A method and apparatus for training a noise recognition model

Abstract

The application discloses a kind of training method and device of noise identification model, it is related to big data technical field.The specific embodiment of the method includes: according to interactive scene division rule, the interactive scene corresponding to the audio segment obtained is determined;For each audio segment, according to the interactive scene of audio segment, corresponding marking rule is obtained, and audio segment is marked according to marking rule;Each audio segment after marking is extracted, to obtain training feature set;Model training is carried out based on training feature set, generates noise identification model, and noise identification model is used to carry out noise identification to the audio segment generated in voice interaction process.The embodiment is trained by dividing interactive scene, and marking rule is marked for different interactive scene, to carry out noise identification model, can improve the accuracy of marking, while giving consideration to the recognition effect of voice and noise dimension, improve the training effect and recognition accuracy of noise identification model.

Description

Technical Field

[0001] This invention relates to the field of big data technology, and in particular to a training method and apparatus for a noise recognition model. Background Technology

[0002] In intelligent voice customer service robot systems, audio data needs to be preprocessed using a noise recognition model. The evaluation criterion for this noise recognition model is to calculate the accuracy of each dimension, specifically speech and noise. Currently, the training method for noise recognition models involves generating training data based on manual labeling, and then using this training data to train the noise recognition model.

[0003] In the process of realizing this invention, the inventors discovered at least the following problems in the prior art:

[0004] Manual labeling is highly subjective and cannot simultaneously address the recognition of both speech and noise dimensions, resulting in poor training performance and recognition accuracy for noise recognition models. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide a training method and apparatus for a noise recognition model. By dividing the interaction scenarios and labeling according to the labeling rules of different interaction scenarios, training data is generated to train the noise recognition model, which can improve the accuracy of labeling, while taking into account the recognition effect of speech and noise dimensions, thereby improving the training effect and recognition accuracy of the noise recognition model.

[0006] To achieve the above objectives, according to one aspect of the present invention, a method for training a noise recognition model is provided.

[0007] A method for training a noise recognition model includes: determining the interaction scenario corresponding to an acquired audio segment based on interaction scenario segmentation rules; for each audio segment, obtaining a corresponding labeling rule based on the interaction scenario of the audio segment, and labeling the audio segment according to the labeling rule; extracting features from each labeled audio segment to obtain a training feature set; and training a model based on the training feature set to generate a noise recognition model, wherein the noise recognition model is used to identify noise in audio segments generated during voice interaction.

[0008] Optionally, before determining the interaction scenario corresponding to the acquired audio segment according to the interaction scenario segmentation rules, the method further includes: determining that the audio segment includes user voice audio, and that the ratio of the length of the user voice audio to the length of the audio segment is greater than a preset threshold.

[0009] Optionally, the interaction scenario includes a user interrupting a customer service speaking process to obtain a speaking right; in the case of the interrupting scenario, the marking the audio segment according to the marking rule includes: in the case that the audio segment includes background noise audio or background user voice audio, marking the audio segment as noise.

[0010] Optionally, the feature extraction on each marked audio segment includes: for each marked audio segment, extracting a pure voice audio segment in the audio segment, and segmenting the pure voice audio segment according to a preset audio segment length threshold to obtain a plurality of audio segmented segments; and performing feature extraction on each audio segmented segment respectively to generate feature information.

[0011] Optionally, the interaction scenario includes a filtering scenario of effectively identifying user voice; in the case of the filtering scenario, the marking the audio segment according to the marking rule includes: in the case that the audio segment includes background user voice audio, marking the audio segment as voice.

[0012] Optionally, the feature extraction on each marked audio segment includes: for each marked audio segment, extracting a pure voice audio segment in the audio segment, and performing feature extraction on the pure voice audio segment to generate feature information.

[0013] Optionally, the training feature set includes feature information and a marking result, and the feature information has a plurality of dimension features; the model training based on the training feature set to generate a noise recognition model includes: taking the feature information as a training input and taking the marking result as a training target to perform model training to obtain model parameters, and generating an importance ranking of the plurality of dimension features; according to the importance ranking, adjusting the model parameters, and re-performing model training based on the training feature set until an identification accuracy rate of a model meets a preset requirement, and taking the model with the identification accuracy rate meeting the preset requirement as the noise recognition model.

[0014] According to another aspect of the embodiment of the present application, a noise recognition model training device is provided.

[0015] The training device of the noise identification model comprises: an interactive scene determination module configured to determine an interactive scene corresponding to an acquired audio segment according to an interactive scene division rule; a marking module configured to, for each audio segment, acquire a corresponding marking rule according to the interactive scene of the audio segment, and mark the audio segment according to the marking rule; a training feature set generation module configured to extract features from each marked audio segment to obtain a training feature set; and a model training module configured to train a model based on the training feature set to generate a noise identification model, which is used for noise identification of an audio segment generated in a voice interaction process.

