Howling suppression method, electronic device, and computer-readable storage medium

Abstract

The application discloses a howling suppression method, an electronic device and a computer readable storage medium. The method comprises the following steps: obtaining input audio, determining initial audio features of the input audio, adjusting the initial audio features by using reference audio features to obtain adjusted audio features; wherein the reference audio features are obtained based on sample audio superimposed with feedback audio, and the feedback audio comprises audio obtained after at least one round of playing; determining a howling detection result based on the adjusted audio features, converting the adjusted audio features into output audio; and determining an output volume matched with the output audio based on the howling detection result, and playing the output audio by using the output volume. In the foregoing manner, the howling suppression effect can be improved.

Description

Technical Field

[0001] This application relates to the field of audio processing technology, and in particular to a method for suppressing howling, an electronic device, and a computer-readable storage medium. Background Technology

[0002] With technological advancements, voice calls have expanded to include multi-person conversations within the same physical space, where multiple microphones and speakers are present. In such cases, the transmitting microphone may pick up sound from the receiving speaker, creating an acoustic loop and causing feedback. This not only severely impacts call quality but can also cause physical damage to the hardware. However, current technologies typically suppress feedback only after it has already occurred, resulting in ineffective suppression. Therefore, improving feedback suppression has become a pressing issue. Summary of the Invention

[0003] The main technical problem addressed by this application is to provide a howling suppression method, electronic device, and computer-readable storage medium that can improve howling suppression effect.

[0004] To address the aforementioned technical problems, this application provides a feedback suppression method, comprising: acquiring input audio; determining initial audio features of the input audio; adjusting the initial audio features using reference audio features to obtain adjusted audio features; wherein the reference audio features are obtained based on sample audio superimposed with feedback audio, the feedback audio including audio obtained after at least one round of playback; determining a feedback detection result based on the adjusted audio features; converting the adjusted audio features into output audio; determining an output volume matching the output audio based on the feedback detection result; and playing the output audio using the output volume.

[0005] To address the aforementioned technical problems, a second aspect of this application provides an electronic device comprising: a memory and a processor coupled to each other, wherein the memory stores program data, and the processor invokes the program data to execute the method described in the first aspect.

[0006] To address the aforementioned technical problems, a third aspect of this application provides a computer-readable storage medium storing program data thereon, wherein the program data, when executed by a processor, implements the method described in the first aspect.

[0007] The beneficial effects of this application are as follows: Unlike existing technologies, this application acquires input audio collected by a microphone, determines the initial audio features of the input audio, and adjusts the initial audio features using reference audio features to obtain adjusted audio features. The reference audio features are obtained based on sample audio overlaid with feedback audio, which includes audio obtained after at least one playback cycle. By comparing the reference audio features with the initial audio features, features that match the feedback audio can be effectively identified. This allows for adjustment of the initial audio features when feedback is imminent, suppressing features that match the feedback audio and reducing the probability of feedback. Based on the adjusted audio features, feedback is detected, the detection result is determined, and the adjusted audio features are converted into output audio. Based on the feedback detection result, a matching output volume is set for the output audio, adaptively adjusting the output volume according to the feedback suppression effect. The output audio is then played using the output volume. Through early feedback suppression and adaptive volume adjustment, the probability of feedback is minimized to improve the feedback suppression effect. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart illustrating one embodiment of the howling suppression method of this application; Figure 2 This is a flowchart illustrating another embodiment of the howling suppression method of this application; Figure 3 This is a schematic diagram of one embodiment of the howling suppression model of this application; Figure 4 This is a schematic diagram of the structure of one embodiment of the detection network of this application; Figure 5 This is a schematic diagram of the structure of one embodiment of the electronic device of this application; Figure 6 This is a schematic diagram of one embodiment of the computer-readable storage medium of this application. Detailed Implementation

[0009] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments, and different implementation methods can be adaptively combined. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0010] In this paper, the terms "system" and "network" are often used interchangeably. The term "and / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship. Furthermore, "many" in this paper means two or more.

