Methods, apparatus, computer equipment and storage media for speech intent recognition

Abstract

This application relates to the field of artificial intelligence and discloses a method, apparatus, computer device, and storage medium for voice intent recognition. When executing the method, it first determines whether the audio data contains the voices of multiple people. If so, it performs voice separation to obtain individual audio data of different people and extracts the corresponding voiceprint features. Then, it searches for a target voiceprint feature that is identical to the pre-stored voiceprint features of a first ID user among the extracted voiceprint features. Finally, it obtains the target individual audio data corresponding to the target voiceprint feature. This target individual audio data can be considered as the voice spoken by the first ID user. Finally, it performs text conversion based on this target individual audio data and performs intent recognition. This allows it to obtain the true intent of the first ID user after excluding interference from other people's audio data, improving the accuracy of intent recognition when the intelligent outbound call robot encounters multiple voices.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence, and in particular to a method, apparatus, computer device, and storage medium for speech intent recognition. Background Technology

[0002] Intelligent outbound calling robots are an important communication tool for businesses to provide effective services to customers. For intelligent outbound calling robots, accurately recognizing the intent behind a customer's voice is crucial.

[0003] When existing intelligent outbound call robots identify customer intent, they first need to convert the customer's voice into text using a speech-to-text model, and then the intent recognition model identifies the text corresponding to the customer's voice, thereby identifying the customer's intent.

[0004] The inventors discovered that when using speech-to-text (STO) systems, customers are often in crowded public places. The presence of other people's voices in the background causes the converted text to be mixed with other voices, leading to errors in the intent model's recognition of customer intent. For example, if a customer says, "Check my balance," but someone else is saying, "I want to watch a movie," the received speech, after being superimposed from both voices, becomes "Check my balance and movie ticket," making intent recognition impossible. Therefore, solving the problem of low accuracy in intent recognition models when intelligent outbound call robots encounter multiple voices is a pressing issue. Summary of the Invention

[0005] The main purpose of this application is to provide a method, apparatus, computer device and storage medium for voice intent recognition, aiming to solve the technical problem of low accuracy of intent recognition model when intelligent outbound call robots encounter multiple voices.

[0006] To achieve the aforementioned objectives, this application proposes a method for voice intent recognition, the method comprising:

[0007] The system acquires the audio data to be identified sent by the first ID user and searches the preset voiceprint library for the pre-stored voiceprint features corresponding to the first ID user.

[0008] Determine whether the audio data contains the voices of multiple people;

[0009] If the audio data contains the voices of multiple people, then the audio data is separated into individual audio data of multiple different people by using a pre-trained voice separation model;

[0010] The voiceprint features of multiple different people are obtained by analyzing the individual audio data of multiple different people;

[0011] Extract the target voiceprint feature that is the same as the pre-stored voiceprint feature from the voiceprint features of multiple different people;

[0012] From the individual audio data of multiple different people, extract the individual audio data corresponding to the target voiceprint feature as the target individual audio data;

[0013] The target audio data is converted into text data, and the text data is input into a pre-trained intent recognition model to obtain the intent of the audio data to be recognized.

[0014] Furthermore, determining whether the audio data contains the voices of multiple people includes:

[0015] Extract a first audio data segment of a first time length from the audio data;

[0016] The first audio data is divided into second audio data and third audio data;

[0017] The voiceprint features of the first audio data, the second audio data, and the third audio data are extracted respectively to obtain the first voiceprint feature, the second voiceprint feature, and the third voiceprint feature;

[0018] Determine whether the first voiceprint feature, the second voiceprint feature, and the third voiceprint feature are the same;

[0019] If they are not the same, it is determined that the audio data contains the voices of multiple people.

[0020] Further, before extracting the target voiceprint feature that is identical to the pre-stored voiceprint feature corresponding to the first ID user from the voiceprint features of multiple different people, the process includes:

[0021] Determine whether the pre-stored voiceprint features of the first ID user are stored;

[0022] If so, then execute the step of "extracting the target voiceprint feature that is the same as the pre-stored voiceprint feature corresponding to the first ID user from the voiceprint features of multiple different people";

[0023] If not, then obtain the first historical voice data of the first ID user, establish the pre-stored voiceprint features of the first ID user based on the first historical voiceprint data, and then perform the step of "extracting the target voiceprint features that are the same as the pre-stored voiceprint features corresponding to the first ID user from the voiceprint features of multiple different people".

[0024] Furthermore, before obtaining the audio data to be identified sent by the first ID user, the process includes:

[0025] Obtain the second historical audio data for each historical user;

[0026] Based on the preset voiceprint registration model, the voiceprint features corresponding to each second historical sound data are extracted for database entry.