[0016] Optionally, the device further comprises an audio segment determination module configured to determine that the audio segment comprises user voice audio, and that a ratio of a length of the user voice audio to a length of the audio segment is greater than a preset threshold.

[0017] Optionally, the interactive scene comprises a breaking scene in which a user obtains a right to speak during a customer service speaking process; and in the case of the breaking scene, the marking module is further configured to, in the case that the audio segment comprises background noise audio or background user voice audio, mark the audio segment as noise.

[0018] Optionally, the training feature set generation module is further configured to, for each marked audio segment, extract a pure voice audio segment in the audio segment, and segment the pure voice audio segment according to a preset audio segment length threshold to obtain a plurality of audio segmented segments; and perform feature extraction on each audio segmented segment to generate feature information.

[0019] Optionally, the interactive scene comprises a filtering scene in which the effectiveness of user voice is identified; and in the case of the filtering scene, the marking module is further configured to, in the case that the audio segment comprises background user voice audio, mark the audio segment as voice.

[0020] Optionally, the training feature set generation module is further configured to, for each marked audio segment, extract a pure voice audio segment in the audio segment, and perform feature extraction on the pure voice audio segment to generate feature information.

[0021] Optionally, the training feature set includes feature information and a label result, the feature information has multiple dimension features; the model training module is further configured to: take the feature information as a training input, take the label result as a training target, perform model training to obtain model parameters, and generate an importance ranking of the multiple dimension features; according to the importance ranking, adjust the model parameters, and perform model training again based on the training feature set until the recognition accuracy of the model meets a preset requirement, and take the model whose recognition accuracy meets the preset requirement as the noise recognition model.

[0022] According to a further aspect of the embodiments of the present application, an electronic device is provided.

[0023] An electronic device includes one or more processors; a memory for storing one or more programs that, when executed by the one or more processors, cause the one or more processors to implement the method for training a noise recognition model provided by the embodiments of the present application.

[0024] According to a further aspect of the embodiments of the present application, a computer readable medium is provided.

[0025] A computer readable medium having stored thereon a computer program that, when executed by a processor, implements the method for training a noise recognition model provided by the embodiments of the present application.

[0026] An embodiment of the above invention has the following advantages or beneficial effects: the obtained audio segment is determined to correspond to an interaction scene according to an interaction scene division rule; for each audio segment, a corresponding tagging rule is obtained according to the interaction scene of the audio segment, and the audio segment is tagged according to the tagging rule; feature extraction is performed on each audio segment after tagging to obtain a training feature set; model training is performed based on the training feature set to generate a noise recognition model, which is used for noise recognition of audio segments generated in a voice interaction process; by dividing the interaction scene, tagging rules for different interaction scenes are tagged, training data is generated to train the noise recognition model, the accuracy of tagging can be improved, the recognition effect of voice and noise dimensions is taken into account, and the training effect and recognition accuracy of the noise recognition model are improved.

[0027] The further effects of the above non-conventional optional mode will be described in the following in combination with the specific embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0028] The accompanying drawings are used to better understand the present application and do not constitute an improper limitation on the present application. Among them:

[0029] Figure 1is a schematic diagram of main steps of a training method of a noise recognition model according to an embodiment of the present application;

[0030] Figure 2 is a schematic diagram of interaction of a break scene according to an embodiment of the present application;

[0031] Figure 3 is a schematic diagram of interaction of a filtering scene according to an embodiment of the present application;

[0032] Figure 4 is a schematic diagram of audio clip tagging according to an embodiment of the present application;

[0033] Figure 5 is a schematic diagram of a flow of a training method of a noise recognition model according to an embodiment of the present application;

[0034] Figure 6 is a schematic diagram of main modules of a training device of a noise recognition model according to an embodiment of the present application;

[0035] Figure 7 is an exemplary system architecture diagram to which embodiments of the present application can be applied;

[0036] Figure 8 is a structural schematic diagram of a computer system of a terminal device or a server suitable for implementing embodiments of the present application. DETAILED DESCRIPTION

[0037] Exemplary embodiments of the present application are described below with reference to the accompanying drawings, which include various details of the embodiments of the present application to assist in understanding, which should be considered in the context of the present application. Therefore, those of ordinary skill in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present application. Also, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.

