CN110556114B - Caller identification method and device based on attention mechanism - Google Patents

Abstract

The invention discloses a speaker identification method and a speaker identification device based on an attention mechanism, which comprise the following steps: collecting the call records of a plurality of tested callers and the call records of the tested callers; establishing a speaker voice library according to the call record corresponding to the tested speaker; training the voice of the tested talking person by adopting a neural network based on attention to obtain a training model; storing the call record of the test caller to obtain a record file; and identifying whether the tested speaker is the target speaker or not by using the recording file by adopting the training model. The voice of the tested caller is trained by adopting the attention-based neural network to obtain a training model, the tested caller is identified by adopting the training model, the consistency of the master corresponding to the dialing number is confirmed, the potential communication safety hazard caused by the imitation of the identity of the caller is avoided, and the information safety in the calling process is further improved.

Description

Caller identification method and device based on attention mechanism

technical field

The present invention relates to the field of speech recognition, in particular to a method and device for identifying a caller based on an attention mechanism.

Background technique

Voice is the most direct and most important way of communication in human production and life. Human voice contains semantic information, language or dialect information, channel information, etc. With the continuous progress and development of computer technology and the advent of the Internet age, there are more and more measures to disguise the identity of the speaker.

During the communication process, determining the identity of the speaker is beneficial to determine the security of the communication. Speaker recognition methods in the prior art include using the approximation of KL divergence as a measure of similarity between speakers, using BP neural network for speaker recognition, and using mixed features of MFCC and GFCC for speaker recognition. However, most of the application scenarios of technologies are based on smart homes, such as sweeping robots, etc., to identify whether the speaker is the owner of the speaker based on the voice. In the prior art, there is no speaker recognition method in the communication process, and no consideration is given to the silence process of the speech in the communication process.

Most of the voice recognition methods in the prior art are suitable for the scene of recognizing the voice signal obtained directly from the outside world, and lack the voice recognition method for the caller during the communication process, and there is a security risk that the caller's identity can be counterfeited.

SUMMARY OF THE INVENTION

The purpose of the present invention is to propose a method and device for identifying a caller based on an attention mechanism, so as to solve the above-mentioned technical problems.

For achieving the above object, the present invention provides the following scheme:

A first aspect of the embodiments of the present invention provides a method for identifying a caller based on an attention mechanism, including the following steps:

Collect call recordings of multiple tested callers and test callers;

According to the call recording corresponding to the tested caller, establish the caller's voice library;

Use the attention-based neural network to train the voice of the tested caller to obtain the training model;

Store the call recording of the test caller and obtain the recording file;

The recording file is used to train the model to identify whether the tested caller is the target caller.

Optionally, the step is to collect call recordings of multiple tested callers and call recordings of test callers, including:

The test party uses the built-in recording function of the smartphone during the call to record the voice of the person under test during the call; when using the built-in call recording function of the system during the call, it is necessary to specify the mobile phone device model, the recording file storage format, and the call recording function. External environment characteristics; and save call recordings as lossless files in Wave format.

Optionally, the step is to establish a caller's voice library according to the call recording corresponding to the tested caller, including:

Obtain the correspondence between the identity of the tested caller and the tested voice;

The caller's voice database is established according to the corresponding relationship. The caller's voice database includes the audio data of the call of the identified party, the identity information of the tested caller, the environmental characteristics of the tested caller, the information of the recording equipment of the tester, the environmental characteristics of the tester, the call duration, Call time, call volume.

Optionally, the step adopts an attention-based neural network to train the voice of the tested caller to obtain a training model, including:

The recording file is denoised with Wiener filter to obtain the preprocessed recording file;

The preprocessed recording file is trained by using an attention mechanism-based temporal recurrent neural network to obtain a training model.

Optionally, the step adopts the attention mechanism-based temporal recurrent neural network training to preprocess the recording files to obtain a training model, including:

Extracting speech features from the preprocessed recording file through the input layer of the time recurrent neural network, and obtaining the Mel cepstral coefficient feature vector of the speech in the preprocessed recording file;

The feature vector of Mel cepstral coefficients is sent to the fully connected layer, and the fully connected layer (which can be regarded as an autoencoder) performs feature extraction on the feature vector of Mel cepstral coefficients to obtain the second feature vector of the speech in the preprocessed recording file ;

sending the second feature vector to the attention-based time recurrent neural network layer, the attention-based time recurrent neural network layer includes multiple LSTM layers, and processing the second feature vector through multiple LSTM layers to obtain processing data;

The processed data is sent to the normalized exponential function layer, and the normalized exponential function layer converts the processed data corresponding to the person's name to obtain the person's name corresponding to the processed data.

