A self-learning method and system for voiceprint recognition

Abstract

The application discloses a self-learning method for voiceprint recognition, comprising the following steps: S1, collecting speech by using a recording device to obtain audio of a speaker; S2, identifying the identity of the speaker by using a voiceprint recognition module based on the audio in S1, and entering voiceprint self-learning after audio recognition; and S3, performing self-learning threshold judgment before voiceprint self-learning, performing voiceprint learning rate calculation according to a set learning rate formula when the condition is met, fusing the obtained voiceprint with old voiceprint to obtain new voiceprint, and storing the new voiceprint in a voiceprint library. The application is used for solving the problem of low voiceprint recognition rate caused by changes in environmental conditions and physiological conditions of a speaker and insufficient voiceprint registration, and the longer the time, the worse the recognition effect. The main steps include voiceprint self-learning threshold judgment, self-learning rate calculation, new and old voiceprint fusion and voiceprint library updating.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of voiceprint recognition, in particular to a self-learning method and system for voiceprint recognition. BACKGROUND

[0002] Due to the differences in vocal cord structure, living environment, emotion and other factors, the voice of each person has unique characteristics, which is technically referred to as voiceprint. Voiceprint can be used to distinguish the identity of the speaker, and has many applications in identity login, smart home, smart cockpit, smart office and other fields.

[0003] The existing voiceprint recognition system has the following two defects: first, the voiceprint system is subject to many conditions. Due to the differences in environmental conditions such as the environment, the model of the recording device, the angle between the recording device and the speaker during recording, the distance of the sound collection, and the environmental noise, and due to the physiological conditions such as the age change of the speaker, different time periods, and tone changes, the voice characteristics of the speaker themselves also change. The irregular changes in environmental conditions and physiological conditions objectively lead to poor robustness of voiceprint application, and there is a great difference between actual application and laboratory testing. Second: when registering voiceprint, only limited audio can be selected, and the voiceprint features obtained by registration cannot fully describe the voice characteristics of the speaker. No matter what model is used to extract the voiceprint, a stable and strong anti-interference voiceprint cannot be obtained, resulting in poor generalization ability in actual use. Due to the above two reasons, the accuracy of voiceprint recognition will become lower and lower with the change of time and environment, until it reaches a completely unusable state. SUMMARY

[0004] The purpose of the present application is to provide a self-learning method for voiceprint recognition, which solves the problem of decreasing recognition accuracy in actual use of voiceprint recognition caused by incomplete environmental conditions, physiological conditions and voiceprint registration.

[0005] The purpose of the present application can be achieved by the following technical solutions:

[0006] A self-learning method for voiceprint recognition, comprising the following steps:

[0007] S1: using a recording device to collect speech to obtain audio of the speaker;

[0008] S2: based on the audio in S1, identifying the identity of the speaker through a voiceprint recognition module, and after audio recognition, entering voiceprint self-learning;

[0009] S3: before voiceprint self-learning, performing self-learning threshold judgment, and when the condition is met, calculating the voiceprint learning rate according to a set learning rate formula, fusing the obtained voiceprint with the old voiceprint to obtain a new voiceprint, and storing the new voiceprint in a voiceprint library;

[0010] In S3, the condition for voiceprint self-learning threshold judgment is:

[0011] x top1 >x learn

[0012] or

[0013]

[0014] wherein, Δx is an adjustable parameter, the range is [0, 0.4], x top1 is the highest similarity of the speaker calculated in the voiceprint recognition module, x top is the second highest similarity of the speaker calculated in the voiceprint recognition module, x learn is a self-learning threshold, and one of the two conditions is met;

[0015] The learning rate formula is:

[0016]

[0017] wherein k(x) is the calculated learning rate, e is a natural constant, x is the highest similarity of the speaker calculated in the voiceprint recognition module, r is a learning constant in the range of [0, 1], and a and b are preset adjustment parameters;

[0018] The acquisition process of the new voiceprint is:

[0019] v neω =v i ·k(x)+υ old ·(1-v i ·k(x))

[0020] wherein, v new is the speaker voiceprint feature vector after self-learning, v i is the voiceprint feature vector of the recognized audio, k(x) is the learning rate calculated by the learning rate formula, and v old is the original speaker voiceprint feature vector before self-learning.