[0038] It should be noted that in the technical solutions of the present application, the collection, collection, update, analysis, processing, use, transmission, storage, etc. of user personal information involved in the technical solutions of the present application comply with relevant laws and regulations, are used for legal purposes, and do not violate public order and good customs. Necessary measures are taken for user personal information to prevent illegal access to user personal information data, and to maintain user personal information security, network security and national security.

[0039] The intelligent voice customer service field has become increasingly prominent in replacing traditional manual customer service and solving core interaction scenarios such as a large number of bills and outbound calls in recent years with the rapid iteration of artificial intelligence voice and natural language processing technology. The result of voice recognition of the intelligent voice customer service robot system usually needs to go through a series of audio pre-processing to improve the recognition accuracy of the system. In the audio pre-processing process, audio enhancement is performed through signal processing technology, and the core of the noise reduction technology in the audio enhancement is a noise recognition model, so the accuracy of the noise recognition model will directly determine whether the entire intelligent voice dialogue interaction process can proceed normally.

[0040] Currently, the effect optimization of the noise model is iteratively updated in two directions: one common method is to implement a noise confidence scoring method based on an acoustic model and a neural network engine, to determine whether the current frame of audio is voice or noise by giving a confidence score of each frame of audio, and to give a confidence score of the corresponding frame; another more reliable solution is to perform post-processing based on the noise model, that is, to add more feature information to the score calculated by the noise model, to re-score and predict the result, so as to further optimize the effect in different scenarios.

[0041] However, both solutions need to be based on the standard for evaluating the noise model, that is, the accuracy / recall rate of each of the voice and noise dimensions needs to be statistically calculated and the overall effect of the model needs to be evaluated. No matter which solution is iterated and evolved, it is always difficult to solve the problem of "trade-off" when optimizing the noise model, that is, when the voice index is improved, the noise index will often decrease, and even through threshold adjustment, it is difficult to greatly improve the overall index. If you want to consider the voice and noise results at the same time, you need to introduce more data, but this may cause overfitting of the training parameters.

[0042] The fundamental reason why the voice / noise effect cannot be considered at the same time is that when the data set is labeled during the training of the noise model, there will be some labeling standards that are difficult to define, and the uncertainty between the audio and the label will cause a negative impact on the model training. Therefore, the problem to be solved by the present application is to propose a model training method based on multi-scene classification labeling for the technical problem of being unable to consider the voice / noise effect at the same time.

[0043] Figure 1 is a main step schematic diagram of the noise recognition model training method according to an embodiment of the present application.

[0044] As shown in Figure 1 the noise recognition model training method according to an embodiment of the present application mainly includes the following steps S101 to S104.

[0045] Step S101: According to the interactive scene division rule, the interactive scene corresponding to the obtained audio segment is determined. The interactive scene can include a breaking scene in which the user obtains the right to speak during the speech of the customer service, and a filtering scene in which the effectiveness of the user's speech is identified.

[0046] Specifically, the breaking scene is a scene in which the user side (the person who is talking on the phone) wants to obtain the right to speak during the speech of the intelligent voice customer service; and the filtering scene is a scene in which the effectiveness of the speech content of the user side is identified.

[0047] During the speech of the voice customer service, the user can interrupt the speech content being performed by the customer service and insert the speech content that the user wants to express. This type of scene is referred to as a "breaking scene" in intelligent human-computer dialogue, that is, the behavior in which the user side forcibly interrupts the broadcast of the system during the broadcast of the speech of the intelligent robot.

[0048] After the speech of the voice customer service is completed, the speech of the user side is waited for, and the next round of reply is performed according to the speech content of the user. In intelligent human-computer dialogue, the effectiveness of the speech of the user side will affect natural language understanding and intent recognition, and therefore it is necessary to judge whether the speech content of the current user is meaningful, whether the speech content of the current user side is spoken by the main speaker, or whether the current user side is background noise. For this type of user side speech content identification scene, it is referred to as a "filtering scene".

[0049] In an embodiment, before the interactive scene corresponding to the obtained audio segment is determined according to the interactive scene division rule, it can further include: determining that the audio segment includes user voice audio, and that the ratio of the length of the user voice audio to the length of the audio segment is greater than a preset threshold.

[0050] Specifically, it is judged whether the audio segment contains user voice audio (i.e., human voice audio). In the case where the audio segment contains user voice audio, if the ratio of the length of the user voice audio to the length of the audio segment is greater than a preset threshold (such as 10%), the audio segment is marked as speech. In the case where the ratio of the length of the user voice audio to the length of the audio segment does not reach the preset threshold, or the audio segment does not contain user voice audio, the audio segment is marked as noise. In the VAD (Voice Activity Detection, audio activity detection) algorithm, if the audio segment contains a large number of silent silent zones, the VAD will not cut most of the silent zones into the audio segment, and therefore when the audio segment is labeled, it is necessary to distinguish according to the proportion of the user voice audio in the audio in advance.