[0011] The feedback suppression method provided in this application is used to reduce the probability of feedback when multiple sets of microphones and speakers exist in the same physical space, and the corresponding execution subject is a processing unit capable of data processing.

[0012] Please see Figure 1 , Figure 1 This is a flowchart illustrating one embodiment of the howling suppression method of this application, which includes: S101: Acquire input audio, determine the initial audio features of the input audio, adjust the initial audio features using reference audio features, and obtain the adjusted audio features; wherein, the reference audio features are obtained based on sample audio with feedback audio superimposed, and the feedback audio includes audio obtained after at least one round of playback.

[0013] Specifically, the input audio collected by the microphone is acquired, the initial audio features of the input audio are determined, and the initial audio features are adjusted using reference audio features to obtain adjusted audio features. The reference audio features are obtained based on sample audio overlaid with feedback audio, which includes audio obtained after at least one round of playback.

[0014] It should be noted that by comparing the reference audio features with the initial audio features, the features in the initial audio features that match the feedback audio can be effectively identified. This allows the initial audio features to be adjusted using the reference audio features when a feedback is about to occur, thereby suppressing the features that match the feedback audio and reducing the probability of feedback.

[0015] It is understandable that the sample audio corresponds to the original audio without any added interference, while the feedback audio corresponds to the audio played through the propagation links between different groups of microphones and speakers, and the feedback audio is played at least once in the corresponding propagation link. Therefore, the sample audio superimposed with the feedback audio corresponds to the initial stage where feedback is about to occur.

[0016] In one embodiment, a reference audio feature is associated with a feedback audio, an input audio sampled by a microphone is acquired, features of the input audio are extracted to obtain an initial audio feature, features similar to the reference audio feature are extracted from the initial audio feature as features to be adjusted, and the initial audio feature is adjusted using the features to be adjusted to eliminate the features to be adjusted in the initial audio feature, thereby obtaining the adjusted audio feature.

[0017] In one embodiment, a reference audio feature is associated with a sample audio, an input audio collected by a microphone is acquired, features of the input audio are extracted to obtain an initial audio feature, the feature difference between the initial audio feature and the reference audio feature is obtained as the feature to be adjusted, and the initial audio feature is adjusted using the feature to be adjusted to eliminate the feature to be adjusted in the initial audio feature, thereby obtaining the adjusted audio feature.

[0018] Optionally, a suppression network is used to extract features from the input audio to obtain initial audio features. After supervised training, the suppression network retains reference audio features. The initial audio features are then adjusted using the reference audio features in the suppression network to obtain adjusted audio features. The suppression network is trained using sample audio with superimposed feedback audio.

[0019] In some implementation scenarios, multiple sample audio segments are acquired, some of the sample audio segments are delayed and convolved using the acoustic feedback path generated by the closed-loop system to obtain feedback audio. The feedback audio is superimposed with the undelayed sample audio to obtain training audio. The sample audio and feedback audio are used as training labels, and the training audio is used to train the suppression network to suppress the features associated with the feedback audio, so that the suppression network records the reference audio features associated with the feedback audio after training.

[0020] In some implementation scenarios, individual sample audio is collected from an interference-free recording environment, while sample audio with feedback audio is collected from a sample environment where feedback occurs. The sample audio collected from the recording environment is used as training labels, and the suppression network is trained using the sample audio with feedback audio to suppress features unrelated to the sample audio, so that the suppression network records reference audio features associated with the sample audio after training.

[0021] S102: Based on the adjusted audio features, determine the feedback detection result and convert the adjusted audio features into output audio.

[0022] Specifically, based on the adjusted audio features, the system detects whether feedback will occur, determines the feedback detection result, and converts the adjusted audio features into output audio.