[0021] As a further scheme of the application: in S1, the recording device for collecting audio includes a conference system microphone, a mobile phone, a lapel microphone, and a voice mouse.

[0022] As a further scheme of the application: in S1, the audio of the speaker is a segment of voice audio of the speaker or a continuous recording audio stream.

[0023] As a further scheme of the application: in S1, the audio is preprocessed before entering the voice recognition module, including hardware-level echo cancellation and noise reduction processing on the audio.

[0024] As a further scheme of the present application: in S2, the voiceprint recognition module is built-in a voiceprint recognition model, the voiceprint recognition model extracts a feature vector from the audio to obtain an audio voiceprint feature vector v i ;

[0025] The cosine method is improved to calculate the voiceprint feature vector v i The similarity of the voiceprint with the registered speaker in the voiceprint library is calculated, the highest similarity is recorded as x top1 , the second highest similarity is recorded as x to After traversal, if there is a voiceprint with x top1 reaching the threshold x re , the identity is marked, and the voiceprint self-learning is entered, otherwise, the recognition failure is output.

[0026] As a further scheme of the present application: in S3, the generated new voiceprint v new is replaced with the old voiceprint v old in the voiceprint library, a self-learning flow of a standard process is completed, and the old voiceprint v old in the voiceprint library is updated.

[0027] As a further scheme of the present application: a voiceprint recognition system comprises:

[0028] An audio acquisition module is configured to acquire a speech audio of a speaker or continuously record an audio stream to obtain the audio of the speaker;

[0029] A voiceprint recognition module is configured to identify the identity of the speaker, calculate a feature vector of the audio and a voiceprint similarity, and enter voiceprint self-learning after the identification;

[0030] A voiceprint self-learning module is configured to perform voiceprint self-learning, perform self-learning threshold judgment before the voiceprint self-learning, perform voiceprint learning rate calculation according to a set learning rate formula when a condition is met, fuse the old voiceprint with the obtained voiceprint to obtain a new voiceprint, and store the new voiceprint in the voiceprint library.

[0031] As a further scheme of the present application: the audio acquisition module comprises an audio acquisition unit and an audio processing unit;

[0032] The audio acquisition unit is an audio acquisition device;

[0033] The audio processing unit is configured to perform echo cancellation and noise reduction processing on the audio.

[0034] As a further scheme of the present application: the voice recognition module comprises a voiceprint recognition model, the voiceprint recognition module is configured to extract a feature vector from the audio, and calculate the audio and the registered speaker in the voiceprint library for comparison;

[0035] If there is a voiceprint reaching the threshold after traversal, the identity of the speaker is marked, and the voiceprint self-learning module is entered.

[0036] Otherwise, output recognition failure.

[0037] As a further scheme of the present application: the functions of the voiceprint library based on voiceprint include storage, query, update and deletion.

[0038] The present application has the following advantages: the present application uses a recording device to collect voice, and after the voiceprint recognition system identifies the speaker's identity, the voiceprint self-learning module is entered, in which the self-learning threshold is first judged, the learning rate is calculated if the condition is met, and finally the voiceprint self-learning is performed. With the use of the voiceprint recognition system, the voiceprint self-learning module is repeatedly called, which is equivalent to multiple registrations of the speaker's voiceprint. The voiceprint characteristics of the speaker are recorded comprehensively, and with the increase of use time, the voiceprint of the speaker will be more complete, which enhances the stability and anti-interference ability of the voiceprint, and improves the voiceprint recognition accuracy under different use environment conditions and physiological conditions.

[0039] The problem of low voiceprint recognition accuracy under different use environment conditions and physiological conditions is solved, and the problem of insufficient voiceprint registration is also solved. The longer the voiceprint system is used, the higher the recognition accuracy is, and the robustness is gradually improved, achieving the effect of being commonly used and constantly new. BRIEF DESCRIPTION OF DRAWINGS

[0040] The present application will be further described below with reference to the accompanying drawings.