[0027] Each historical user's ID is mapped one-to-one with its corresponding voiceprint feature and placed into the voiceprint database.

[0028] Furthermore, the voiceprint registration model is an improved model that adds a SpecAug layer before the first layer of the ecapa-tdnn model. The SpecAug layer is used to randomly mask the input fbank vector.

[0029] Furthermore, before obtaining the audio data to be identified sent by the first ID user, the process includes:

[0030] Train a basic voice separation model to obtain a pre-trained voice separation model. The basic voice separation model is an improved model that adds a SpecAug layer before the first layer of the ecapa-tdnn model and replaces the softmax layer of the capa-tdnn model with a focaloss layer.

[0031] Furthermore, before training the basic voice separation model to obtain the pre-trained voice separation model, the process includes:

[0032] Retrieve audio clips of all historical users with IDs, wherein each audio clip has a corresponding ID tag;

[0033] By splicing the various sound segments in pairs, multiple spliced ​​sound segments are obtained;

[0034] The spliced ​​audio segment is divided into segments according to a preset time length to obtain segmented audio samples. Each spliced ​​audio segment is divided into at least two segmented audio samples with individual ID tags, and one segmented audio sample with both ID tags.

[0035] This will yield a set of all the segmented audio samples, forming a sample set for training the basic model of human voice separation.

[0036] This application also provides a device for voice intent recognition, the device comprising:

[0037] The acquisition unit is used to acquire the audio data to be identified sent by the first ID user, and to search for the pre-stored voiceprint features corresponding to the first ID user in the preset voiceprint library.

[0038] A judgment unit is used to determine whether the audio data contains the voices of multiple people;

[0039] The voice separation unit is used to separate the voices of multiple people in the audio data by using a pre-trained voice separation model if the audio data contains the voices of multiple people, so as to obtain individual audio data of multiple different people.

[0040] The parsing unit is used to analyze the voiceprint features of individual audio data from multiple different people to obtain the voiceprint features of multiple different people.

[0041] The first extraction unit is used to extract the target voiceprint feature that is the same as the pre-stored voiceprint feature from the voiceprint features of multiple different people;

[0042] The second extraction unit is used to extract individual audio data corresponding to the target voiceprint features from the individual audio data of multiple different people as target individual audio data;

[0043] An intent recognition unit is used to convert the target individual audio data into text data and input the text data into a pre-trained intent recognition model to obtain the intent of the audio data to be recognized.

[0044] This application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods described above.

[0045] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0046] The method, apparatus, computer device, and storage medium for voice intent recognition disclosed in this application, when executing the method, first determine whether the audio data contains the voices of multiple people. If so, voice separation is performed to obtain individual audio data of different people and extract the corresponding voiceprint features. Then, a target voiceprint feature that is the same as the pre-stored voiceprint feature of the first ID user is searched among the extracted voiceprint features. Finally, the target individual audio data corresponding to the target voiceprint feature is obtained. This target individual audio data can be considered as the voice spoken by the first ID user. Finally, text conversion is performed based on the target individual audio data, and intent recognition is performed. This can obtain the true intent of the first ID user after eliminating the interference of other people's audio data, thereby improving the accuracy of intent recognition when the intelligent outbound call robot encounters multiple voices. Attached Figure Description

[0047] Figure 1 This is a flowchart illustrating a method for voice intent recognition according to an embodiment of this application;

[0048] Figure 2 This is a schematic diagram illustrating the process of determining whether audio data contains the voices of multiple people, according to an embodiment of this application.

[0049] Figure 3 This is a schematic block diagram of a speech intent recognition device according to an embodiment of this application;

[0050] Figure 4 This is a schematic block diagram of the structure of a computer device according to one embodiment of the application.

[0051] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0053] Reference Figure 1 This application provides a method for voice intent recognition, including the following steps S1 to S7:

[0054] S1. Obtain the audio data to be identified sent by the first ID user, and search for the pre-stored voiceprint features corresponding to the first ID user in the preset voiceprint library.

[0055] S2. Determine whether the audio data contains the voices of multiple people;

[0056] S3. If the audio data contains the voices of multiple people, then the audio data is separated into individual audio data of multiple different people by using a pre-trained voice separation model.

[0057] S4. Analyze the voiceprint features of individual audio data from multiple different people to obtain the voiceprint features of multiple different people;

[0058] S5. Extract the target voiceprint feature that is the same as the pre-stored voiceprint feature from the voiceprint features of multiple different people;

[0059] S6. From the individual audio data of multiple different people, extract the individual audio data corresponding to the target voiceprint feature as the target individual audio data;

[0060] S7. Convert the target individual audio data into text data, and input the text data into a pre-trained intent recognition model to obtain the intent of the audio data to be recognized.