[0023] In one embodiment, the adjusted audio features are classified into two categories to obtain two feedback detection results: feedback occurs or feedback does not occur. The adjusted audio features are then converted into data to obtain the output audio.

[0024] In one embodiment, standard audio features without feedback are acquired, the feature similarity between the adjusted audio features and the standard audio features is determined, the probability value of feedback is determined based on the feature similarity, feedback detection results are obtained, and the adjusted audio features are decoded to obtain output audio. The initial audio features are obtained through encoding.

[0025] Optionally, the adjusted audio features are input into the detection network to obtain the feedback detection result output by the detection network. The detection network and the suppression network are connected and jointly trained using sample audio with superimposed feedback audio. The training loss includes the network losses corresponding to the suppression network and the detection network, respectively. The suppression network ultimately converts the adjusted audio features into the output audio.

[0026] S103: Based on the feedback detection results, determine the output volume that matches the output audio, and play the output audio using the output volume.

[0027] Specifically, based on the feedback detection results, a matching output volume is set for the output audio, thereby adaptively setting the output volume according to the feedback suppression effect. The output audio is then played using the output volume. By suppressing feedback in advance and adjusting the volume adaptively, the probability of feedback occurring is reduced as much as possible to improve the feedback suppression effect.

[0028] In one embodiment, the feedback detection result corresponds to feedback occurring or not occurring. Based on the current feedback detection result, a current volume adjustment value is determined. The set volume at the speaker end is adjusted using the volume adjustment value to obtain an output volume that matches the output audio. The output audio is then played using the output volume. Specifically, when the feedback detection result indicates feedback has occurred, the volume adjustment value is a preset value used to reduce the set volume.

[0029] In one embodiment, the feedback detection result corresponds to feedback occurring and feedback not occurring, the input audio corresponds to the audio segment segmented by the audio collected by the microphone, each audio segment corresponds to a feedback detection result, based on the feedback detection result obtained up to the current time as the number of feedbacks, a volume adjustment value positively correlated with the number is determined, the volume adjustment value is used to reduce the set volume at the speaker end, the output volume matching the output audio is obtained, and the output audio is played using the output volume.

[0030] In one embodiment, the feedback detection result corresponds to the probability value of feedback occurrence. The historical volume of the previous moment is obtained, and the volume adjustment value that is positively correlated with the probability value is obtained. The output volume matching the output audio is obtained by subtracting the volume adjustment value from the historical volume, and the output audio is played using the output volume.

[0031] Optionally, a volume controller is used to determine the output volume that matches the output audio. The volume controller is connected to the detection network and the suppression network. The volume controller obtains the feedback detection result output by the detection network, and then sets the matching output volume for the output audio based on the feedback detection result, and applies the output volume to the final output audio output by the suppression network.

[0032] Understandably, feedback suppression methods are based on feedback suppression models, which include suppression networks, detection networks, and volume controllers. By deploying feedback suppression models on the propagation links corresponding to each group of microphones and speakers, the probability of feedback can be reduced, thus improving the feedback suppression effect in the entire multi-person dialogue scenario.

[0033] The above scheme acquires input audio from the microphone, determines the initial audio features of the input audio, and adjusts the initial audio features using reference audio features to obtain adjusted audio features. The reference audio features are obtained based on sample audio overlaid with feedback audio, which includes audio obtained after at least one playback cycle. By comparing the reference audio features with the initial audio features, features that match the feedback audio can be effectively identified. This allows for adjustment of the initial audio features when feedback is imminent, suppressing features that match the feedback audio and reducing the probability of feedback. Based on the adjusted audio features, feedback is detected, and the adjusted audio features are converted into output audio. Based on the feedback detection result, a matching output volume is set for the output audio, adaptively adjusting the output volume according to the feedback suppression effect. The output audio is then played using the output volume. Through early feedback suppression and adaptive volume adjustment, the probability of feedback is minimized to improve feedback suppression effectiveness.