[0041] Figure 1 is a flowchart of the present application;

[0042] Figure 2 is a voiceprint self-learning flowchart in the present application;

[0043] Figure 3 is a program block diagram of the present application;

[0044] Figure 4 is a voiceprint self-learning rate example curve diagram of the present application. DETAILED DESCRIPTION

[0045] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0046] Please refer to Figures 1-2 The present application is a voiceprint recognition self-learning method, which includes the following steps:

[0047] S1: voice collection is performed using a recording device to obtain audio of a speaker;

[0048] S2: based on the audio in S1, the identity of the speaker is identified through a voiceprint recognition module, and after audio recognition, voiceprint self-learning is entered;

[0049] S3: self-learning threshold judgment is performed before voiceprint self-learning, and if the conditions are met, the voiceprint learning rate is calculated according to the set learning rate formula, the obtained voiceprint is fused with the old voiceprint to obtain a new voiceprint, and the new voiceprint is stored in a voiceprint library;

[0050] In S3, the conditions for voiceprint self-learning threshold judgment are:

[0051] x top1 >x learn

[0052] or

[0053]

[0054] wherein Δx is an adjustable parameter, the range is [0, 0.4], x top1 is the highest similarity of the speaker calculated in the voiceprint recognition module, x top is the second highest similarity of the speaker calculated in the voiceprint recognition module, and x learn is the self-learning threshold, and one of the two conditions is met;

[0055] The learning rate formula is:

[0056]

[0057] wherein k(x) is the calculated learning rate, e is a natural constant, x is the highest similarity of the speaker calculated in the voiceprint recognition module, r is a learning constant, the range is [0, 1], and a and b are preset adjustment parameters;

[0058] The process of obtaining a new voiceprint is:

[0059] υ neω =v i ·k(x)+v old ·(1-v i ·k(x))

[0060] wherein v new is the speaker voiceprint feature vector after self-learning, v i is the voiceprint feature vector of the recognized voice audio, k(x) is the learning rate calculated by the learning rate formula, and v old is the original speaker voiceprint feature vector before self-learning.

[0061] In S1, the recording equipment includes, but is not limited to, conference system microphones, mobile phones, lavalier microphones, voice mice and other devices capable of recording audio.

[0062] In S1, the speaker's audio is either a segment of the speaker's voice or a continuously recorded audio stream;

[0063] In S1, the audio is preprocessed before entering the sound recognition module. The preprocessing includes hardware-level echo cancellation and noise reduction to improve the audio quality.

[0064] In S2, the process of the voiceprint recognition module identifying the speaker involves inputting the received audio into the recognition model of the voiceprint recognition module, using the built-in voiceprint recognition model to extract feature vectors from the audio, and obtaining the audio voiceprint feature vector v. i Calculate the voiceprint feature vector v i Similarity is calculated between the speaker and those already registered in the voiceprint database using the cosine similarity method, with the highest similarity score denoted as x. top1 The second highest similarity is denoted as x. top2 If x is found after traversal top1 Reaching threshold x re If the voiceprint is recognized, the user's identity is identified, and the user continues to the voiceprint self-learning module; otherwise, the recognition fails.

[0065] In this embodiment, ECAPA-TDNN is used as the voiceprint recognition model.

[0066] In S3, the newly generated voiceprint v will be stored in the voiceprint library. new Replace old voiceprint v old Based on its position in the voiceprint database, it completes a standard self-learning process, while simultaneously updating the v in the voiceprint database. old This allows the effects in the voiceprint recognition module to be updated in real time.

[0067] Example 2

[0068] See Figure 3 A voiceprint recognition system includes a sound acquisition module, a voiceprint recognition module, a voiceprint self-learning module, and a voiceprint library;

[0069] The sound acquisition module is used to acquire the speaker's audio.

[0070] The voiceprint recognition module is used for speaker identification and to calculate the learning parameters of the voice self-learning module;

[0071] The learning parameters include the feature vector of the audio and the voiceprint similarity.

[0072] The voiceprint self-learning module includes the voiceprint self-learning operation logic to update the voiceprint;

[0073] The voiceprint library provides storage, query, update and deletion functions of voiceprints.

[0074] The audio collected by the sound collection module is a segment of speech audio of the speaker or a continuous recording audio stream.

[0075] The sound collection device of the sound collection module includes but is not limited to a conference system microphone, a mobile phone, a lapel microphone, a voice mouse and other devices capable of recording audio. The sound collection device performs echo cancellation and noise reduction processing on the collected audio to improve the audio quality, and then transmits the processed audio audio-1 to the sound recognition module.