[0051] Step S102: For each audio segment, a corresponding marking rule is obtained according to the interactive scene of the audio segment, and the audio segment is marked according to the marking rule.

[0052] Figure 2 is an interaction schematic diagram of a breaking scenario according to an embodiment of the present application.

[0053] As shown in Figure 2 the noise recognition model is used to identify the current user-side voice and determine whether it is the user's real speech voice and intention. If the noise recognition model determines that it is voice, it indicates that the user wants to interrupt the speech of the intelligent customer service robot; if the noise recognition model determines that it is noise, it indicates that it is only environmental background noise or there is a lot of noisy voice around the user, and it does not mean that the user wants to interrupt the speech of the customer service. Therefore, in the breaking scenario, if the noise recognition model determines that the noise is voice, it will cause the customer service speech to be frequently interrupted, which is unacceptable and should be avoided as much as possible; if the noise recognition model determines that the voice is noise, it will only cause part of the user-side voice to fail to interrupt the broadcast of the customer service, and the user has no real intention to interrupt, which is acceptable in the breaking scenario. Therefore, for the breaking scenario, the optimization focus is to improve the prediction accuracy of the noise recognition model for voice.

[0054] In the case of the breaking scenario, the audio segment is labeled according to the labeling rule, which can include: in the case that the audio segment includes background noise audio or background user voice audio, the audio segment is labeled as noise.

[0055] Specifically, during the labeling process, each audio segment includes various complex situations, and the two types of background noise and background voice need to be further refined. In the voice robot breaking scenario, frequent false interruptions have a greater impact on system interaction, so when generating training data biased towards the breaking scenario, the labeling of audio segments including background noise and background voice of non-main speakers (i.e. background user voice audio) needs to be modified to noise.

[0056] Figure 3 is an interaction schematic diagram of a filtering scenario according to an embodiment of the present application.

[0057] As shown in Figure 3As shown, in the filtering scenario, after the user speaks in the current round, the noise recognition model needs to recognize the sound emitted by the user side in the current round to determine whether the current user is really speaking. If the noise recognition model judges that it is speech, it means that the current user's conversation in this round is normal, and needs to answer the questions of the customer service; if the noise recognition model judges that it is noise, it means that the sound on the user side at present is not real human voice speaking, but caused by background noise. Therefore, in the filtering scenario, if the noise recognition model predicts speech as noise, the speech spoken by the current user cannot be converted into text through ASR (Automatic Speech Recognition), which will cause serious problems in the subsequent intent recognition of NLP (Natural Language Processing), which is unacceptable and should be avoided as much as possible; if the noise recognition model judges that the noise is speech, even if the user does not really speak in the current round, it can be filtered out as meaningless by NLU (Natural Language Understanding) after being converted into text through ASR, which does not affect the normal interaction of the system. Therefore, for the filtering scenario, the optimization focus is to improve the prediction accuracy of the noise recognition model for noise.

[0058] In the case of the filtering scenario, the audio segment is labeled according to the labeling rule, which can include: in the case that the audio segment includes background user speech audio, the audio segment is labeled as speech.

[0059] Specifically, since false detection often leads to misjudgment of user speech as noise, resulting in lack of user speech input in the robot interaction process, leading to smooth interaction, it is necessary to modify the label of the audio segment including background user speech audio (i.e. background human voice) to speech.

[0060] In one embodiment, the user speech audio, the background noise audio, and the background user speech audio can correspond to one or more audio labeling types respectively, wherein the audio labeling types include: dialect data of the main speaker, audible voice of the main speaker, inaudible voice of the main speaker, heavy accent voice of the main speaker, main speaker data frame loss, background noise, inaudible voice of the secondary speaker, secondary speaker data frame loss, audible voice of the secondary speaker, multiple people speaking at the same time in the same time period, customer service personnel voice, and synthesized voice. For example, the user speech audio can correspond to the dialect data of the main speaker and the audible voice of the main speaker; the background noise audio can correspond to the background noise; and the background user speech audio can correspond to the inaudible voice of the secondary speaker, the audible voice of the secondary speaker, and multiple people speaking at the same time in the same time period.

[0061] Step S103: Extract features from each tagged audio segment to obtain a training feature set. The training feature set may include feature information and labeling results.

[0062] Feature information can have multiple dimensions, including one or more of the following: noise confidence, language model score, acoustic model score, Bayesian minimum risk confidence score, candidate word results, audio signal state, and audio start and end times. Specifically, noise confidence is extracted using MFCC (Mel-frequency cepstral coefficients); language model score, acoustic model score, Bayesian minimum risk confidence score, and candidate word results are extracted using ASR (Automatic Speech Recognition); and audio signal state, audio start and end times are extracted using VAD (Audio Activity Detection).

