A method, device, electronic device and storage medium for separating talkers

Abstract

The embodiments of the present application are a method, device, electronic device and storage medium for speaker separation, and belong to the field of speech recognition technology. The method comprises: obtaining a plurality of speech segments to be processed, dividing the plurality of speech segments to be processed into two speech piles; respectively selecting two groups of first long speech segments whose speech duration is greater than a first preset duration from the two speech piles; respectively registering the two groups of first long speech segments as two voiceprint samples, wherein different voiceprint samples represent different voice users; respectively matching the two voiceprint samples with two groups of short speech segments whose speech duration is less than a second preset duration, so as to determine the voiceprint samples to which the two groups of short speech segments belong, wherein the second preset duration is less than or equal to the first preset duration. Using a speaker separation method provided by the present application, the accuracy of identifying short speech segments can be improved.

Description

Technical Field

[0001] The embodiments of the present application relate to the field of speech recognition technology, and in particular, to a speaker separation method, device, electronic device and storage medium. Background Art

[0002] Speaker separation, also known as speaker identification, is a recognition technology similar to speech recognition processing technology. It is usually used to separate at least two people in an audio segment with at least two people in order to identify different speakers.

[0003] Traditional speaker separation technology uses the idea of ​​clustering to separate different speakers. However, when using the idea of ​​clustering to identify speakers of short-speaking voice clips in audio, the accuracy of the identified speakers is low. Summary of the invention

[0004] The embodiments of the present application provide a speaker separation method, device, electronic device and storage medium, which are intended to solve the problem of low speaker recognition accuracy when recognizing short speech segments.

[0005] A first aspect of an embodiment of the present application provides a method for separating speakers, the method comprising:

[0006] Acquire a plurality of speech segments to be processed, and divide the plurality of speech segments to be processed into two speech piles;

[0007] Selecting two groups of first long voice segments whose voice duration is greater than a first preset duration from the two voice piles respectively;

[0008] Registering the two groups of first long voice segments into two voiceprint samples respectively, wherein different voiceprint samples represent different voice users;

[0009] The two voiceprint samples are matched respectively with two groups of short voice segments whose voice duration is less than a second preset duration to determine the voiceprint samples to which the two groups of short voice segments belong, wherein the second preset duration is less than or equal to the first preset duration.

[0010] Optionally, the two groups of first long voice segments are registered as two voiceprint samples respectively, including:

[0011] According to the ivocar features of the two groups of first long speech segments, the two groups of first long speech segments are clustered again to obtain two groups of second long speech segments;

[0012] The two groups of second long voice segments are registered as two voiceprint samples respectively.

[0013] Optionally, according to the ivocar features of the two groups of first long speech segments, the two groups of first long speech segments are clustered twice to obtain two groups of second long speech segments, including:

[0014] Using the PLDA model, the ivocar features of each of the two first long speech segments are scored to obtain the PLDA score between each of the two first long speech segments;

[0015] According to the plda score, the first long speech segment is clustered twice to obtain two groups of second long speech segments.

[0016] Optionally, a plurality of speech segments to be processed are obtained, and the plurality of speech segments to be processed are divided into two speech piles, including:

[0017] Extracting features from the plurality of speech segments to be processed respectively to obtain the mfcc features of the plurality of speech segments to be processed;

[0018] According to the respective MFCC features of the multiple speech segments to be processed, the multiple speech segments to be processed are clustered once to divide the multiple speech segments to be processed into two speech piles.

[0019] Optionally, clustering the multiple voice segments to be processed according to the respective MFCC features of the multiple voice segments to be processed to divide the multiple voice segments to be processed into two voice piles, including:

[0020] Performing VAD silence detection on the MFCC features of the multiple speech segments to be processed, and extracting the MFCC features of multiple speech activity segments in the speech segments to be processed;

[0021] According to the MFCC features of the multiple voice activity segments, the multiple voice activity segments are clustered to divide the multiple voice activity segments into two voice piles.

[0022] Optionally, the method further comprises:

[0023] The first preset time length is set according to the accuracy of the secondary clustering; wherein the first preset time length is directly proportional to the accuracy of the secondary clustering.