[0061] The speech intent recognition method of this application is applied to artificial intelligence fields such as intelligent speech recognition and intelligent voice response. For example, this application is applied to intelligent outbound calling robots, which are devices or applications that connect to audio data input by users and then provide corresponding responses.

[0062] As described in step S1 above, the first ID user is the user who owns the first ID. The first ID user is the user currently inputting voice, and their input voice (audio data to be recognized) generally includes their own voice and sounds from their surrounding environment, such as wind, rain, car engine sounds, and sounds from other people. In other words, the audio data input by the first ID user may contain a large amount of interference, which is the main reason why the recognition of their input voice intent is currently inaccurate. The aforementioned preset voiceprint library is a database that pre-stores a large number of voiceprint features, and each voiceprint feature in the library corresponds to an ID. In this application, in different application scenarios, users generally register an account in advance, and a unique ID is generated during account registration. When the first ID user inputs audio data, the corresponding ID can be determined, and then the corresponding pre-stored voiceprint feature can be found in the voiceprint library.

[0063] As described in step S2 above, "multiple people" refers to two or more people. The method for determining whether audio data contains the voices of multiple people can be as follows: The audio data is segmented into multiple parts, then the voiceprint features of each segment are extracted. The voiceprint features of each segment are then compared to see if they are identical. If they are identical, it is determined that only one person is speaking; otherwise, it is determined that multiple people are speaking. When it is determined that only one person is speaking, the audio data can be considered to be the voice of the first ID user, and text conversion and intent recognition can be performed directly. When it is determined that multiple people are speaking, the audio data needs to be processed to obtain the individual audio data corresponding to the first ID user.

[0064] As described in step S3 above, once it is determined that there are multiple voices in the audio data to be identified, voice separation is performed. Voice separation is the process of separating the voices of different people. The aforementioned voice separation model is a model trained on a large amount of sample data. The sample data consists of combined speech data containing the voices of different people, and is labeled accordingly. Supervised training is generally used during training. In this embodiment, the voice separation model is a model trained based on the ECAPA-TDNN model.

[0065] As described in steps S4-S6 above, after extracting the individual audio data of different people, it is not possible to determine which individual audio data belongs to the first ID user. Therefore, it is necessary to extract the voiceprint features of each individual audio data and compare them with the pre-stored voiceprint features of the first ID user to obtain the target voiceprint features. Then, the individual audio data corresponding to the target voiceprint features is taken as the target individual audio data and recognized as the individual audio data of the first ID user.

[0066] As described in step S7 above, the target individual audio data is converted into text data, and the text data is input into a pre-trained intent recognition model to obtain the intent of the audio data to be recognized.

[0067] The voice intent recognition method of this application embodiment first determines whether the audio data contains the voices of multiple people. If so, it performs voice separation to obtain individual audio data of different people and extracts the corresponding voiceprint features. Then, it searches for target voiceprint features that are the same as the pre-stored voiceprint features of the first ID user among the extracted voiceprint features. Finally, it obtains the target individual audio data corresponding to the target voiceprint features. This target individual audio data can be considered as the voice of the first ID user. Finally, it performs text conversion based on the target individual audio data and performs intent recognition, that is, it obtains the true intent of the first ID user after excluding the interference of other people's audio data, thereby improving the accuracy of intent recognition when the intelligent outbound call robot encounters multiple voices.

[0068] Reference Figure 2 In one embodiment, step S2, which determines whether the audio data contains the voices of multiple people, includes the following steps S21 to S25:

[0069] S21. Extract first audio data of a first time length from the audio data;

[0070] S22. Divide the first audio data into second audio data and third audio data;

[0071] S23. Extract the voiceprint features of the first audio data, the second audio data and the third audio data respectively to obtain the first voiceprint feature, the second voiceprint feature and the third voiceprint feature;

[0072] S24. Determine whether the first voiceprint feature, the second voiceprint feature, and the third voiceprint feature are the same;

[0073] S25. If they are not the same, then the audio data is determined to contain the voices of multiple people.

