the speech signal, performs speech recognition processing on each first speech segment, divides each first speech segment into each word, and determines whether to further divide the first speech segment into a plurality of second speech segments according to the interval length of each word in any first speech segment. The dividing unit divides the speech segments into long speech segments and short speech segments according to whether the duration is greater than a second threshold, wherein the duration of the long speech segments is greater than the second threshold, and the duration of the short speech segments is less than or equal to the second threshold. The clustering unit performs hierarchical clustering on each long speech segment, determines each cluster center, and performs clustering processing on each speech segment based on the cluster centers. The short speech segments do not participate in the hierarchical clustering. The clustering processing includes K-means clustering.

8. The distinguishing device according to claim 7, wherein: The fusion unit determines the voiceprint feature based on at least one of the fundamental frequency features or short-time energy features of each speech segment, and determines the spectrum feature based on at least one of the frequency distribution features or peak features of the signal envelope of each speech segment.

9. The distinguishing device according to claim 8, wherein: The fusion unit extracts the frequency cepstral coefficients of each speech segment, performs first-order difference on the frequency cepstral coefficients, determines the frequency distribution characteristics, performs linear prediction coefficient analysis on each speech segment, and performs line spectrum pair solution processing to determine the line spectrum pair characteristics as the peak characteristics of the signal envelope.

10. The distinguishing device according to claim 7, wherein: The division unit further divides the first speech segment into a plurality of second speech segments according to the interval lengths of the segmented words in the first speech segment when the interval lengths of the segmented words in the first speech segment are all greater than a first threshold.

11. The distinguishing device according to claim 7, wherein: The clustering unit performs hierarchical clustering on each long speech segment to determine each candidate cluster center, and re-determines the belonging relationship of each long speech segment to each candidate cluster center based on the state probability and transition probability of each long speech segment to each candidate cluster center. The state probability is determined based on the distance from the feature vector of the long speech segment to the candidate cluster center, and the cluster centers are determined based on the feature vector of each long speech segment belonging to each candidate cluster center.

12. The distinguishing device according to any one of claims 7 to 11, further comprising: A filtering unit, configured to perform at least one of the following steps: filtering out noise segments in each of the speech segments according to whether the voiceprint features of the speech segments belong to the noise voiceprint range; or The voice segments are matched using a preset voice template, and the system announcement voice segments in the voice segments are filtered.

13. A device for distinguishing a voice originator, comprising: Memory; and A processor coupled to the memory, wherein the processor is configured to execute the method for distinguishing a voice originator according to any one of claims 1 to 6 based on instructions stored in the memory.

14. A non-volatile computer-readable storage medium having a computer program stored thereon, wherein when the program is executed by a processor, the method for distinguishing a voice originator according to any one of claims 1 to 6 is implemented.

Images