Optionally, the step uses the recording file to identify whether the tested caller is the target caller by using a training model, including:

Determine whether there is a voice to be tested in the caller's voice library, and if so, identify the existing testee in the caller's voice library; otherwise, if the audio file to be recognized belongs to a new tester, it will be identified as the closest one. Existing tested person, that is, the existing classification with the largest confidence value.

In order to achieve the above object, the present invention also provides the following scheme:

Caller identification device based on attention mechanism, including:

The collection module is used to collect the call recordings of multiple tested callers and the call recordings of the test callers;

The voice library establishment module is used to establish the caller's voice library according to the call recording corresponding to the tested caller;

The training module is used to train the voice of the tested caller by using the attention-based neural network to obtain the training model;

The file storage module is used to store the call recording of the test caller and obtain the recording file;

The test module is used to identify whether the tested caller is the target caller by using the training model on the recording file.

Optionally, the collection module specifically includes:

The test party unit is used for the test party to use the built-in recording function of the smartphone during the call to record the voice of the tested caller during the call;

The recording unit is used to use the built-in call recording function of the system during the call. It is necessary to specify the mobile phone device model, the recording file storage format, and the external environment characteristics during the call; and save the call recording as a lossless file in Wave format.

Optionally, the voice library establishment module specifically includes:

A corresponding relationship obtaining unit, used for obtaining the corresponding relationship between the identity of the tested caller and the tested voice;

The voice library establishment unit is used to establish a caller voice library according to the corresponding relationship. The caller voice library includes the audio data of the call of the identified party, the identity information of the tested caller, the environmental characteristics of the tested caller, the information of the test party's recording equipment, and the test. Party environment characteristics, call duration, call time, call volume.

Optionally, the training module specifically includes:

The preprocessing unit is used to denoise the recording file by using the Wiener filter to obtain the preprocessing recording file;

The training model establishment unit is used to train the preprocessing recording file by using the attention mechanism-based time recurrent neural network to obtain the training model.

According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

The invention discloses a speaker identification method and device based on an attention mechanism. By using an attention-based neural network to train the voice of the tested caller, a training model is obtained, and the training model is used to identify the tested caller. In the case of audio, the existing communication voice library is used to determine the speech in the communication process, so that the user can match the caller with the actual caller through the number display, so as to make a reliability judgment, which can effectively prevent fraud by counterfeiting the caller's voice. and other behaviors to indirectly protect the security of user communications.

Description of drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings required in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some of the present invention. In the embodiments, for those of ordinary skill in the art, other drawings can also be obtained according to these drawings without creative labor.

1 is a schematic flowchart of an embodiment of a method for identifying a caller based on an attention mechanism according to Embodiment 1 of the present invention;

2 is a flowchart of a method for speaker identification based on an attention mechanism provided in Embodiment 2 of the present invention;

3 is a schematic structural diagram of a caller identification device based on an attention mechanism according to Embodiment 3 of the present invention;

FIG. 4 is a structural diagram of a neural network provided by the present invention.

Detailed ways

In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some, but not all, embodiments of the present invention. Based on the described embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the protection scope of the present invention.

Example 1

Embodiment 1 of the present invention provides an embodiment of a method for speaker recognition based on an attention mechanism. As shown in FIG. 1 , the method includes the following steps:

S101: Collect call recordings of multiple tested callers and call recordings of test callers;

S102: Establish a caller's voice library according to the call recording corresponding to the tested caller;

S103: Use an attention-based neural network to train the voice of the tested caller to obtain a training model;

S104: Store the call recording of the test caller, and obtain the recording file;

S105: Use the recording file to identify whether the tested caller is the target caller by using the training model.

During the answering process, the test party records the call, and then corresponds the call recording to the tested caller, and builds the tested caller's voice library;

Use the attention-based LSTM neural network to learn the characteristics of the caller's voice to generate a caller recognition model;

When the caller is identified, the test party records the call process and saves the recording file as a wave audio file;

The saved audio file is passed through the Wiener filter, and the Mel cepstral coefficients are extracted and input into the trained caller identification model to identify the tested caller.