[0063] Figure 4 This is a schematic diagram of audio segment tagging according to an embodiment of the present invention.

[0064] like Figure 4 As shown, audio segments obtained through VAD (Audio Activity Detection) are generally not very accurate. Audio segments include silent regions and clean speech regions. Silence regions are typically located at the beginning and end of the audio (e.g.,...). Figure 4 The audio segment contains silent regions 1 and 2. Therefore, it is necessary to extract the pure speech from the audio segment to obtain the pure speech audio segment. This can be achieved by using the time of the first and last character (e.g., ...). Figure 4 (The message in the text is "Hello, could you please tell me where my package is now?")

[0065] In one embodiment, in the case of interrupted scene, feature extraction for each tagged audio segment may include: for each tagged audio segment, extracting the pure speech audio segment from the audio segment, and segmenting the pure speech audio segment according to a preset audio segment length threshold to obtain multiple audio segment segments; and performing feature extraction on each audio segment segment to generate feature information.

[0066] Specifically, interrupted scenarios are highly sensitive to user speech detection, requiring minimizing the data length of audio stream segments. To ensure the effectiveness of the noise recognition model, the length of audio segments needs to be rationally set. After extracting pure speech segments from the audio stream, the data is processed according to a preset audio segment length threshold (Inter-pausal Unit, i.e., ... Figure 4 The interruption granularity in the code can be set to 200ms to segment the pure speech audio segment, resulting in multiple audio segment segments (e.g., ...). Figure 4IPU 1, IPU 2, …, IPU N in the audio segmentation result). Feature extraction is performed on each audio segment respectively, to generate feature information corresponding to each audio segment, and the label of each audio segment is the label of the audio segment.

[0067] In an embodiment, in the filtering scenario, the feature extraction on each audio segment after labeling can include: for each audio segment after labeling, extracting a pure speech audio segment in the audio segment, and performing feature extraction on the pure speech audio segment to generate feature information.

[0068] Specifically, in the filtering scenario, since noise detection is performed on the entire speech content of the speaker, the pure speech part in the entire speech needs to be obtained for feature extraction, that is, the filtering scenario granularity is the entire pure speech audio segment. After extracting the pure speech audio segment in the audio segment, feature extraction is directly performed on the pure speech audio segment.

[0069] Step S104: model training based on the training feature set to generate a noise recognition model, the noise recognition model being used for noise recognition on an audio segment generated in a voice interaction process.

[0070] In an embodiment, the model training based on the training feature set to generate the noise recognition model can include: taking the feature information as a training input, taking the label result as a training target, performing model training to obtain model parameters, and generating an importance ranking of the multiple dimension features; according to the importance ranking, adjusting the model parameters, and re-performing model training based on the training feature set until the recognition accuracy of the model meets a preset requirement, and taking the model with the recognition accuracy meeting the preset requirement as the noise recognition model.

[0071] Specifically, the training feature set can include feature information and label results of training audio segments, wherein the training audio segments are audio segmentation segments or pure speech audio segments. The model training method uses XGBoost (a kind of machine learning algorithm) based on a gradient descent algorithm. XGBoost generally has better performance in model stability, and the model training process is relatively simple. By importing the training feature set, XGBoostSaveModel (i.e., saving the final gradient iteration decision tree after training) is used for training and saving. The model parameters in the model training process can mainly include max_depth (the depth of constructing a tree, the greater the easier overfitting), num_round (the number of iterations, the more the slower the training process), and min_child_weight (the minimum sample weight sum of a child node, if the weight sum of a leaf node is less than the value, the splitting process ends).

[0072] The trained multiple models are respectively used for accuracy rate statistics and calculation with the training feature set, the stability of the models is verified, and the importance of the dimension features in the model training process is saved. Through model experiment, the importance of the dimension features is in the order of: language model score approximately acoustic model score > Bayes score approximately candidate word length > noise confidence > speech signal start and end time. According to the importance order, the model parameters are adjusted, and the model training is performed again based on the training feature set until the recognition accuracy of the model meets the preset requirement (for example, the recognition accuracy reaches 90%), and the model meeting the preset requirement of the recognition accuracy is used as the noise recognition model.

[0073] The embodiment of the present application can effectively improve the robustness of the noise recognition model, thereby improving the overall interactive experience of the intelligent voice customer service dialogue system.

[0074] Figure 5 It is a flowchart of a training method of a noise recognition model according to an embodiment of the present application.