[0074] As described in steps S21 to S25 above, firstly, a segment of first audio data is extracted from the audio data. Secondly, the first audio data is split into second and third audio data. Then, the voiceprint features of these three audio data segments are extracted and compared. If the voiceprint features of the three audio data segments are all the same, it indicates that the voice in the first audio data is entirely from the same person. If they are different, it indicates that the voices in the first audio data are from multiple people, which in turn indicates that the audio data to be identified sent by the first ID user contains the voices of multiple people. In a specific embodiment, the first time length is 3 seconds, while the second and third audio data are 1.5 seconds each. The method of determining whether the audio data contains the voices of multiple people in this embodiment only needs to analyze a small segment of audio data to determine whether there are multiple voices, which is accurate and saves computing resources. In a specific embodiment, if the intelligent outbound call robot needs to identify audio data in real time, it can make a judgment every 3 seconds to prevent the input audio data from being interspersed with the voices of other people.

[0075] In one embodiment, before step S4 of extracting the target voiceprint feature that is identical to the pre-stored voiceprint feature corresponding to the first ID user from the voiceprint features of multiple different people, the following steps are included:

[0076] Determine whether the pre-stored voiceprint features of the first ID user are stored;

[0077] If so, then execute the step of "extracting the target voiceprint feature that is the same as the pre-stored voiceprint feature corresponding to the first ID user from the voiceprint features of multiple different people";

[0078] If not, then obtain the first historical voice data of the first ID user, establish the pre-stored voiceprint features of the first ID user based on the first historical voiceprint data, and then perform the step of "extracting the target voiceprint features that are the same as the pre-stored voiceprint features corresponding to the first ID user from the voiceprint features of multiple different people".

[0079] As mentioned above, the first ID user may be an unregistered user, but there are historical interaction records. If no pre-stored voiceprint feature corresponding to the first ID user is found in the voiceprint database, a new "ID-voiceprint feature" data needs to be created in the voiceprint database. The method for creating this data is to retrieve the user's historical interaction records (voice interaction records), and then extract a voice data point from the historical interaction records that only contains the voice of one person, i.e., the aforementioned first historical voice data. Then, based on the first historical voiceprint data, the pre-stored voiceprint feature of the first ID user is created. After creation, step S4 is executed. To determine which voice data in the historical interaction records are from one person, the methods described in steps S21 to S25 above can be used.

[0080] In one embodiment, prior to obtaining the audio data to be identified sent by the first ID user, the process includes:

[0081] Obtain the second historical audio data for each historical user;

[0082] Based on the preset voiceprint registration model, the voiceprint features corresponding to each second historical sound data are extracted for database entry.

[0083] Each historical user's ID is mapped one-to-one with its corresponding voiceprint feature and placed into the voiceprint database.

[0084] As described above, this is the process of building a voiceprint database. Voiceprint features are extracted from the second historical voice data of all historical users, and then bound one-to-one with their corresponding IDs to form the voiceprint database. It is important to note that each second historical voice data point represents the voice of a single individual.

[0085] In one embodiment, during the ID creation process, users can actively input audio data of a person's voice into the intelligent outbound calling robot. The robot then directly obtains the voiceprint features corresponding to the ID based on this audio data and stores them in the voiceprint database in a one-to-one mapping. The aforementioned voiceprint registration model is a model for extracting voiceprint features, which can be trained using the ECAPA-TDNN model.

[0086] In one embodiment, the above-mentioned voiceprint registration model is an improved model that adds a SpecAug layer before the first layer of the ecapa-tdnn model. The SpecAug layer is used to perform random masking on the input fbank vector.

[0087] As mentioned above, when training the voiceprint registration model, the sample size is first preprocessed to obtain the fbank vector. Then, the fbank vector is input into the SpecAug layer, which randomly masks the input fbank vector. This random masking of the original generated fbank vector before inputting it into the subsequent networks of the ECAPA-TDNN model effectively augments the data, increasing the model's generalization ability.

[0088] In one embodiment, the process of preprocessing the sample number to obtain the fbank vector includes: pre-emphasis, framing, windowing, performing fast Fourier transform, and Mel filter processing.

[0089] In one embodiment, prior to obtaining the audio data to be identified sent by the first ID user, the process includes:

[0090] Retrieve audio clips of all historical users with IDs, wherein each audio clip has a corresponding ID tag;

[0091] By splicing the various sound segments in pairs, multiple spliced ​​sound segments are obtained;

[0092] The spliced ​​audio segment is divided into segments according to a preset time length to obtain segmented audio samples. Each spliced ​​audio segment is divided into at least two segmented audio samples with individual ID tags, and one segmented audio sample with both ID tags.

[0093] This will yield the complete set of audio samples from all segments, forming a sample set for training the basic human voice separation model.