[0076] After receiving the audio audio-1 transmitted by the sound collection module, the voiceprint recognition module inputs it into the recognition model to extract its feature vector using the built-in voiceprint recognition model, and obtains its voiceprint feature vector v i . Then, the voiceprint feature vector v i is calculated for similarity with the registered speakers in the voiceprint library. The similarity calculation uses the cosine method, the highest similarity is recorded as x top1 , the second highest similarity is recorded as x to , and after traversal, if there is a voiceprint with x top1 reaching the threshold x re , the identity is marked, and the voiceprint self-learning module is entered, otherwise the recognition fails is output.

[0077] The voiceprint recognition model is an ECAPA-TDNN voiceprint recognition model.

[0078] After entering the voiceprint self-learning module, the self-learning threshold x learn is first determined, the learning rate is calculated after the condition is met, and then the self-learning of the voiceprint is performed. The old voiceprint v old is fused with v i to form a new voiceprint v new , which is stored in the voiceprint library.

[0079] The specific process is as follows:

[0080] The self-learning threshold needs to be determined because the speaker voiceprint has learned the voiceprint with a wrong marked identity, which will cause the voiceprint feature to be biased to the wrong speaker, and the cumulative error will cause the speaker voiceprint to fail.

[0081] Condition 1:

[0082] x top1 > x learn .

[0083] Condition 2

[0084]

[0085] wherein, Δx is an adjustable parameter, the range is [0, 0.4], the larger, the more relaxed, the more stringent;

[0086] The above two conditions meet one can be determined, condition 2 is used when the number of voiceprint library is small, which can greatly improve the efficiency of voiceprint self-learning;

[0087] The voiceprint self-learning rate is calculated, and the calculation formula is as follows:

[0088]

[0089] wherein, k(x top1 ) is the calculated learning rate, e is a natural constant, x top1 is the highest similarity of the speaker calculated in the voiceprint recognition module, r is a learning constant (range is [0, 1]) for adjusting the learning efficiency, the higher the learning efficiency, a, b are two adjustable parameters, which are adjusted according to the needs of the project, mainly adjusting the effective interval of the learning function curve on the x-axis;

[0090] In a specific embodiment, a = 20, b = 14, r = 0.1, then the voiceprint self-learning curve is shown in Figure 4 Through the curve, it can be observed that the main change interval x top is located in [0.4, 1], generally, x learn is greater than 0.4, so when x to is less than 0.4, it will not enter this function, so again only to explain the part greater than 0.4; when x top1 is located in [0.4, 0.6], the learning rate decreases slowly, which is set because when x top1 is smaller, the similarity between the old speaker voiceprint v old and the newly generated v i is lower, at this time we expect v old to learn more differences in v i , we think that [0.4, 0.6] is a more appropriate learning interval, so the decrease is slow; when x top1 is located in (0.6, 0.8], the learning curve decreases rapidly, because as x top1 becomes larger, the learnable differences in v i become less sharply; when x top1 is located in (0.8, 1], the learning rate gradually tends to 0, the meaning of self-learning in this interval is not great. The above is only an example of a scene, the three variables in different scenes can be adjusted accordingly.

[0091] The voiceprint self-learning module performs self-learning of the voiceprint, and the old voiceprint vold With v i Algorithm fusion, generate a new voiceprint v new , the new voiceprint v new contains a small part of the characteristics of v i , v new When encountering this v i The sound environment and physiological conditions can more accurately identify the speaker identity, the specific learning formula is as follows:

[0092] υ new =υ i ·k(x top1 )+υ o l d ·(1-v i ·k(x top1 ))

[0093] Where v new Is the speaker's voiceprint feature vector after self-learning, v i Is the voiceprint feature vector of the recognized voice audio, k(x top1 ) is the learning rate calculated by the learning rate formula, v old Is the original speaker's voiceprint feature vector before self-learning;

[0094] v old , v i , v new In essence, a set of feature vectors, which can be generally matrix operation, complete the required operation of the formula.

[0095] Replace the newly generated voiceprint v new Replace v old In the voiceprint library, complete a standard process of self-learning flow. At the same time, update v old In memory, the effect of the voiceprint recognition module is updated in real time.

[0096] Those skilled in the art should understand that the above discussion of any embodiment is only exemplary and is not intended to limit the scope of the disclosure (including claims); under the idea of the disclosure, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other changes of different aspects of the disclosure as described above. In order to be brief, they are not provided in detail.