[0034] Please see Figure 2 , Figure 2 This is a flowchart illustrating another embodiment of the howling suppression method of this application, the method comprising: S201: Acquire input audio, determine the initial audio features of the input audio, adjust the initial audio features using reference audio features, and obtain the adjusted audio features; wherein, the reference audio features are obtained based on sample audio with feedback audio superimposed, and the feedback audio includes audio obtained after at least one round of playback.

[0035] Specifically, the input audio is acquired via a microphone, and the initial audio features of the input audio are determined. Based on the reference audio features and the initial audio features, the features to be adjusted that need to be adjusted for the initial audio features are determined, and the initial audio features are adjusted to obtain the adjusted audio features. The reference audio features are obtained based on sample audio overlaid with feedback audio, which includes audio obtained after at least one playback cycle.

[0036] In one embodiment, adjusting initial audio features using reference audio features to obtain adjusted audio features includes: determining a feature to be adjusted in the initial audio features based on the reference audio features and the initial audio features; wherein the feature to be adjusted is associated with noise audio and feedback audio in the input audio; and adjusting the initial audio features using the feature to be adjusted to obtain the adjusted audio features.

[0037] Specifically, based on reference audio features and initial audio features, the features to be adjusted are parsed from the initial audio features. The features to be adjusted are associated with noise audio and feedback audio in the input audio so as to further eliminate noise audio in the environment in addition to suppressing feedback audio.

[0038] Furthermore, the initial audio features are adjusted using the features to be adjusted, suppressing the features to be adjusted in the initial audio features to obtain adjusted audio features. This suppresses features in the input audio that are associated with noise audio and feedback audio, ensuring the accuracy of the adjusted audio features and improving the purity of the output audio obtained based on the adjusted audio features.

[0039] It should be noted that the reference audio features include feedback features associated with the feedback audio and sample features associated with the sample audio. Based on the reference audio features and the initial audio features, the features to be adjusted in the initial audio features are determined, including: based on the feedback features and the initial audio features, extracting features from the initial audio features whose similarity to the feedback features exceeds a similarity threshold, as the first candidate features; based on the sample features and the initial audio features, extracting features from the initial audio features that deviate from the sample features, as the second candidate features; and based on the first candidate features and the second candidate features, obtaining the features to be adjusted in the initial audio features.

[0040] Specifically, the reference audio features include feedback features associated with the feedback audio and sample features associated with the sample audio. The feedback features represent the features associated with the feedback audio when a howling is about to occur, and the sample features represent the features corresponding to the audio that has not yet experienced a howling.

[0041] It is understandable that the feedback features are compared with the initial audio features, and features with a similarity exceeding a similarity threshold are extracted from the initial audio features as the first candidate features. Thus, the first candidate features are obtained by mining features similar to the feedback features in the initial audio features.

[0042] In addition, by comparing the sample features with the initial audio features, the deviation between the initial audio features and the sample features is determined, and the features that deviate from the sample features are extracted as second candidate features. Thus, the second candidate features are obtained by comparing the deviation between the initial audio features and the sample features.

[0043] Furthermore, by combining the first and second candidate features obtained from the two approaches, the features to be adjusted in the initial audio features are obtained, thereby improving the accuracy of the features to be adjusted.

[0044] It is understandable that the first and second candidate features are fused to obtain the fused features, and the fused features are used as the features to be adjusted.

[0045] In some implementation scenarios, the weights of the first candidate feature and the second candidate feature are obtained, and the first candidate feature and the second candidate feature are weighted and fused to obtain the feature to be adjusted. Among them, the weight of the first candidate feature is positively correlated with the similarity, and the sum of the weights of the first candidate feature and the second candidate feature is a preset value, thereby adaptively adjusting the weights of the two to improve the fusion accuracy.