[0094] Based on the sample set, a basic voice separation model is trained to obtain a pre-trained voice separation model. The basic voice separation model is an improved model that adds a SpecAug layer before the first layer of the ecapa-tdnn model and replaces the softmax layer of the capa-tdnn model with a focaloss layer.

[0095] As mentioned above, before training the voice separation model, a voice separation sample dataset is first constructed. When constructing this dataset, audio segments from historical users are directly selected, as they each have a unique ID. Then, audio segments from two different users are concatenated to form the concatenated audio segment. This concatenated audio segment is then segmented to obtain segmented audio samples. For example, a concatenated audio segment might be formed by concatenating audio segments with IDs no.1 and no.2. The boundary frame at the concatenation point is denoted as ch. During segmentation, the concatenated audio segment is segmented with each frame having a length of 25 milliseconds. This results in multiple segmented audio samples labeled no.1, multiple segmented audio samples labeled no.2, and one segmented audio sample containing both no.1 and no.2 (i.e., the segmented audio sample containing the boundary frame ch). The resulting segmented audio sample set can then be represented as [no.1, no.1, ..., no.1, ch, no.2, ... no.2]. After segmenting multiple spliced ​​audio segments, a sufficiently large sample dataset can be constructed for training the voice separation model. As described above, the sample dataset contains relatively few 'ch' tags, so the original softmax loss function is improved to focaloss, which is more effective against imbalanced samples. Similarly, a SpecAug layer is added before the first layer of the ECAPA-TDNN model because randomly masking the fbank vector generated from the original data effectively augments the data and increases the model's generalization ability.

[0096] Reference Figure 3 This application also provides a device for voice intent recognition, the device comprising:

[0097] The acquisition unit 10 is used to acquire the audio data to be identified sent by the first ID user, and to search for the pre-stored voiceprint features corresponding to the first ID user in the preset voiceprint library.

[0098] The judgment unit 20 is used to determine whether the audio data contains the voices of multiple people;

[0099] The voice separation unit 30 is used to separate the voices of multiple people in the audio data by using a pre-trained voice separation model to obtain individual audio data of multiple different people if the audio data contains the voices of multiple people.

[0100] The parsing unit 40 is used to analyze the voiceprint features of individual audio data from multiple different people to obtain the voiceprint features of multiple different people.

[0101] The first extraction unit 50 is used to extract the target voiceprint feature that is the same as the pre-stored voiceprint feature from the voiceprint features of multiple different people;

[0102] The second extraction unit 60 is used to extract individual audio data corresponding to the target voiceprint feature from the individual audio data of multiple different people as target individual audio data;

[0103] The intent recognition unit 70 is used to convert the target individual audio data into text data and input the text data into a pre-trained intent recognition model to obtain the intent of the audio data to be recognized.

[0104] In one embodiment, the aforementioned determining unit is specifically configured to: extract first audio data of a first time length from the audio data; divide the first audio data into second audio data and third audio data; extract the voiceprint features of the first audio data, the second audio data, and the third audio data respectively to obtain first voiceprint features, second voiceprint features, and third voiceprint features; determine whether the first voiceprint features, the second voiceprint features, and the third voiceprint features are the same; if they are not the same, determine that the audio data contains the voices of multiple people.

[0105] In one embodiment, the above-mentioned speech intent recognition device further includes:

[0106] The first judgment unit is used to determine whether the pre-stored voiceprint features of the first ID user are stored.

[0107] The execution unit is configured to, if applicable, execute the step of "extracting a target voiceprint feature that is identical to the pre-stored voiceprint feature corresponding to the first ID user from the voiceprint features of multiple different people"; otherwise, acquire the first historical voice data of the first ID user, establish the pre-stored voiceprint feature of the first ID user based on the first historical voiceprint data, and then execute the step of "extracting a target voiceprint feature that is identical to the pre-stored voiceprint feature corresponding to the first ID user from the voiceprint features of multiple different people".

[0108] In one embodiment, the above-mentioned speech intent recognition device further includes:

[0109] The first acquisition unit is used to acquire the second historical audio data of each historical user;

[0110] The third extraction unit is used to extract the voiceprint features corresponding to each second historical sound data based on the preset voiceprint registration model.

[0111] The input unit is used to map the ID of each historical user to the corresponding voiceprint feature and input it into the voiceprint database.

[0112] In one embodiment, the above-mentioned voiceprint registration model is an improved model that adds a SpecAug layer before the first layer of the ecapa-tdnn model. The SpecAug layer is used to perform random masking on the input fbank vector.