[0024] A second aspect of an embodiment of the present application provides a talker separation device, the device comprising:

[0025] A division module, used for acquiring a plurality of speech segments to be processed, and dividing the plurality of speech segments to be processed into two speech piles;

[0026] A screening module, used to screen out two groups of first long voice segments whose voice duration is greater than a first preset duration from the two voice piles respectively;

[0027] A registration module, used to register the two groups of first long voice segments into two voiceprint samples respectively, wherein different voiceprint samples represent different voice users;

[0028] The matching module is used to match the two voiceprint samples with two groups of short voice segments whose voice duration is less than a second preset duration, so as to determine the voiceprint samples to which the two groups of short voice segments belong respectively, wherein the second preset duration is less than or equal to the first preset duration.

[0029] Optionally, the registration module includes:

[0030] A secondary clustering module, used for performing secondary clustering on the two groups of first long speech segments according to respective ivictor features of the two groups of first long speech segments, so as to obtain two groups of second long speech segments;

[0031] The first registration submodule is used to register the two groups of second long voice segments into two voiceprint samples respectively.

[0032] Optionally, the secondary clustering module includes:

[0033] A first scoring module is used to score the ivocar features of each two first long speech segments by using the PLDA model to obtain a PLDA score between each two first long speech segments;

[0034] The secondary clustering submodule is used to perform secondary clustering on the first long speech segment according to the plda score to obtain two groups of second long speech segments.

[0035] Optionally, the division module includes:

[0036] A feature extraction module, used to extract features from the plurality of speech segments to be processed respectively, to obtain the mfcc features of the plurality of speech segments to be processed respectively;

[0037] The primary clustering module is used to perform primary clustering on the multiple speech segments to be processed according to the respective MFCC features of the multiple speech segments to be processed, so as to divide the multiple speech segments to be processed into two speech piles.

[0038] Optionally, the primary clustering module includes:

[0039] A silence detection module, used to perform VAD silence detection on the MFCC features of the multiple voice segments to be processed, and extract the MFCC features of multiple voice activity segments in the voice segments to be processed;

[0040] The primary clustering submodule is used to perform primary clustering on the multiple voice activity segments according to the MFCC features of the multiple voice activity segments, so as to divide the multiple voice activity segments into two voice piles.

[0041] Optionally, the device further comprises:

[0042] The setting module is used to set the first preset time length according to the accuracy of the secondary clustering; wherein the first preset time length is directly proportional to the accuracy of the secondary clustering.

[0043] A third aspect of an embodiment of the present application provides an electronic device, including:

[0044] one or more processors; and

[0045] One or more machine-readable media having instructions stored thereon, when executed by the one or more processors, enable the electronic device to perform the speaker separation method as described in the first aspect of the embodiment of the present application.

[0046] A fourth aspect of an embodiment of the present application provides one or more machine-readable storage media having instructions stored thereon, which, when executed by one or more processors, enable the processors to perform the speaker separation method as described in the first aspect of the embodiment of the present application.

[0047] By adopting a speaker separation method provided in the present application, after obtaining the speech to be processed, the speech to be processed can be divided into multiple speech segments, and the multiple speech segments can be divided into two speech piles, and then the first long speech segment in the two speech piles can be registered, and the short speech segments in the two speech piles can be matched with the first long speech segment respectively to determine the voiceprint samples to which the two groups of short speech segments belong, thereby identifying different speakers.

[0048] In this process, since the registered voiceprint samples are two groups of first long voice segments, not all voiceprint samples of multiple voice segments to be processed, nor short voice segments of multiple voice segments to be processed, the registered voiceprint samples do not include short voice segments. The registered voiceprint samples contain the first long voice segments with rich information and clear meaning, which will not be affected by the short voice segments. Therefore, the accuracy of voiceprint recognition can be improved. Accordingly, even when the short voice segments with shorter voice duration are matched with the two groups of first long voice segments, the speaker corresponding to the short voice segments can be accurately identified.

[0049] In addition, in the embodiment of the present application, the voiceprint recognition technology is applied to the scenario of speaker separation. Compared with the application of voiceprint recognition technology to the voiceprint unlocking scenario in the prior art, the short voice segment of the user in the same voice environment is matched with the first long voice segment, rather than registering the voice segment in one voice environment as a sample and matching the voice segment in another voice environment with the sample, thereby reducing the interference caused by different voice environments and further improving the accuracy of speaker recognition separation. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the description of the embodiments of the present application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0051] Figure 1 It is a flowchart of a method for separating talkers proposed in one embodiment of the present application;

[0052] Figure 2 is another step flow chart of a method for separating talkers proposed in one embodiment of the present application;

[0053] Figure 3 It is a logical schematic diagram of dividing into two speech piles proposed in an embodiment of the present application;

[0054] Figure 4 It is a logic diagram of a speaker separation method proposed in an embodiment of the present application;

[0055] Figure 5 It is a structural block diagram of a speaker separation device proposed in one embodiment of the present application. DETAILED DESCRIPTION

[0056] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0057] See also Figure 1 The present application embodiment provides a method for separating talkers, the method comprising the following steps:

[0058] Step 101: Acquire multiple speech segments to be processed, and divide the multiple speech segments to be processed into two speech piles.