[0046] It should be noted that there are multiple sample audios, and these sample audios are superimposed with feedback audios of different delays. The feedback features are obtained by fusing the features corresponding to the feedback audios of different delays, and the sample features are obtained by fusing the features corresponding to the multiple sample audios.

[0047] Specifically, there are multiple sample audio files, each superimposed with feedback audio, and multiple sample audio files are superimposed with feedback audio at different delays. The feedback features are obtained by fusing features corresponding to feedback audio at different delays, in order to obtain representative features corresponding to feedback audio at different delays, thus ensuring the accuracy of the first candidate features extracted under different network environments. The sample features are obtained by fusing features corresponding to multiple sample audio files, in order to obtain representative features corresponding to sample audio with different expressions, thus ensuring that relatively accurate second candidate features can be extracted for input audio from different speakers.

[0048] S202: Based on the adjusted audio features, determine the feedback detection result and convert the adjusted audio features into output audio.

[0049] Specifically, based on the adjusted audio features, the system detects whether feedback will occur, determines the feedback detection result, and converts the adjusted audio features into output audio.

[0050] In some implementation scenarios, the howling suppression method is based on the howling suppression model, which includes a suppression network, a detection network, and a volume controller. The suppression network includes an encoder, an amplifier, and a decoder connected in sequence. The volume controller is connected after the detection network, and both are connected in parallel with the decoder.

[0051] Specifically, please refer to Figure 3 , Figure 3 This is a schematic diagram of an embodiment of the howling suppression model of this application. The suppression network includes an encoder, an enhancer, and a decoder connected in sequence. The encoder includes multiple coding layers, the enhancer includes a recurrent convolutional layer, and the decoder includes multiple decoding layers. The input of the detection network is connected to the output of the enhancer, the input of the volume controller is connected to the output of the detection network, and the output of the volume controller is connected to the output of the decoder.

[0052] It should be noted that the howling suppression model is trained using multiple sample audios with superimposed feedback audio. The encoder is used to determine the initial audio features, the enhancer is used to adjust the initial audio features to obtain adjusted audio features, the encoder is used to convert the adjusted audio features into output audio, the detection network is used to determine the howling detection result, and the volume controller is used to determine the output volume. The detection network includes pooling layers, convolutional layers, and an output layer. The pooling layer is used to extract multi-dimensional key information of the adjusted audio features, including channels, time, and frequency. The convolutional layer is used to fuse the multi-dimensional key information to obtain fused information, and the output layer is used to output the howling detection result based on the fused information.

[0053] Specifically, the howling suppression model is trained using multiple sample audios superimposed with feedback audio. During the training process of the howling suppression model, the sample audios are used as labels, and the training loss includes the network losses corresponding to the suppression network and the detection network, respectively. When the training loss converges, the trained howling suppression model is obtained.

[0054] Understandably, there are multiple sample audio files, each superimposed with feedback audio. These multiple sample audio files are superimposed with feedback audio files of varying delays. The reference audio features include feedback features associated with the feedback audio files and sample features associated with the sample audio files. During the training of the feedback suppression model, the model continuously optimizes the feedback features obtained by fusing features corresponding to feedback audio files of different delays, as well as the sample features obtained by fusing features corresponding to multiple sample audio files. Finally, after the feedback suppression model training is complete, the final feedback features and sample features are recorded.

[0055] Furthermore, after obtaining the trained howling suppression model, the input audio is input to the encoder to obtain the initial audio features output by the encoder. The initial audio features are then input to the enhancer to adjust the initial audio features using reference audio features, resulting in adjusted audio features. The adjusted audio features are then input to the detection network to obtain the howling detection results output by the detection network. The adjusted audio features are then passed to the encoder to obtain the output audio output by the encoder. The volume controller determines the output volume based on the howling detection results and matches the output volume with the output audio, thereby playing the output audio according to the output volume.