[0097] The above describes one embodiment of the present application in detail, but the content described is only the preferred embodiment of the present application, and cannot be considered as limiting the scope of the present application. Any equivalent changes and improvements made in the scope of the present application should still belong to the patent scope of the present application.

Claims

1. A self-learning method for voiceprint recognition, characterized in that, Includes the following steps: S1: Use recording equipment to capture the speaker's audio; S2: Based on the audio in S1, the speaker's identity is identified through the voiceprint recognition module, and after audio recognition, the voiceprint self-learning is initiated. S3: Before voiceprint self-learning, a self-learning threshold is determined. If the condition is met, the voiceprint learning rate is calculated according to the set learning rate formula. The new voiceprint is obtained by fusing the old voiceprint with the new voiceprint and storing the new voiceprint in the voiceprint library. In S3, the condition for determining the voiceprint self-learning threshold is: or in, It is an adjustable parameter with a range of [0, 0.4]. It is the highest speaker similarity calculated in the voiceprint recognition module. It is the second highest speaker similarity calculated in the voiceprint recognition module. yes Either of the two conditions must be met; The learning rate formula is: in Here, is the calculated learning rate, e is the natural constant, x is the highest speaker similarity calculated in the voiceprint recognition module, r is the learning constant ranging from [0,1], and a and b are preset adjustment parameters. The process of obtaining a new voiceprint is as follows: in, It is the speaker's voiceprint feature vector after self-learning. It is the voiceprint feature vector for audio recognition. The learning rate is calculated using the learning rate formula. It is the original speaker's voiceprint feature vector before self-learning.

2. The self-learning method for voiceprint recognition according to claim 1, characterized in that, In S1, the recording devices used to capture audio include a conference system microphone, a mobile phone, a lavalier microphone, and a voice mouse.

3. The self-learning method for voiceprint recognition according to claim 1, characterized in that, In S1, the speaker's audio is either a segment of the speaker's voice or a continuously recorded audio stream.

4. The self-learning method for voiceprint recognition according to claim 1, characterized in that, In S1, the audio is preprocessed before entering the sound recognition module, including hardware-level echo cancellation and noise reduction.

5. The self-learning method for voiceprint recognition according to claim 1, characterized in that, In S2, the voiceprint recognition module has a built-in voiceprint recognition model. This model extracts feature vectors from the audio data to obtain audio voiceprint feature vectors. ; Improving cosine method for calculating voiceprint feature vector The similarity score with speakers already registered in the voiceprint database is used to determine the highest similarity score. The second highest similarity is denoted as If after traversal there are Reaching the threshold If the voiceprint is recognized, the user's identity is identified and the user enters the voiceprint self-learning process; otherwise, the recognition fails.

6. The self-learning method for voiceprint recognition according to claim 1, characterized in that, In S3, the new voiceprint will be generated Replace old voiceprint Based on its position in the voiceprint database, it completes a standard self-learning process while updating old voiceprints in the database. .

7. A voiceprint recognition system based on the self-learning method of voiceprint recognition as described in claim 1, characterized in that, include: The audio acquisition module is used to acquire a segment of the speaker's voice or to continuously record an audio stream to obtain the speaker's audio. The voiceprint recognition module is used to identify the speaker's identity and calculate the feature vector of the audio and the voiceprint similarity. After recognition, it enters the voiceprint self-learning stage. The voiceprint self-learning module is used to perform voiceprint self-learning. Before voiceprint self-learning, a self-learning threshold is judged. If the condition is met, the voiceprint learning rate is calculated according to the set learning rate formula. The obtained voiceprint is fused with the old voiceprint to obtain a new voiceprint, and the new voiceprint is stored in the voiceprint library.

8. A voiceprint recognition system according to claim 7, characterized in that, The audio acquisition module includes an audio acquisition unit and an audio processing unit; The audio acquisition unit is an audio acquisition device; The audio processing unit is used for audio echo cancellation and noise reduction.

9. A voiceprint recognition system according to claim 7, characterized in that, The voice recognition module includes a voiceprint recognition model, which is used to extract feature vectors from audio and compare the audio with the speakers registered in the voiceprint database. If a voiceprint reaches the threshold after traversal, the speaker's identity is marked and the system enters the voiceprint self-learning module. Otherwise, the output will fail to recognize the data.

10. A voiceprint recognition system according to claim 7, characterized in that, The voiceprint database's voiceprint-based functions include storage, querying, updating, and deletion.

Images