[0113] In one embodiment, the above-mentioned speech intent recognition device further includes:

[0114] A voice separation training unit is used to train a basic voice separation model to obtain a pre-trained voice separation model. The basic voice separation model is an improved model that adds a SpecAug layer before the first layer of the ecapa-tdnn model and replaces the softmax layer of the capa-tdnn model with a focaloss layer.

[0115] In one embodiment, the above-mentioned speech intent recognition device further includes:

[0116] The dataset construction unit is used to acquire audio segments of all historical users with IDs, wherein each audio segment has a corresponding ID tag; concatenate each audio segment in pairs to obtain multiple concatenated audio segments; segment the concatenated audio segments according to a preset time length to obtain segmented audio samples, wherein each concatenated audio segment is divided into at least two segmented audio samples with individual ID tags, and one segmented audio sample with both ID tags; the collection of all obtained segmented audio samples forms a sample set for training the basic human voice separation model.

[0117] Reference Figure 4 This application also provides a computer device, which can be the aforementioned management server or the server corresponding to the management node, and its internal structure can be as follows: Figure 4 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores data such as voiceprint features. The network interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements a method for voice intent recognition.

[0118] The processor performing the above-mentioned method for voice intent recognition includes:

[0119] The system acquires the audio data to be identified sent by the first ID user and searches the preset voiceprint library for the pre-stored voiceprint features corresponding to the first ID user.

[0120] Determine whether the audio data contains the voices of multiple people;

[0121] If the audio data contains the voices of multiple people, then the audio data is separated into individual audio data of multiple different people by using a pre-trained voice separation model;

[0122] The voiceprint features of multiple different people are obtained by analyzing the individual audio data of multiple different people;

[0123] Extract the target voiceprint feature that is the same as the pre-stored voiceprint feature from the voiceprint features of multiple different people;

[0124] From the individual audio data of multiple different people, extract the individual audio data corresponding to the target voiceprint feature as the target individual audio data;

[0125] The target audio data is converted into text data, and the text data is input into a pre-trained intent recognition model to obtain the intent of the audio data to be recognized.

[0126] In one embodiment, determining whether the audio data contains the voices of multiple people includes:

[0127] Extract a first audio data segment of a first time length from the audio data;

[0128] The first audio data is divided into second audio data and third audio data;

[0129] The voiceprint features of the first audio data, the second audio data, and the third audio data are extracted respectively to obtain the first voiceprint feature, the second voiceprint feature, and the third voiceprint feature;

[0130] Determine whether the first voiceprint feature, the second voiceprint feature, and the third voiceprint feature are the same;

[0131] If they are not the same, it is determined that the audio data contains the voices of multiple people.

[0132] In one embodiment, before extracting the target voiceprint feature that is identical to the pre-stored voiceprint feature corresponding to the first ID user from the voiceprint features of multiple different people, the process includes:

[0133] Determine whether the pre-stored voiceprint features of the first ID user are stored;

[0134] If so, then execute the step of "extracting the target voiceprint feature that is the same as the pre-stored voiceprint feature corresponding to the first ID user from the voiceprint features of multiple different people";

[0135] If not, then obtain the first historical voice data of the first ID user, establish the pre-stored voiceprint features of the first ID user based on the first historical voiceprint data, and then perform the step of "extracting the target voiceprint features that are the same as the pre-stored voiceprint features corresponding to the first ID user from the voiceprint features of multiple different people".

[0136] In one embodiment, prior to obtaining the audio data to be identified sent by the first ID user, the process includes:

[0137] Obtain the second historical audio data for each historical user;

[0138] Based on the preset voiceprint registration model, the voiceprint features corresponding to each second historical sound data are extracted for database entry.

[0139] Each historical user's ID is mapped one-to-one with its corresponding voiceprint feature and placed into the voiceprint database.

[0140] In one embodiment, the voiceprint registration model is an improved model that adds a SpecAug layer before the first layer of the ecapa-tdnn model. The SpecAug layer is used to perform a random mask on the input fbank vector.

[0141] In one embodiment, prior to obtaining the audio data to be identified sent by the first ID user, the process includes:

[0142] Train a basic voice separation model to obtain a pre-trained voice separation model. The basic voice separation model is an improved model that adds a SpecAug layer before the first layer of the ecapa-tdnn model and replaces the softmax layer of the capa-tdnn model with a focaloss layer.

[0143] In one embodiment, before training the basic voice separation model to obtain the pre-trained voice separation model, the process includes:

[0144] Retrieve audio clips of all historical users with IDs, wherein each audio clip has a corresponding ID tag;

[0145] By splicing the various sound segments in pairs, multiple spliced ​​sound segments are obtained;

[0146] The spliced ​​audio segment is divided into segments according to a preset time length to obtain segmented audio samples. Each spliced ​​audio segment is divided into at least two segmented audio samples with individual ID tags, and one segmented audio sample with both ID tags.