[0056] It should be noted that due to limited deployment resources, the detection network adopts a lightweight network structure, consisting of pooling layers, convolutional layers, and an output layer. Specifically, determining the feedback detection result based on adjusted audio features involves: inputting the adjusted audio features into the pooling layer for information extraction, obtaining multi-dimensional key information output by the pooling layer; these multi-dimensional features include channel, time, and frequency; inputting this multi-dimensional key information into the convolutional layer for information fusion, obtaining fused information output by the convolutional layer; and inputting this fused information into the output layer for result judgment, obtaining the feedback detection result output by the output layer.

[0057] Specifically, please refer to Figure 4 , Figure 4 This is a schematic diagram of one embodiment of the detection network in this application. The pooling layer can reduce the amount of computation, extract and adjust key information in the audio features, and reduce the frequency dimension through pooling or convolution, as shown below: ,in, The corresponding dimensions are channel, time, and frequency, respectively. The feature map is averaged in the frequency dimension through a pooling layer, and the dimension is adjusted to... Convolutional layers typically include one-dimensional convolutions to achieve feature fusion between channels. By fusing information between channels using convolutions, fused information is obtained. Output layers typically include linear layers and a sigmoid function to output the howling detection results. Thus, the lightweight network structure described above can save deployment resources and improve detection efficiency.

[0058] S203: In response to the current feedback detection result being feedback, obtain the first statistical number of consecutive feedback detection results being feedback, and determine the adjustment ratio of the output audio based on the first statistical number and a preset number threshold.

[0059] Specifically, the input audio corresponds to the audio segments segmented by the audio collected by the microphone. Each audio segment corresponds to a feedback detection result. When the current feedback detection result indicates that feedback has occurred, the first statistical number of consecutive feedback detection results indicating that feedback has occurred is obtained. Based on the relationship between the first statistical number and the preset number threshold, the adjustment ratio of the output audio is adaptively determined in order to accurately control the output volume.

[0060] In some implementation scenarios, quantity threshold matching has a lower limit of proportion; based on the first statistical quantity and the preset quantity threshold, the adjustment proportion of the output audio is determined, including: in response to the first statistical quantity being zero, determining the adjustment proportion of the output audio as the maximum proportion; in response to the first statistical quantity being between zero and the quantity threshold, generating the adjustment proportion of the output audio based on the lower limit of proportion and the ratio of the first statistical quantity to the quantity threshold; in response to the first statistical quantity reaching the quantity threshold, determining the adjustment proportion of the output audio as the lower limit of proportion.

[0061] Specifically, the quantity threshold matching has a proportional lower limit. When the first statistical quantity is zero, the adjustment ratio of the output audio is determined to be the maximum ratio, where the maximum ratio is usually 1, meaning that the volume remains unchanged when the first statistical quantity is zero. When the first statistical quantity is between zero and the quantity threshold, the ratio of the first statistical quantity to the quantity threshold is obtained. Using this ratio and the proportional lower limit, an adjustment ratio that adapts to the first statistical quantity is generated, and this adjustment ratio is between the proportional lower limit and the maximum ratio. When the first statistical quantity reaches the quantity threshold, the adjustment ratio of the output audio is determined to be the proportional lower limit, thereby reducing the output volume as much as possible and avoiding continuous volume reduction leading to mute.

[0062] Understandably, comparing the first statistical count with the count threshold allows for a more accurate adjustment ratio based on continuous howling detection results, avoiding the impact of errors caused by single detection results. The above process is expressed by the following formula: (1) in, For output volume, To set the volume, This is the first statistical quantity. This is the lower limit of the ratio. This is the quantity threshold.

[0063] S204: Adjust the set volume of the output audio matching by adjusting the adjustment ratio to obtain the output volume matching the output audio, and play the output audio using the output volume.

[0064] Specifically, the speaker end corresponding to the output audio has a set volume. The set volume is adjusted by adjusting the ratio to obtain the output volume of the output audio. The output audio is played using the output volume. Thus, if the feedback detection result indicates the presence of feedback, the set volume can be adjusted in time by adjusting the ratio so that the output audio can be played with a more suitable output volume, thereby reducing the probability of feedback.