[0075] As shown in Figure 5 , the embodiment of the present application proposes an overall implementation scheme for optimizing the effect of the noise recognition model based on multiple scenes, noise recognition model feature selection, optimization strategy and model training method in different scenes. According to the interactive scene division rule, the interactive scene corresponding to the obtained audio segment is determined, the labeling rule corresponding to the interactive scene is obtained, and the audio segment is labeled according to the labeling rule. Feature extraction is performed on each audio segment after labeling in multiple dimensions to obtain a training feature set. Model training is performed based on the training feature set to generate a noise recognition model.

[0076] Figure 6 It is a main module schematic diagram of a training device of a noise recognition model according to an embodiment of the present application.

[0077] As shown in Figure 6 , the training device 600 of the noise recognition model of one embodiment of the present application mainly includes: an interactive scene determination module 601, a labeling module 602, a training feature set generation module 603 and a model training module 604.

[0078] The interactive scene determination module 601 is used for determining the interactive scene corresponding to the obtained audio segment according to the interactive scene division rule.

[0079] The labeling module 602 is used for obtaining the corresponding labeling rule according to the interactive scene of the audio segment for each audio segment, and labeling the audio segment according to the labeling rule.

[0080] The training feature set generation module 603 is used for performing feature extraction on each audio segment after labeling to obtain a training feature set.

[0081] The model training module 604 is configured to perform model training based on the training feature set, and generate a noise recognition model, which is used for noise recognition on an audio segment generated in a voice interaction process.

[0082] In an embodiment, the audio segment determination module (not shown in the figure) can also be included, and is configured to: determine that the audio segment includes user voice audio, and that a ratio of a length of the user voice audio to a length of the audio segment is greater than a preset threshold.

[0083] In an embodiment, the interaction scenario can include a breaking scenario in which the user obtains the right to speak during the customer service speaking process; in the case of the breaking scenario, the marking module 602 is specifically configured to: in a case where the audio segment includes background noise audio or background user voice audio, mark the audio segment as noise.

[0084] In an embodiment, the training feature set generation module 603 is specifically configured to: for each marked audio segment, extract a pure voice audio segment in the audio segment, and segment the pure voice audio segment according to a preset audio segment length threshold to obtain a plurality of audio segmentation segments; and perform feature extraction on each audio segmentation segment respectively to generate feature information.

[0085] In an embodiment, the interaction scenario can include a filtering scenario in which the effectiveness of the user voice is recognized; in the case of the filtering scenario, the marking module 602 is specifically configured to: in a case where the audio segment includes background user voice audio, mark the audio segment as voice.

[0086] In an embodiment, the training feature set generation module 603 is specifically configured to: for each marked audio segment, extract a pure voice audio segment in the audio segment, and perform feature extraction on the pure voice audio segment to generate feature information.

[0087] In an embodiment, the training feature set can include feature information and a marking result, and the feature information can have a plurality of dimension features; the model training module 604 is specifically configured to: take the feature information as a training input, take the marking result as a training target, perform model training to obtain model parameters, and generate an importance ranking of the plurality of dimension features; according to the importance ranking, adjust the model parameters, and perform model training again based on the training feature set until an identification accuracy of the model meets a preset requirement, and take the model whose identification accuracy meets the preset requirement as the noise recognition model.

[0088] In addition, the specific implementation content of the training device of the noise recognition model in the embodiment of the present application has been described in detail in the above training method of the noise recognition model, and therefore the repeated content will not be described here.

[0089] Figure 7An exemplary system architecture 700 of a training method of a noise recognition model or a training device of a noise recognition model to which embodiments of the present application can be applied is shown.

[0090] As shown in Figure 7 , the system architecture 700 can include terminal devices 701, 702, 703, a network 704 and a server 705. The network 704 is a medium for providing a communication link between the terminal devices 701, 702, 703 and the server 705. The network 704 can include various connection types, such as wired, wireless communication links or fiber optic cables, etc.

[0091] A user can use the terminal devices 701, 702, 703 to interact with the server 705 through the network 704 to receive or send messages, etc. Various communication client applications can be installed on the terminal devices 701, 702, 703, such as noise recognition applications, voice interaction applications, intelligent customer service applications, instant messaging tools, email clients, social platform software, etc. (only as examples).

[0092] The terminal devices 701, 702, 703 can be various electronic devices with display screens and supporting web browsing, including but not limited to smartphones, tablet computers, laptop computers and desktop computers, etc.