[0059] In an embodiment of the present application, the speech segment to be processed is divided from the audio to be processed. When the audio to be processed is divided into multiple speech segments to be processed, the audio to be processed can be divided into multiple speech segments to be processed, so that each speech segment to be processed includes only the speech of one speaker, or only includes environmental speech such as noise, thereby improving the accuracy of subsequent secondary clustering.

[0060] Among them, after dividing the audio to be processed into multiple voice segments to be processed, since the voice segments to be processed include voice segments of two speakers, it is necessary to perform preliminary division on the multiple voice segments to be processed to divide them into two voice piles to preliminarily separate two different speakers.

[0061] The embodiments of the present application may be applied to a scenario where voice input from two different speakers needs to be separated, wherein the speakers may be speakers in a telephone conference or a video conference.

[0062] When it is specifically applied to a scenario where two people are talking, the video to be processed of the two people talking is first obtained, and then the audio to be processed is divided into multiple voice segments to be processed through voice framing technology, for example, into multiple voice segments to be processed such as "hello", "how are you", "how are you doing recently", "I'm doing well, how about you", etc., and then the two voice segments to be processed "hello" and "how are you doing recently" are divided into the first voice pile, and the two voice segments to be processed "hello" and "I'm doing well, how about you" are divided into the second voice pile.

[0063] In this way, after being divided into two speech piles, it can be preliminarily identified that the first speech pile is uttered by user A and the second speech pile is uttered by user B, thereby preliminarily realizing the identification of the interlocutors.

[0064] Step 102: Select two groups of first long voice segments whose voice duration is greater than a first preset duration from the two voice piles.

[0065] In the embodiment of the present application, the duration of the voice segments to be processed contained in the two voice piles is different. Specifically, the two voice piles contain long voice segments and short voice segments, and in the short voice segments such as "um" and "hello", the short voice segments contain less information, and the information represented by the sentence is not very clear. Therefore, if the short voice segments are registered together with the long voice segments such as "How are you doing recently?" and "I'm doing well, how about you?", the registered voiceprint samples will contain short voice segments with less information. Since the voiceprint samples of the registered short voice segments themselves have less information, if the short voice uttered by the speaker is matched with the voiceprint samples with less information, the accuracy of the identified speaker will be lower.

[0066] Therefore, in the embodiment of the present application, before registering the first long voice segment, the first long voice segment will be screened. Specifically, the first long voice segment with a voice duration longer than the first preset duration is screened out from the first voice pile, and the first long voice segment with a voice duration longer than the first preset duration is screened out from the second voice pile, and the long voice segments in the first voice pile and the second voice pile are combined into two groups of first long voice segments.

[0067] In specific application, the first long voice segment of more than 1S is selected from the first voice pile corresponding to "Hello" and "How are you recently?"; the first long voice segment of "I'm doing well, how about you" is selected from the second voice pile corresponding to "Hello" and "I'm doing well, how about you?". In this way, the first long voice segments selected are all voice segments with longer voice duration, clear semantics, and more information, and will not be affected by fewer short voice segments with shorter voice duration, thereby improving the accuracy of voiceprint recognition.

[0068] The first preset duration may be any preset duration between 1 and 2 seconds.

[0069] Among them, voiceprint registration refers to registering and saving the voice features of multiple speakers, and then identifying a segment of speech in which one person speaks to identify which one of the multiple speakers it belongs to. It is used in voiceprint password scenarios. For example, the voice features of user A, user B, and user C are registered and saved in a mobile phone. User A is the owner of the mobile phone. When user A speaks to the mobile phone, the mobile phone can be unlocked because the voice features of user A are registered in the mobile phone.

[0070] Step 103: Register the two groups of first long voice segments into two voiceprint samples respectively, wherein different voiceprint samples represent different voice users.

[0071] In an embodiment of the present application, after obtaining the pre-recorded voice to be processed, the voice to be processed is immediately divided into multiple voice segments to be processed, and then the multiple voice segments to be processed are divided into two voice piles, and two groups of first long voice segments in the two voice piles whose voice duration is greater than the first preset duration are registered as two voiceprint samples, so that voiceprint samples of different users are stored in the client.