[0147] This will yield a set of all the segmented audio samples, forming a sample set for training the basic model of human voice separation.

[0148] Those skilled in the art will understand that Figure 4 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer equipment on which the present application is applied.

[0149] In the computer device of this application embodiment, when the processor executes the above method, a verification rule interface is set on the execution engine of the front-end system. This allows the user-created first rule to be directly sent to the corresponding back-end testing terminal for verification, eliminating the need for the user to copy the rule via a USB flash drive or other storage device and then find the developer for rule verification, or to send an email to the developer, thus improving user efficiency. Similarly, developers are not required to connect to the aforementioned USB flash drive or other storage device to read the user-created rule, saving their workload.

[0150] One embodiment of this application also provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements a method for voice intent recognition, including: acquiring audio data to be recognized sent by a first ID user, and searching for pre-stored voiceprint features corresponding to the first ID user in a preset voiceprint database; determining whether the audio data contains the pronunciation of multiple people; if the audio data contains the pronunciation of multiple people, performing voice separation on the audio data using a pre-trained voice separation model to obtain individual audio data of multiple different people; parsing the voiceprint features of the individual audio data of multiple different people to obtain voiceprint features of multiple different people; extracting a target voiceprint feature that is the same as the pre-stored voiceprint feature from the voiceprint features of multiple different people; extracting individual audio data corresponding to the target voiceprint feature from the individual audio data of multiple different people as target individual audio data; converting the target individual audio data into text data, and inputting the text data into a pre-trained intent recognition model to obtain the intent of the audio data to be recognized.

[0151] When the processor executes the above method, it first determines whether the audio data contains the voices of multiple people. If so, it performs voice separation to obtain individual audio data of different people and extracts the corresponding voiceprint features. Then, it searches for target voiceprint features that are the same as the pre-stored voiceprint features of the first ID user among the extracted voiceprint features. Finally, it obtains the target individual audio data corresponding to the target voiceprint features. This target individual audio data can be considered as the voice of the first ID user. Finally, it performs text conversion and intent recognition based on the target individual audio data. This allows it to obtain the true intent of the first ID user after eliminating the interference of other people's audio data, thus improving the accuracy of intent recognition when the intelligent outbound call robot encounters multiple voices.

[0152] In one embodiment, determining whether the audio data contains the voices of multiple people includes: extracting a first audio data of a first time length from the audio data; dividing the first audio data into a second audio data and a third audio data; extracting the voiceprint features of the first audio data, the second audio data, and the third audio data respectively to obtain a first voiceprint feature, a second voiceprint feature, and a third voiceprint feature; determining whether the first voiceprint feature, the second voiceprint feature, and the third voiceprint feature are the same; if they are not the same, then determining that the audio data contains the voices of multiple people.

[0153] In one embodiment, before extracting the target voiceprint feature that is the same as the pre-stored voiceprint feature corresponding to the first ID user from the voiceprint features of multiple different people, the method includes: determining whether the pre-stored voiceprint feature of the first ID user is stored; if so, performing the step of "extracting the target voiceprint feature that is the same as the pre-stored voiceprint feature corresponding to the first ID user from the voiceprint features of multiple different people"; if not, obtaining the first historical voice data of the first ID user, establishing the pre-stored voiceprint feature of the first ID user based on the first historical voiceprint data, and then performing the step of "extracting the target voiceprint feature that is the same as the pre-stored voiceprint feature corresponding to the first ID user from the voiceprint features of multiple different people".

[0154] In one embodiment, before obtaining the audio data to be identified sent by the first ID user, the process includes: obtaining the second historical sound data of each historical user; extracting the corresponding voiceprint features for each second historical sound data based on a preset voiceprint registration model; and mapping each historical user's ID to the corresponding voiceprint features and placing them into the voiceprint library.

[0155] In one embodiment, the voiceprint registration model is an improved model that adds a SpecAug layer before the first layer of the ecapa-tdnn model. The SpecAug layer is used to perform a random mask on the input fbank vector.

[0156] In one embodiment, before obtaining the audio data to be identified sent by the first ID user, the process includes: training a basic voice separation model to obtain a pre-trained voice separation model, wherein the basic voice separation model is an improved model that adds a SpecAug layer before the first layer of the ecapa-tdnn model and replaces the softmax layer of the capa-tdnn model with a focaloss layer.