[0093] The server 705 can be a server providing various services, such as a background management server supporting a noise recognition website browsed by a user using a terminal device 701, 702, 703 (only as an example). The background management server can divide the received data such as the training request of the noise recognition model according to the interaction scene division rule, determine the interaction scene corresponding to the obtained audio segment; for each audio segment, obtain the corresponding labeling rule according to the interaction scene of the audio segment, and label the audio segment according to the labeling rule; extract features from each labeled audio segment to obtain a training feature set; perform model training based on the training feature set to generate a noise recognition model, the noise recognition model is used for noise recognition and other processing of audio segments generated in a voice interaction process, and the processing result (such as the training result of the noise recognition model - only as an example) is fed back to the terminal device.

[0094] It should be noted that the training method of the noise recognition model provided by the embodiments of the present application is generally executed by the server 705, and correspondingly, the training device of the noise recognition model is generally provided in the server 705.

[0095] It should be understood that Figure 7 the number of terminal devices, networks and servers in is only illustrative. According to the needs of implementation, there can be any number of terminal devices, networks and servers.

[0096] Reference will now be made to Figure 8 which shows a structural diagram of a computer system 800 suitable for use in implementing the terminal device or server of embodiments of the present application. Figure 8 The terminal device or server shown is merely an example and should not impose any limitation on the functions and scope of use of embodiments of the present application.

[0097] As shown in Figure 8 , the computer system 800 includes a central processing unit (CPU) 801 which can perform various appropriate actions and processes in accordance with a program stored in a read only memory (ROM) 802 or a program loaded from a storage section 808 into a random access memory (RAM) 803. In the RAM 803, various programs and data required for the operation of the system 800 are also stored. The CPU 801, the ROM 802, and the RAM 803 are connected to each other through a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0098] The following components are connected to the I / O interface 805: an input section 806 including a keyboard, a mouse, etc.; an output section 807 including a display such as a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 808 including a hard disk, etc.; and a communication section 809 including a network interface card such as a LAN card, a modem, etc. The communication section 809 performs communication processing via a network such as the Internet. A drive 810 is also connected to the I / O interface 805 as necessary. A removable media 811 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc. is attached to the drive 810 as necessary, so that a computer program read therefrom is installed into the storage section 808 as necessary.

[0099] In particular, the processes described above with reference to the flowcharts can be implemented as a computer software program in accordance with embodiments of the present application. For example, embodiments of the present application include a computer program product comprising a computer program carried on a computer readable medium, the computer program containing program code for executing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network by the communication section 809 and / or installed from the removable media 811. When the computer program is executed by the central processing unit (CPU) 801, the above-described functions defined in the system of the present application are performed.

[0100] It should be noted that the computer-readable medium shown in the present application can be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium may, for example, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or component, or any combination of the above. More specific examples of computer-readable storage media can include, but are not limited to, an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present application, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, device or component. In the present application, the computer-readable signal medium can include a data signal carried in a baseband or as a part of a carrier wave, which carries computer-readable program code. Such a propagated data signal can take various forms, including but not limited to an electromagnetic signal, an optical signal or any suitable combination of the above. The computer-readable signal medium can also be any computer-readable medium other than the computer-readable storage medium, which can send, propagate or transmit a program for use by or in conjunction with an instruction execution system, device or component. The program code contained on the computer-readable medium can be transmitted by any suitable medium, including but not limited to wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0101] The flowcharts and block diagrams in the drawings illustrate the possible implementation architectures, functions and operations of the systems, methods and computer program products according to various embodiments of the present application. In this regard, each block in the flowcharts or block diagrams can represent a module, a program segment or a part of code containing one or more executable instructions for implementing the specified logical functions. It should also be noted that in some alternative implementations, the functions noted in the blocks can occur in different order than that shown in the drawings. For example, two blocks that are shown in succession can actually be executed substantially in parallel, and sometimes in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams or flowcharts, and the combination of blocks in the block diagrams or flowcharts, can be implemented by a dedicated hardware-based system that performs the specified functions or operations, or can be implemented by a combination of dedicated hardware and computer instructions.

[0102] The modules described in the embodiments of the present application can be implemented in the form of software or in the form of hardware. The described modules can also be arranged in a processor, for example, a processor can be described as including an interactive scene determination module, a tagging module, a training feature set generation module, and a model training module. In some cases, the names of these modules do not constitute a limitation on the modules themselves, for example, the interactive scene determination module can also be described as "a module for determining the interactive scene corresponding to the obtained audio segment according to the interactive scene division rule".