[0072] In traditional voiceprint registration and recognition technology, the user's voice clips in a voice environment (such as a quiet place like a library) are registered in advance on the client. When voiceprint unlocking is performed, the same user can make a voice in another environment (such as a noisy place like a vegetable market) to unlock the client with the voiceprint.

[0073] In the present application, compared with the traditional voiceprint registration and recognition technology, what is obtained is the complete speech to be processed. Therefore, after obtaining the speech to be processed, it is necessary to process the speech to be processed into two groups of first long voice segments for registration in real time, rather than registering the user's voice segments in advance. Therefore, the present application can identify different speakers in a segment of speech to be processed in real time.

[0074] In addition, the present application divides the to-be-processed speech emitted by the user in the same voice environment (for example, a quiet place such as a library) into short voice segments and a first long voice segment. Therefore, matching the short voice segment with the first long voice segment in the same voice environment can reduce the differences brought by different voice environments, reduce the interference of speaker recognition brought by different voice environments, and improve the accuracy of speaker recognition.

[0075] Among them, the two voiceprint samples contain a large number of first long voice segments of two different users, so the two voiceprint samples can represent the voice characteristics of different voice users.

[0076] Step 104: Match the two voiceprint samples with two groups of short voice segments whose voice duration is less than a second preset duration respectively to determine the voiceprint samples to which the two groups of short voice segments belong, wherein the second preset duration is less than or equal to the first preset duration.

[0077] In the embodiment of the present application, two groups of long voice segments with a voice duration longer than the first preset duration can be registered as two voiceprint samples to improve the accuracy of voiceprint recognition; and then two groups of short voice segments with a voice duration shorter than the second preset duration are matched with the two samples respectively to determine the voiceprint samples to which the two groups of short voice segments belong. In this way, since the embodiment of the present application matches the short voice segments with the voiceprint samples with richer information, longer voice duration, and clear information meaning after the short voice segments are removed, rather than matching the short voice segments with all the voiceprint samples of the user, the accuracy of identifying different speakers is improved.

[0078] Among them, all the voiceprint samples of the user include voiceprint samples whose speech duration is longer than a first preset duration and voiceprint samples whose speech duration is shorter than a second preset duration.

[0079] The second preset time length is less than or equal to the first preset time length. For example, the second preset time length and the first preset time length may both be 1 second, or the first preset time length is 2 seconds and the second preset time length is 1 second.

[0080] In specific application, the first long voice segment in the first voice pile "How are you doing recently?" and the first long voice segment in the second voice pile "I'm doing well, how about you?" are registered as two samples respectively. The first voice pile is sent by user A, and the second voice pile is sent by user B.

[0081] Then match "hello" in the first voice pile with the two samples respectively, and determine that the voice features of the short voice segment corresponding to "hello" are compatible with the voice features of "how are you doing recently" uttered by user A. It can be determined that the speaker corresponding to "hello" is user A.

[0082] Then match "Hello" in the second voice pile with the two samples respectively, and determine that the voice features of the short voice segment corresponding to "Hello" are compatible with "I'm doing well, how about you" uttered by user B. It can be determined that the speaker corresponding to "Hello" is user B.

[0083] In this way, in an embodiment of the present application, after obtaining the speech to be processed, the speech to be processed can be divided into multiple speech segments, and the multiple speech segments can be divided into two speech piles, and then the first long speech segment in the two speech piles can be registered, and the short speech segments in the two speech piles can be matched with the first long speech segment respectively to identify different speakers.

[0084] In this process, since the registered voiceprint samples are two groups of first long voice segments, not all voiceprint samples of multiple voice segments to be processed, nor short voice segments of multiple voice segments to be processed, the registered voiceprint samples do not include short voice segments. The registered voiceprint samples contain the first long voice segments with rich information and clear meaning, which will not be affected by the short voice segments. Therefore, the accuracy of voiceprint recognition can be improved. Accordingly, even when the short voice segments with shorter voice duration are matched with the two groups of first long voice segments, the speaker corresponding to the short voice segments can be accurately identified.

[0085] In addition, in the embodiment of the present application, the voiceprint sample is first registered, and then the voiceprint recognition technology corresponding to the voiceprint sample is matched with the speaker's voice is applied to the scenario of speaker separation. Compared with the application of voiceprint recognition technology to the voiceprint unlocking scenario in the prior art, the short voice segment of the user in the same voice environment is matched with the first long voice segment, rather than registering the voice segment in one voice environment as a sample and matching the voice segment in another voice environment with the sample, thereby reducing the interference caused by different voice environments and further improving the accuracy of speaker recognition separation.