[0157] In one embodiment, before training the basic voice separation model to obtain the pre-trained voice separation model, the process includes: acquiring all audio segments of historical users with IDs, wherein each audio segment has a corresponding ID tag; concatenating each audio segment in pairs to obtain multiple concatenated audio segments; segmenting the concatenated audio segments according to a preset time length to obtain segmented audio samples, wherein each concatenated audio segment is divided into at least two segmented audio samples with individual ID tags, and one segmented audio sample with both ID tags; and obtaining the set of all segmented audio samples to form a sample set for training the basic voice separation model.

[0158] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media provided in this application and in the embodiments may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-speed SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0159] The above description is only a preferred embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural changes made based on the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for speech intent recognition, characterized in that, The method includes: The system acquires the audio data to be identified sent by the first ID user and searches the preset voiceprint library for the pre-stored voiceprint features corresponding to the first ID user. Determine whether the pre-stored voiceprint features of the first ID user are stored; if not, obtain the first historical voice data of the first ID user, and establish the pre-stored voiceprint features of the first ID user based on the first historical voice data. Determine whether the audio data contains the voices of multiple people; If the audio data contains the voices of multiple people, then the audio data is separated by adding a SpecAug layer before the first layer of the ecapa-tdnn model and replacing the softmax layer of the ecapa-tdnn model with a focaloss layer in a pre-trained voice separation model, so as to obtain individual audio data of multiple different people. The voiceprint features of multiple different people are obtained by analyzing the individual audio data of multiple different people; Extract the target voiceprint feature that is the same as the pre-stored voiceprint feature from the voiceprint features of multiple different people; From the individual audio data of multiple different people, extract the individual audio data corresponding to the target voiceprint feature as the target individual audio data; The target audio data is converted into text data, and the text data is input into a pre-trained intent recognition model to obtain the intent of the audio data to be recognized.

2. The method for speech intent recognition according to claim 1, characterized in that, The step of determining whether the audio data contains the voices of multiple people includes: Extract a first audio data segment of a first time length from the audio data; The first audio data is divided into second audio data and third audio data; The voiceprint features of the first audio data, the second audio data, and the third audio data are extracted respectively to obtain the first voiceprint feature, the second voiceprint feature, and the third voiceprint feature; Determine whether the first voiceprint feature, the second voiceprint feature, and the third voiceprint feature are the same; If they are not the same, it is determined that the audio data contains the voices of multiple people.

3. The method for speech intent recognition according to claim 1, characterized in that, Before obtaining the audio data to be identified sent by the first ID user, the process includes: Obtain the second historical audio data for each historical user; Based on the preset voiceprint registration model, the voiceprint features corresponding to each second historical sound data are extracted for database entry. Each historical user's ID is mapped one-to-one with its corresponding voiceprint feature and placed into the voiceprint database.

4. The method for speech intent recognition according to claim 1, characterized in that, Before obtaining the audio data to be identified sent by the first ID user, the process includes: training a basic voice separation model to obtain a pre-trained voice separation model. Before obtaining the pre-trained voice separation model by training the basic voice separation model, the process includes: Retrieve audio clips of all historical users with IDs, wherein each audio clip has a corresponding ID tag; By splicing the various sound segments in pairs, multiple spliced ​​sound segments are obtained; The spliced ​​audio segment is divided into segments according to a preset time length to obtain segmented audio samples. Each spliced ​​audio segment is divided into at least two segmented audio samples with individual ID tags, and one segmented audio sample with both ID tags. This will yield a set of all the segmented audio samples, forming a sample set for training the basic model of human voice separation.

5. A device for speech intent recognition, used to implement the speech intent recognition method as described in any one of claims 1-4, characterized in that, The device includes: The acquisition unit is used to acquire the audio data to be identified sent by the first ID user, and to search for the pre-stored voiceprint features corresponding to the first ID user in the preset voiceprint library. A judgment unit is used to determine whether the audio data contains the voices of multiple people; The voice separation unit is used to separate the voices of multiple people in the audio data by using a pre-trained voice separation model if the audio data contains the voices of multiple people, so as to obtain individual audio data of multiple different people. The parsing unit is used to analyze the voiceprint features of individual audio data from multiple different people to obtain the voiceprint features of multiple different people. The first extraction unit is used to extract the target voiceprint feature that is the same as the pre-stored voiceprint feature from the voiceprint features of multiple different people; The second extraction unit is used to extract individual audio data corresponding to the target voiceprint features from the individual audio data of multiple different people as target individual audio data; An intent recognition unit is used to convert the target individual audio data into text data and input the text data into a pre-trained intent recognition model to obtain the intent of the audio data to be recognized.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 4.

Images