The use of attention-based temporal recurrent neural network can be directly implemented using the open source tool tensorflow.

The network parameters used in the example of the present invention are: rnn_size is any real number, such as 64,

attn_length is any real number, such as 64. It should be emphasized that the core of the present invention is a caller identification method based on an attention mechanism, and operations for the network such as modifying network parameters are all included in the present invention.

The caller identification based on the attention mechanism provided by the embodiment of the present invention can use the existing communication voice library to determine the speech in the communication process under the condition of only audio, so that the user can match the caller and the actual caller through the number display, Thereby, it can judge the reliability and indirectly protect the user's communication security. By using the training model to identify whether the tested caller is the target caller, the behavior of making a call by imitating the caller's identity by voice has a certain recognition ability, which improves the user's call process. security of call information.

Example 2

Embodiment 2 of the present invention provides a preferred embodiment of an attention mechanism-based speaker identification method. Referring to Figure 2, in this embodiment, the method includes the steps:

S201: Collect the call recording of the user, correspond the call recording with the tested caller, and construct a caller's voice library.

The test party uses the built-in recording device of the smartphone during the call or uses the headset with recording function to record the voice of the tested caller during the call.

Android phones can use the built-in call recording function during a call. It is necessary to clarify the phone device model, the recording file storage format, and the external environment characteristics (such as quiet and noisy) during the call; while the Apple phone does not provide a call due to the privacy setting system. For the recording function, you can record through the headphones with the recording function. You need to clarify the brand of the headphones, the model of the headphones, the recording file storage format, and the external environment characteristics (such as quiet, noisy).

When the test party is the calling party, you can turn on the smartphone's built-in recording device or the headset with recording function after the called party answers the call; when the test party is the called party, you can turn on the smartphone's built-in recording device or include Headphones with recording function.

The voice library is the association library between the tested caller's identity and the tested voice, and the purpose is to provide data for subsequent model training.

The voice library contains the audio data of the call of the recognized party; the identity information of the recognized party (such as phone number, name, location); the environmental characteristics of the called party (such as: indoor, street, store, etc.); the test party's recording equipment information (such as : sampling frequency, noise reduction characteristics, audio storage format, etc.); test party environmental characteristics (for example: indoor, street, store, etc.); call duration; call time; call volume.

S202: Use an attention-based neural network to train the communication person's speech to generate a training model.

Specifically, the speech files in the speech library are preprocessed, and the Wiener filter is used to perform simple noise reduction of speech, so as to avoid the influence of noise on the whole experimental process.

The neural network used is based on the temporal recurrent neural network of the attention mechanism, and the network structure is shown in Figure 4.

S203: During the test, the test party records the call process, and saves the recording file.

Specifically, the audio file format required to be saved is lossless, such as WAVE, FLAC, APE, ALAC, WavPack, etc.

Among them, WAVE usually uses three parameters to represent the sound, the number of quantization bits, the sampling frequency and the sampling point amplitude. The quantization bits are divided into 8 bits, 16 bits, and 24 bits. The channels are divided into mono and stereo. The mono amplitude data is n*1 matrix point, and the stereo is n*2 matrix point. The sampling frequency is general. There are 11025Hz (11kHz), 22050Hz (22kHz) and 44100Hz (44kHz) three, the sound quality is excellent, but the file size is large. And record the encoding method of the recording file.

S204: Input the recording file into the trained model to identify the tested caller.

Specifically, the saved audio file should be preprocessed first, and the voice signal will be denoised by Wiener filtering, and then cut into voice fragments of the same length, such as 10 seconds, and the Mel frequency cepstrum is extracted from each voice fragment. After counting the features, input the trained model for classification. The neural network model is shown in Figure 4.

The neural network structure includes:

S301 is a feature input layer, which performs feature engineering on the voice in the communication voice library, and extracts the required features, where the Mel cepstral coefficients of the voice in the voice library are extracted as a feature vector;

S302 is a fully connected layer;

S303 is an attention-based temporal recurrent neural network layer;

The result of the normalized exponential function layer calculation in S304;

S305 is an output layer, which converts the code and the person's name, and is used to output the result calculated by the normalized exponential function layer in S304.

This embodiment can only identify the existing tested person, that is, the existing tested person in the voice library of the caller. If the audio file to be recognized belongs to a new test person, it will be recognized as the closest existing tested person. people.