[0086] See also Figure 2 A speaker separation method provided by an embodiment of the present application is shown, and Figure 4A logic diagram of a method for separating speakers is provided, the method comprising the following steps:

[0087] Step 201: extracting features from the plurality of speech segments to be processed respectively to obtain the mfcc features of the plurality of speech segments to be processed respectively.

[0088] In the present application, please refer to Figure 3 , we can first use the speech framing technology to divide the speech to be processed into multiple speech segments to be processed, and then sequentially subject the speech segments to be processed to windowing, fast Fourier transform, triangular bandpass filter, calculation of the logarithmic energy of each bandpass filter group output, discrete cosine transform, and other steps to obtain the mfcc features (Mel-scale Frequency Cepstral Coefficients).

[0089] Among them, the mfcc feature is a speech feature that can be recognized by a computer.

[0090] Step 202: clustering the multiple speech segments to be processed according to the respective MFCC features of the multiple speech segments to be processed, so as to divide the multiple speech segments to be processed into two speech piles.

[0091] In the embodiment of the present application, the process of obtaining two speech piles is achieved through a clustering process, which is a bottom-up clustering algorithm in agglomerative-cluster. Figure 3 , specifically including the following sub-steps:

[0092] Sub-step 2021: Perform VAD silence detection on the MFCC features of the multiple speech segments to be processed, and extract the MFCC features of multiple speech activity segments in the speech segments to be processed.

[0093] In this step, since the speech to be processed includes not only the speaker's voice, but also the pauses of the speaker and the silence periods caused by switching between different speakers, the VAD silence detection technology (Voice Activity Detection, VAD), also known as voice endpoint detection or voice boundary detection, etc., can be used to eliminate long silence periods from the speech segment to be processed and extract multiple voice activity segments.

[0094] The speech activity segment refers to a speech segment in which a speaker speaks in the speech segment to be processed, rather than a speech segment in which no speaker speaks.

[0095] By extracting the speech activity segment from the speech segment to be processed, the silence period can be identified and eliminated to obtain a more accurate speech segment.

[0096] Sub-step 2022: Clustering the multiple voice activity segments according to the MFCC features of the multiple voice activity segments to divide the multiple voice activity segments into two voice piles.

[0097] In this step, see Figure 3 First, the trained UBM model is used to process the MFCC features of multiple speech activity segments and extract the Ivctor features of multiple speech activity segments; secondly, the trained PLDA model can be used to score the Ivctor features of every two speech activity segments in multiple speech activity segments to obtain the PLDA scores. Based on the PLDA scores, multiple speech activity segments are clustered to obtain two speech piles.

[0098] Among them, ivertor features are information extracted from speech activity segments that can represent speech features. Specifically, they may include short-time average energy, short-time average amplitude, short-time average zero-crossing rate, resonance peaks, glottal waves, speech rate, etc. of speech activity segments, as long as they can represent the speech features of different speech activity segments.

[0099] Among them, the plda score refers to the dimensional distance between each ivert feature. The larger the plda score, the greater the difference between the two ivert features, and the smaller the plda score, the smaller the difference between the two ivert features.

[0100] Specifically, when clustering multiple voice activity segments based on the PLDA score, the PLDA model can be used to score each two voice activity segments in the multiple voice activity segments, and the two voice activity segments with scores less than the threshold are clustered into one voice pile, and the two voice activity segments with scores greater than or equal to the threshold are clustered into different voice piles. The threshold can be 10 points, 20 points, etc., and this application does not limit it here.

[0101] In specific application, for example, there are voice activity segments 1, 2, 3, 4, and 5. Since the scoring is done by scoring every two voice activity segments, the above 5 voice activity segments will have 7 combinations: (1 and 2), (1 and 3), (1 and 4), (1 and 5), (2 and 3), (2 and 4), (2 and 5), (3 and 4), and (3 and 5).

[0102] When performing a clustering, the plda model is first used to score the above 7 combinations respectively. Finally, the plda scores of (1 and 2), (1 and 5), and (2 and 5) are all less than 10 points; the plda score of (3 and 5) is less than 10 points, and the plda scores of (1 and 4), (2 and 3), (2 and 4), and (3 and 4) are all greater than 90 points. Then, the speech activity segments 1, 2, and 5 can be preliminarily divided into the first speech pile; and the speech activity segments 3 and 4 can be preliminarily divided into the second speech pile.