[0065] In some implementation scenarios, the volume is set to a preset fixed volume, and the fixed value is adjusted according to the adjustment ratio.

[0066] In some implementation scenarios, the device volume is the output volume from the previous moment, and the volume is continuously adjusted in real time according to the adjustment ratio.

[0067] Optionally, after adjusting the set volume of the output audio matching by adjusting the adjustment ratio to obtain the output volume matching the output audio, and playing the output audio using the output volume, the method further includes: in response to the output volume of the previous moment being less than the reference volume and the current feedback detection result being no feedback, obtaining a second statistical number of consecutive feedback detection results up to the current moment that are no feedback, obtaining a recovery ratio based on the second statistical number and the ratio of the output volume of the previous moment to the reference volume; adjusting the output volume of the previous moment using the recovery ratio to obtain the current output volume, and playing the output audio using the current output volume.

[0068] Specifically, the feedback detection results are continuously output. When the output volume at the current moment is less than the reference volume and the current feedback detection result is no feedback, the second statistical number of consecutive feedback detection results up to the current moment with no feedback is obtained, and the ratio of the output volume at the previous moment to the reference volume is determined. Based on the second statistical number and the corresponding ratio, the recovery ratio is obtained. The recovery ratio is greater than 1 and less than or equal to the upper limit of the ratio. The upper limit of the ratio corresponds to the quotient of the reference volume and the output volume at the previous moment. The recovery ratio is positively correlated with the second statistical number and negatively correlated with the ratio of the output volume at the previous moment to the reference volume.

[0069] Furthermore, the output volume of the previous moment is adjusted using the recovery ratio to obtain the current output volume. The output audio is then played using the current output volume. Thus, after the feedback is suppressed, the volume is adaptively and gradually restored based on the second statistical quantity to ensure the clarity of the dialogue.

[0070] Please see Figure 5 , Figure 5This is a schematic diagram of an embodiment of the electronic device of this application. The electronic device 30 includes a memory 301 and a processor 302 coupled to each other. The memory 301 stores program data (not shown in the figure). The processor 302 calls the program data to implement the method in any of the above embodiments. For the description of the relevant content, please refer to the detailed description of the above method embodiments, which will not be repeated here.

[0071] Please see Figure 6 , Figure 6 This is a schematic diagram of a computer-readable storage medium according to an embodiment of the present application. The computer-readable storage medium 40 stores program data 400. When the program data 400 is executed by a processor, it implements the method in any of the above embodiments. For a detailed description of the relevant content, please refer to the detailed description of the above method embodiments, which will not be repeated here.

[0072] It should be noted that the units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0073] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0074] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods of various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0075] The above description is merely an embodiment of this application and does not limit the scope of protection of this application. Any equivalent structural or procedural transformations 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 scope of protection of this application.

Claims

1. A method for suppressing howling, characterized in that, The method includes: The input audio is acquired, the initial audio features of the input audio are determined, and the initial audio features are adjusted using reference audio features to obtain adjusted audio features; wherein, the reference audio features are obtained based on sample audio overlaid with feedback audio, and the feedback audio includes audio obtained after at least one round of playback; Based on the adjusted audio features, the feedback detection result is determined, and the adjusted audio features are converted into output audio. Based on the feedback detection results, the output volume matching the output audio is determined, and the output audio is played using the output volume.

2. The howling suppression method according to claim 1, characterized in that, The step of adjusting the initial audio features using reference audio features to obtain adjusted audio features includes: Based on the reference audio features and the initial audio features, features to be adjusted in the initial audio features are determined; wherein, the features to be adjusted are associated with noise audio and feedback audio in the input audio; The initial audio features are adjusted using the features to be adjusted to obtain the adjusted audio features.