Embodiment 2 provided by the present invention uses the attention-based neural network to train the voice of the tested caller, obtains a training model, uses the training model to identify the tested caller, confirms the consistency of the owner corresponding to the dialed number, and avoids It avoids the security risk of the caller's identity being forged.

Example 3

Embodiment 3 of the present invention further provides a device for identifying a caller based on an attention mechanism, as shown in FIG. 3 .

The collection module 10 is used for constructing the tested caller's voice library corresponding to the tested caller's identity and the audio file. The collection module 10 can be further subdivided into a telephone recording module 11 and a database processing module 12 .

The telephone recording module 11 is used to record the call process. The specific implementation method is that the test party has a planned or unplanned call with the tested caller, records the call process, and saves the recording file as an audio in a Wave format. document.

The test party uses the built-in recording device of the smartphone during the call or uses the headset with recording function to record the voice of the tested caller during the call.

Android phones can use the built-in call recording function during a call. It is necessary to clarify the phone device model, the recording file storage format, and the external environment characteristics (such as quiet and noisy) during the call; while the Apple phone does not provide a call due to the privacy setting system. For the recording function, you can record through the headphones with the recording function. You need to clarify the brand of the headphones, the model of the headphones, the recording file storage format, and the external environment characteristics (such as quiet, noisy).

When the test party is the calling party, you can turn on the smartphone's built-in recording device or the headset with recording function after the called party answers the call; when the test party is the called party, you can turn on the smartphone's built-in recording device or include Headphones with recording function.

The database processing module 12 is used for associating the audio file with the tested person's identity.

The test party stores the collected audio files, the identity of the caller under test, and the related configuration information of the audio files into the database. The configuration information contains the identity information of the identified party (for example: phone number, name, location); the environmental characteristics of the called party (for example: indoor, street, store, etc.); the test party's recording equipment information (for example: sampling frequency, noise reduction characteristics) , audio storage format, etc.); test party environmental characteristics (for example: indoor, street, store, etc.); call duration; call duration; call volume information.

The training module 20 is used to train the audio files in the tested person's voice database. The network structure model used is shown in Figure 4.

Specifically, the audio files are denoised and features extracted before this. The noise reduction processing uses Wiener filtering, and the features use the Mel cepstral coefficients of the audio file. Here you can use the audio processing module python_speech_features provided by Python.

The testing module 30 is used to identify the caller to which the new audio file belongs.

During the testing process, the testing party still records the conversation process with the person being tested through the above recording method. The recording file is preprocessed, that is, the noise reduction processing and feature extraction mentioned in the training module 20 are transferred to the trained model, and the output result of the model is received.

The testing module 30 can only identify the existing tested persons in the training module 20, that is, the existing tested persons in the voice database of the caller. If the audio file to be identified in the testing module 30 belongs to a new test person, it will be identified as the most similar existing tested person, and the closest existing tested person is the existing classified tested person with the largest confidence value.

Compared with the prior art, the present invention has the following technical effects:

The invention discloses a method and device for identifying a speaker based on an attention mechanism. By using an attention-based neural network to train the voice of the tested caller, a training model is obtained, and the training model is used to identify the tested caller. The consistency of the owner corresponding to the dialed number is confirmed, the hidden danger of communication security caused by the counterfeiting of the caller's identity is avoided, and the security of the information during the call is further improved.

The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same and similar parts between the various embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant part can be referred to the description of the method.

The principles and implementations of the invention are described herein by using specific examples. The descriptions of the above embodiments are only used to help understand the method and the core idea of the present invention, and the described embodiments are only a part of the embodiments of the present invention. , rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work shall fall within the protection scope of the present invention.

Claims (7)

1. The method for identifying the talker based on the attention mechanism is characterized by comprising the following steps of:

collecting the call records of a plurality of tested callers and the call records of the tested callers;

establishing a speaker voice library according to the call record corresponding to the tested speaker;

training a neural network based on an attention mechanism by adopting the call record of the tested caller to obtain a training model;

storing the call records of the test caller and the tested caller to obtain a record file;

adopting the training model to identify whether the tested speaker is a target speaker according to the recording file,

wherein, adopt the conversation recording of the person who is tested the neural network based on the attention mechanism trains, obtains the training model, includes:

denoising the call recording of the tested caller by adopting a wiener filter to obtain a preprocessed recording file;