[0103] Thus, the voice activity segments 1, 2, and 5 in the first voice pile are preliminarily determined to be the voice segments uttered by user A; and the voice activity segments 3 and 4 in the second voice pile are preliminarily determined to be the voice segments uttered by user B.

[0104] Step 203: two groups of first long voice segments whose voice duration is greater than a first preset duration are respectively selected from the two voice piles.

[0105] In the embodiment of the present application, after clustering, both the first voice pile and the second voice pile may contain short voice segments, and registering the short voice segments as voiceprint samples will undoubtedly reduce the accuracy of voiceprint recognition.

[0106] Therefore, the first long voice segments with a voice duration longer than the first preset duration, such as the first long voice segments 2 and 5, can be screened out from the first voice pile; the first long voice segments with a voice duration longer than the first preset duration, such as the first long voice segments 3 and 4, can also be screened out from the second voice pile.

[0107] Through the screening in step 203, two groups of first long voice segments can be preliminarily screened out from the first voice pile and the second voice pile, so that the first long voice segments registered as samples do not contain short voice segments as much as possible, thereby improving the accuracy of voiceprint recognition.

[0108] Among them, since two groups of first long voice segments with voice durations longer than the first preset duration will be preliminarily screened out from the two voice piles, the size of the first preset duration will affect the accuracy of the subsequent step 204 secondary clustering and the quality of the voiceprint registration. Therefore, the size of the first preset duration needs to be set.

[0109] Specifically, when the first preset duration is larger, the number of first long voice segments screened out and registered as voiceprint samples is smaller. In this way, although the voiceprint registration sample can include first long voice segments with longer voice duration to ensure the accuracy of secondary clustering, the quality of voiceprint registration will be reduced due to the small number of first long voice segments registered as samples.

[0110] When the first preset duration is shorter, more first long voice segments are screened out and registered as voiceprint samples. In this way, although the number of first long voice segments registered as voiceprint samples is larger, the voiceprint registration quality is improved; however, the registered voiceprint samples contain more first long voice segments with shorter voice durations, which will reduce the accuracy of secondary clustering.

[0111] Therefore, in the embodiment of the present application, the first preset time length is set according to the accuracy of the secondary clustering; wherein the first preset time length is directly proportional to the accuracy of the secondary clustering.

[0112] For example, the first preset duration is set within 1-2 seconds, so that the first long voice segment selected will not cause the quality of voiceprint registration to decrease, and will not cause the accuracy of secondary clustering to decrease.

[0113] Step 204: performing secondary clustering on the two groups of first long speech segments according to the respective ivocar features of the two groups of first long speech segments to obtain two groups of second long speech segments.

[0114] Among them, when performing a clustering in step 202, the clustering process will be affected by the short voice segments, resulting in low clustering accuracy. Then the first voice pile divided may contain the to-be-processed voice segments that originally belonged to the second voice pile. Similarly, the second voice pile may also contain the to-be-processed voice segments that originally belonged to the first voice pile. Correspondingly, there will be a mutual mixing of the two groups of first long voice segments evolved from the second voice pile, which will cause the two registered voiceprint samples to mix with each other, affecting the accuracy of voiceprint recognition. Therefore, in order to make the division of the two groups of first long voice segments more accurate, the embodiment of the present application also performs secondary clustering.

[0115] When performing secondary clustering, the following sub-steps are included:

[0116] Sub-step 2041: Using the PLDA model, score the ivert features of every two first long speech segments to obtain the PLDA score between every two first long speech segments.

[0117] In this step, similar to sub-step 2022, the trained PLDA model can be used to score the ivictor features of every two first long speech segments in the first speech pile; and score the ivictor features of every two first long speech segments in the second speech pile.

[0118] For example, when the PLDA model is used again to score the first long speech segments 2, 3, 4, and 5 selected using the first preset time length in step 203, the four first long speech segments can have five combinations: (2 and 3), (2 and 4), (2 and 5), (3 and 4), and (3 and 5).

[0119] After using the PLDA model to score the ivocar features of every two first long speech segments, the PLDA scores of (2 and 3), (2 and 4), and (3 and 4) are less than 10 points, and the PLDA scores of (2 and 5) and (3 and 5) are greater than 90 points.

[0120] Sub-step 2042: performing secondary clustering on the first long speech segment according to the plda score to obtain two groups of second long speech segments.

[0121] In this step, after scoring the ivocar features of each of the two first long speech segments in the two speech piles, the two first long speech segments with scores below the threshold can be clustered into one speech pile again, while the two first long speech segments with scores above the threshold can be divided into different speech piles again.