3. The howling suppression method according to claim 2, characterized in that, The reference audio features include feedback features associated with the feedback audio and sample features associated with the sample audio; The step of determining the features to be adjusted in the initial audio features based on the reference audio features and the initial audio features includes: Based on the feedback features and the initial audio features, features whose similarity to the feedback features exceeds a similarity threshold are extracted from the initial audio features and used as first candidate features; Based on the sample features and the initial audio features, features that deviate from the sample features are extracted from the initial audio features and used as second candidate features; Based on the first candidate feature and the second candidate feature, the features to be adjusted in the initial audio features are obtained.

4. The howling suppression method according to claim 3, characterized in that, The number of sample audios is multiple, and the multiple sample audios are superimposed with feedback audios with different delays. The feedback features are obtained by fusing the features corresponding to the feedback audios with different delays, and the sample features are obtained by fusing the features corresponding to the multiple sample audios.

5. The howling suppression method according to claim 1, characterized in that, The step of determining the output volume matching the output audio based on the feedback detection result, and playing the output audio using the output volume, includes: In response to the current feedback detection result indicating feedback has occurred, the first statistical number of consecutive feedback detection results indicating feedback has occurred up to the current time is obtained, and the adjustment ratio of the output audio is determined based on the first statistical number and a preset number threshold. The output audio is adjusted using the adjustment ratio to match the set volume, resulting in an output volume that matches the output audio. The output audio is then played using the output volume.

6. The howling suppression method according to claim 5, characterized in that, The quantity threshold matching has a proportional lower limit; Determining the adjustment ratio of the output audio based on the first statistical quantity and a preset quantity threshold includes: In response to the first statistical count being zero, the adjustment ratio of the output audio is determined to be the maximum ratio; In response to the first statistical quantity being between zero and a quantity threshold, an adjustment ratio for the output audio is generated based on the lower limit of the ratio and the ratio of the first statistical quantity to the quantity threshold; In response to the first statistical quantity reaching the quantity threshold, the adjustment ratio of the output audio is determined to be the lower limit of the ratio.

7. The howling suppression method according to claim 5, characterized in that, After adjusting the set volume of the output audio using the adjustment ratio to obtain the output volume matching the output audio, and playing the output audio using the output volume, the method further includes: In response to the previous output volume being less than the reference volume and the current feedback detection result being no feedback, a second statistical number of consecutive feedback detection results up to the current time being no feedback is obtained. Based on the second statistical number and the ratio of the previous output volume to the reference volume, the recovery ratio is obtained. The output volume at the previous moment is adjusted using the recovery ratio to obtain the current output volume, and the output audio is played using the current output volume.

8. The howling suppression method according to claim 1, characterized in that, The feedback suppression method is based on a feedback suppression model, which includes a suppression network, a detection network, and a volume controller. The suppression network includes an encoder, an amplifier, and a decoder connected in sequence. The volume controller is connected after the detection network, and both are connected in parallel with the decoder. The feedback suppression model is trained using multiple sample audio samples superimposed with the feedback audio. The encoder determines the initial audio features, the enhancer adjusts the initial audio features to obtain the adjusted audio features, the encoder converts the adjusted audio features into output audio, the detection network determines the feedback detection result, and the volume controller determines the output volume. The detection network includes a pooling layer, a convolutional layer, and an output layer. The pooling layer extracts multi-dimensional key information from the adjusted audio features, including channels, time, and frequency. The convolutional layer fuses the multi-dimensional key information to obtain fused information, and the output layer outputs the feedback detection result based on the fused information.

9. An electronic device, characterized in that, include: A memory and a processor are coupled to each other, wherein the memory stores program data, and the processor invokes the program data to perform the method as described in any one of claims 1-8.

10. A computer-readable storage medium storing program data thereon, characterized in that, When the program data is executed by the processor, it implements the method as described in any one of claims 1-8.

Images