extracting voice features from the preprocessed sound recording file through an input layer of a time recursive neural network to obtain a Mel cepstrum coefficient feature vector of voice in the preprocessed sound recording file;

sending the Mel cepstrum coefficient feature vector to a full connection layer, and performing feature extraction on the Mel cepstrum coefficient feature vector by the full connection layer to obtain a second feature vector of the voice in the preprocessed audio file;

sending the second feature vector to an attention-based time-recursive neural network layer, wherein the attention-based time-recursive neural network layer comprises a plurality of LSTM layers, and the second feature vector is processed through the plurality of LSTM layers to obtain processed data;

and sending the processing data to a normalization index function layer, wherein the normalization index function layer correspondingly converts the processing data and the name of the person to obtain the name of the person corresponding to the processing data.

2. The method of claim 1, wherein collecting call records of a plurality of tested callers and a test caller comprises:

the testing party records the voice of the tested caller in the call process by using the built-in recording function of the smart phone in the call process; the call recording function of the system is used in the call process, and the model of the mobile phone equipment, the storage format of the recording file and the external environment characteristics during the call need to be determined; and saving the call record in a lossless file form in Wave format.

3. The method for identifying a caller according to claim 1, wherein the establishing a caller voice library according to the call record corresponding to the tested caller comprises:

acquiring the corresponding relation between the identity of the tested caller and the tested voice;

and establishing a speaker voice library according to the corresponding relation, wherein the speaker voice library comprises the voice frequency data of the identified party, the identity information of the tested party, the environmental characteristics of the tested party, the recording equipment information of the tested party, the environmental characteristics of the tested party, the call duration, the call time and the call volume.

4. The method for identifying a caller based on an attention mechanism according to claim 1, wherein identifying whether the tested caller is a target caller by using the training model according to the recording file comprises:

judging whether the voice to be tested exists in the speaker voice library, and if so, identifying the tested person existing in the speaker voice library; otherwise, if the audio file to be identified belongs to the new tester, the existing tester is identified as the closest.

5. A speaker recognition device based on an attention mechanism, comprising:

the collecting module is used for collecting the call records of a plurality of tested callers and the call records of the tested callers;

the voice database establishing module is used for establishing a speaker voice database according to the call record corresponding to the tested speaker;

the training module is used for training the neural network based on the attention mechanism by adopting the call record of the tested caller to obtain a training model;

the file storage module is used for storing the call records of the test caller and the tested caller to obtain a record file;

the test module is used for identifying whether the tested speaker is a target speaker or not by adopting the training model according to the recording file,

wherein the training module is configured to:

denoising the call recording of the tested caller by adopting a wiener filter to obtain a preprocessed recording file;

extracting voice features from the preprocessed sound recording file through an input layer of a time recursive neural network to obtain a Mel cepstrum coefficient feature vector of voice in the preprocessed sound recording file;

sending the Mel cepstrum coefficient feature vector to a full connection layer, and performing feature extraction on the Mel cepstrum coefficient feature vector by the full connection layer to obtain a second feature vector of the voice in the preprocessed audio file;

sending the second feature vector to an attention-based time-recursive neural network layer, wherein the attention-based time-recursive neural network layer comprises a plurality of LSTM layers, and the second feature vector is processed through the plurality of LSTM layers to obtain processed data;

and sending the processing data to a normalization index function layer, wherein the normalization index function layer correspondingly converts the processing data and the name of the person to obtain the name of the person corresponding to the processing data.

6. The apparatus for identifying a human speaker based on attention mechanism as claimed in claim 5, wherein the collecting module comprises:

the testing party unit is used for recording the voice of the tested caller in the calling process by using the built-in recording function of the smart phone in the calling process by the testing party;

the recording unit is used for using the call recording function of the system in the call process and determining the model of the mobile phone equipment, the storage format of a recording file and the external environment characteristics during the call; and saving the call record in a lossless file form in Wave format.

7. The apparatus for identifying a caller according to claim 5, wherein the speech library creating module comprises:

the corresponding relation acquisition unit is used for acquiring the corresponding relation between the identity of the tested caller and the tested voice;

and the voice library establishing unit is used for establishing a speaker voice library according to the corresponding relation, wherein the speaker voice library comprises the voice frequency data of the identified party, the identity information of the tested party, the environmental characteristics of the tested party, the recording equipment information of the tested party, the environmental characteristics of the tested party, the call duration, the call time and the call volume.

Images