[0122] For example, the first long speech segments (2 and 3), (2 and 4), and (3 and 4) are clustered into a first speech pile, and the first long speech segment 5 is clustered into a second speech pile.

[0123] Through the setting of secondary clustering, the two groups of first long voice segments can be divided for a second time to be divided into two groups of second long voice segments, so that the voice segments divided in the two groups of first long voice segments will not be mixed with each other, that is, the voice segments in the first long voice segments in the first voice pile will not be divided into the first long voice segments in the second voice pile, and vice versa, so that the two registered voiceprint samples can accurately belong to two different users. Naturally, when the short voice segments are matched with the two samples respectively in the subsequent process, the accuracy of voiceprint recognition can also be improved based on accurate voiceprint samples.

[0124] The two groups of the second long speech segments are obtained by secondary clustering of the two groups of the first long speech segments.

[0125] Step 205: Register the two groups of second long voice segments into two voiceprint samples respectively.

[0126] Step 206: Match the two voiceprint samples with two groups of short voice segments whose voice duration is less than a second preset duration respectively to determine the voiceprint samples to which the two groups of short voice segments belong, wherein the second preset duration is less than or equal to the first preset duration.

[0127] Through a speaker separation method provided in an embodiment of the present application, the two groups of screened first long voice segments can be clustered again to divide two groups of second long voice segments. After using accurate second long voice segments to register two voiceprint samples respectively, the accuracy of voiceprint recognition can be improved.

[0128] Based on the same inventive concept, the embodiment of the present application further provides a speaker separation device, the device comprising:

[0129] A division module, used for acquiring a plurality of speech segments to be processed, and dividing the plurality of speech segments to be processed into two speech piles;

[0130] A screening module, used to screen out two groups of first long voice segments whose voice duration is greater than a first preset duration from the two voice piles respectively;

[0131] A registration module, used to register the two groups of first long voice segments into two voiceprint samples respectively, wherein different voiceprint samples represent different voice users;

[0132] The matching module is used to match the two voiceprint samples with two groups of short voice segments whose voice duration is less than a second preset duration, so as to determine the voiceprint samples to which the two groups of short voice segments belong respectively, wherein the second preset duration is less than or equal to the first preset duration.

[0133] Optionally, the registration module includes:

[0134] A secondary clustering module, used for performing secondary clustering on the two groups of first long speech segments according to respective ivictor features of the two groups of first long speech segments, so as to obtain two groups of second long speech segments;

[0135] The first registration submodule is used to register the two groups of second long voice segments into two voiceprint samples respectively.

[0136] Optionally, the secondary clustering module includes:

[0137] A first scoring module is used to score the ivocar features of each two first long speech segments by using the PLDA model to obtain a PLDA score between each two first long speech segments;

[0138] The secondary clustering submodule is used to perform secondary clustering on the first long speech segment according to the plda score to obtain two groups of second long speech segments.

[0139] Optionally, the division module includes:

[0140] A feature extraction module, used to extract features from the plurality of speech segments to be processed respectively, to obtain the mfcc features of the plurality of speech segments to be processed respectively;

[0141] The primary clustering module is used to perform primary clustering on the multiple speech segments to be processed according to the respective MFCC features of the multiple speech segments to be processed, so as to divide the multiple speech segments to be processed into two speech piles.

[0142] Optionally, the primary clustering module includes:

[0143] A silence detection module, used to perform VAD silence detection on the MFCC features of the multiple voice segments to be processed, and extract the MFCC features of multiple voice activity segments in the voice segments to be processed;

[0144] The primary clustering submodule is used to perform primary clustering on the multiple voice activity segments according to the MFCC features of the multiple voice activity segments, so as to divide the multiple voice activity segments into two voice piles.

[0145] Optionally, the device further comprises:

[0146] The setting module is used to set the first preset time length according to the accuracy of the secondary clustering; wherein the first preset time length is directly proportional to the accuracy of the secondary clustering.

[0147] Based on the same inventive concept, an embodiment of the present application further provides an electronic device, including:

[0148] one or more processors; and

[0149] One or more machine-readable media having instructions stored thereon, when executed by the one or more processors, enable the electronic device to perform a talker separation method as described above.

[0150] Based on the same inventive concept, an embodiment of the present application also provides one or more machine-readable storage media, characterized in that instructions are stored thereon, and when executed by one or more processors, the processors execute a speaker separation method as described above.