[0103] As another aspect, the present application also provides a computer readable medium, which can be included in the device described in the above embodiments, or can exist independently without being assembled into the device. The computer readable medium carries one or more programs, which, when executed by the device, cause the device to include: determining the interactive scene corresponding to the obtained audio segment according to the interactive scene division rule; for each audio segment, obtaining the corresponding tagging rule according to the interactive scene of the audio segment, and tagging the audio segment according to the tagging rule; performing feature extraction on each tagged audio segment to obtain a training feature set; and performing model training based on the training feature set to generate a noise recognition model, which is used for noise recognition of the audio segment generated in the voice interaction process.

[0104] According to the technical solutions of the embodiments of the present application, the interactive scene corresponding to the obtained audio segment is determined according to the interactive scene division rule; for each audio segment, the corresponding tagging rule is obtained according to the interactive scene of the audio segment, and the audio segment is tagged according to the tagging rule; feature extraction is performed on each tagged audio segment to obtain a training feature set; and model training is performed based on the training feature set to generate a noise recognition model, which is used for noise recognition of the audio segment generated in the voice interaction process. By dividing the interactive scene, tagging according to the tagging rule of different interactive scenes, generating training data to train the noise recognition model, the accuracy of tagging can be improved, and the recognition effect of the voice and noise dimensions is taken into account, and the training effect and recognition accuracy of the noise recognition model are improved.

[0105] The above specific embodiments do not constitute a limitation on the scope of protection of the present application. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made depending on design requirements and other factors. Any modifications, equivalent replacements, and improvements made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A training method for a noise recognition model, characterized in that, The method comprises the following steps: According to the interaction scene division rule, determine the corresponding interaction scene of the obtained audio segment; For each audio segment, obtain the corresponding tagging rule according to the interaction scene of the audio segment, and tag the audio segment according to the tagging rule; Feature extraction is performed on each tagged audio segment to obtain a training feature set; Based on the training feature set, model training is performed to generate a noise recognition model, which is used for noise recognition of audio segments generated in the voice interaction process; The interaction scene includes a filtering scene for validity recognition of user voice; In the case of the filtering scene, the tagging of the audio segment according to the tagging rule comprises: in the case that the audio segment includes background user voice audio, marking the audio segment as voice.

2. The method of claim 1, wherein, Before the step of determining the corresponding interaction scene of the obtained audio segment according to the interaction scene division rule, the method further comprises the steps of: Determine that the audio segment includes user voice audio, and the ratio of the length of the user voice audio to the length of the audio segment is greater than a preset threshold.

3. The method of claim 1, wherein, The interaction scene includes a breaking scene in which the user obtains the right to speak during the customer service speaking process; In the case of the breaking scene, the tagging of the audio segment according to the tagging rule comprises: In the case that the audio segment includes background noise audio or background user voice audio, mark the audio segment as noise.

4. The method of claim 3, wherein, The feature extraction of each tagged audio segment comprises: For each tagged audio segment, extract the pure voice audio segment in the audio segment, and segment the pure voice audio segment according to a preset audio segment length threshold to obtain a plurality of audio segmentation segments; Respectively extract features of each audio segmentation segment to generate feature information.

5. The method of claim 1, wherein, The feature extraction of each tagged audio segment comprises: For each tagged audio segment, extract the pure voice audio segment in the audio segment, and extract features of the pure voice audio segment to generate feature information.

6. The method of claim 1, wherein, The training feature set includes feature information and label results, and the feature information has multiple dimension features; The model training based on the training feature set to generate a noise recognition model comprises: Take the feature information as the training input, take the label result as the training target, perform model training to obtain model parameters, and generate an importance ranking of the multiple dimension features; According to the importance ranking, adjust the model parameters, and perform model training again based on the training feature set until the recognition accuracy of the model meets the preset requirement, and take the model with the recognition accuracy meeting the preset requirement as the noise recognition model. 7.A device for training a noise recognition model, comprising: The method comprises the following steps: An interaction scene determination module is configured to determine the corresponding interaction scene of the obtained audio segment according to the interaction scene division rule; A tagging module is configured to obtain the corresponding tagging rule for each audio segment according to the interaction scene of the audio segment, and tag the audio segment according to the tagging rule; The training feature set generation module is configured to perform feature extraction on each audio segment after marking to obtain a training feature set; The model training module is configured to perform model training based on the training feature set to generate a noise recognition model, which is configured to perform noise recognition on an audio segment generated in a voice interaction process; The interaction scenario includes a filtering scenario in which validity of user voice is recognized; in the filtering scenario, the marking module is specifically configured to: in a case where the audio segment includes background user voice audio, mark the audio segment as voice.

8. An electronic device, comprising: comprising: one or more processors; a memory device for storing one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement the method of any one of claims 1-6.

9. A computer readable medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the method of any one of claims 1-6. The computer program is executed by the processor to implement the method of any one of claims 1-6.

Images