[0151] As for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0152] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0153] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, devices, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0154] The embodiments of the present application are described with reference to the flowcharts and / or block diagrams of the methods, terminal devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing terminal device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0155] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing terminal device to operate in a specific manner, so that the instructions stored in the computer readable memory produce a manufactured product including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0156] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device so that a series of operating steps are executed on the computer or other programmable terminal device to produce a computer-implemented process, thereby providing instructions for executing on the computer or other programmable terminal device to implement the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0157] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the embodiments of the present application.

[0158] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or terminal device including the elements.

[0159] The above is a detailed introduction to a speaker separation method, device, electronic device and storage medium provided by the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea. At the same time, for a person skilled in the art, according to the idea of ​​the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A method for separating speakers, characterized in that: The method comprises: Acquire a plurality of speech segments to be processed, and divide the plurality of speech segments to be processed into two speech piles; Selecting two groups of first long voice segments whose voice duration is greater than a first preset duration from the two voice piles respectively; Registering the two groups of first long voice segments into two voiceprint samples respectively, wherein different voiceprint samples represent different voice users; The two voiceprint samples are matched respectively with two groups of short voice segments whose voice duration is less than a second preset duration to determine the voiceprint samples to which the two groups of short voice segments belong, wherein the second preset duration is less than or equal to the first preset duration.

2. The method according to claim 1, characterized in that: Registering the two groups of first long voice segments into two voiceprint samples respectively includes: According to the ivocar features of the two groups of first long speech segments, the two groups of first long speech segments are clustered again to obtain two groups of second long speech segments; The two groups of second long voice segments are registered as two voiceprint samples respectively.

3. The method according to claim 2, characterized in that According to the ivocar features of the two groups of first long speech segments, the two groups of first long speech segments are clustered twice to obtain two groups of second long speech segments, including: Using the PLDA model, the ivocar features of each of the two first long speech segments are scored to obtain the PLDA score between each of the two first long speech segments; According to the plda score, the first long speech segment is clustered twice to obtain two groups of second long speech segments.

4. The method according to claim 1, characterized in that: Acquire a plurality of speech segments to be processed, and divide the plurality of speech segments to be processed into two speech piles, including: Extracting features from the plurality of speech segments to be processed respectively to obtain the mfcc features of the plurality of speech segments to be processed; According to the respective MFCC features of the multiple speech segments to be processed, the multiple speech segments to be processed are clustered once to divide the multiple speech segments to be processed into two speech piles.

5. The method according to claim 4, characterized in that According to the respective MFCC features of the plurality of speech segments to be processed, clustering the plurality of speech segments to be processed is performed once to divide the plurality of speech segments to be processed into two speech piles, including: Performing VAD silence detection on the MFCC features of the multiple speech segments to be processed, and extracting the MFCC features of multiple speech activity segments in the speech segments to be processed; According to the MFCC features of the multiple voice activity segments, the multiple voice activity segments are clustered to divide the multiple voice activity segments into two voice piles.

6. The method according to claim 2, characterized in that The method further comprises: The first preset time length is set according to the accuracy of the secondary clustering; wherein the first preset time length is directly proportional to the accuracy of the secondary clustering.

7. A speaker separation device, characterized in that: The device comprises: A division module, used for acquiring a plurality of speech segments to be processed, and dividing the plurality of speech segments to be processed into two speech piles; A screening module, used to screen out two groups of first long voice segments whose voice duration is greater than a first preset duration from the two voice piles respectively; A registration module, used to register the two groups of first long voice segments into two voiceprint samples respectively, wherein different voiceprint samples represent different voice users; The matching module is used to match the two voiceprint samples with two groups of short voice segments whose voice duration is less than a second preset duration, so as to determine the voiceprint samples to which the two groups of short voice segments belong respectively, wherein the second preset duration is less than or equal to the first preset duration.

8. The device according to claim 7, characterized in that The registration module includes: A secondary clustering module, used for performing secondary clustering on the two groups of first long speech segments according to respective ivictor features of the two groups of first long speech segments, so as to obtain two groups of second long speech segments; The first registration submodule is used to register the two groups of second long voice segments into two voiceprint samples respectively.

9. An electronic device, characterized in that: include: one or more processors; and One or more machine-readable media having instructions stored thereon, when executed by the one or more processors, cause the electronic device to perform the speaker separation method as described in any one of claims 1-6.

10. One or more machine-readable storage media, characterized in that Instructions are stored thereon, which, when executed by one or more processors, cause the processors to perform the speaker separation method according to any one of claims 1-6.

Images