Audio quality testing methods, equipment, storage media and software products

By integrating handcrafted features, deep audio features, and pre-trained audio features, an audio quality detection method has been developed that addresses the accuracy and generalization issues of speaker audio quality detection, achieving multi-dimensional, comprehensive, and accurate audio quality detection.

CN123075347APending Publication Date: 2026-11-10GOERTEK INC
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611560124.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-10
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

In existing technologies, loudspeaker audio quality testing methods suffer from high subjectivity and low accuracy. In particular, manual listening tests and single-dimensional feature detection are difficult to accurately characterize noise in the audio played by loudspeakers, and lack generalization across sound sources and devices.

Method used

A fusion method of handcrafted features, deep audio features, and pre-trained audio features is adopted. By acquiring the target audio signal, handcrafted features, deep audio features, and pre-trained audio features are extracted, and the audio quality detection result is determined based on the fused features. Multi-dimensional detection is achieved by utilizing the physical directivity of handcrafted features, the local anomaly patterns of deep audio features, and the general audio representation of pre-trained features.

Benefits of technology

It improves the accuracy and generalization ability of audio quality detection, enabling multi-dimensional and comprehensive characterization of audio quality, adapting to different speaker models and audio content, and enhancing the accuracy and adaptability of detection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123075347A_ABST
    Figure CN123075347A_ABST
Patent Text Reader

Abstract

This application discloses an audio quality detection method, device, storage medium, and program product, belonging to the field of computer technology. The method includes: acquiring a target audio signal; extracting features from the target audio signal to obtain hand-crafted features, deep audio features, and pre-trained audio features; fusing the hand-crafted features, deep audio features, and pre-trained audio features; and determining the audio quality detection result based on the fused features. In this application, hand-crafted features have a clear physical orientation towards the nonlinear distortion and structural anomalies of loudspeakers, deep audio features can automatically capture local anomaly patterns in the time-frequency domain, and pre-trained features provide rich general audio representations. Based on these three features, the model can perform audio quality detection in multiple dimensions, from micro to macro, thereby greatly improving the accuracy of the audio quality detection results.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to an audio quality detection method, device, storage medium, and program product. Background Technology

[0002] As a core component of audio equipment, the sound quality of loudspeakers directly impacts the user experience. During loudspeaker manufacturing, factors such as manufacturing processes, material defects, and assembly errors can cause various noises to occur when the loudspeakers play sound, severely affecting product quality. Therefore, how to test loudspeaker quality is a crucial issue. One related technology uses manual listening tests to assess loudspeaker quality. However, this method is highly subjective and has low accuracy. Summary of the Invention

[0003] This application provides an audio quality detection method that can improve the accuracy of audio quality detection. The technical solution is as follows: Firstly, an audio quality detection method is provided, the method comprising: Acquire the target audio signal; Feature extraction is performed on the target audio signal to obtain handcrafted features, deep audio features, and pre-trained audio features; The handcrafted features, the deep audio features, and the pre-trained audio features are fused together, and the audio quality detection result is determined based on the fused features.

[0004] In this application, a target audio signal is acquired, and features are extracted from the target audio signal to obtain handcrafted features, deep audio features, and pre-trained audio features. These handcrafted features, deep audio features, and pre-trained audio features are then fused, and the audio quality detection result is determined based on the fused features. Because handcrafted features have a clear physical orientation to the nonlinear distortion and structural anomalies of loudspeakers, deep audio features can automatically capture local anomaly patterns in the time-frequency domain, and pre-trained features provide rich and general audio representations, the model can perform multi-dimensional, micro-to-macro audio quality detection based on these three types of features, thereby significantly improving the accuracy of the audio quality detection results.

[0005] Optionally, the step of extracting features from the target audio signal to obtain handcrafted features, deep audio features, and pre-trained audio features includes: Based on the target audio signal, one or more acoustic data are determined, including one or more of the following: total harmonic distortion data, spectral flatness, spectral centroid, spectral roll-off point, and zero-crossing rate; the handcrafted features are generated based on the one or more acoustic data. The target audio signal is subjected to time-frequency feature extraction to obtain the depth audio features; The target audio signal is input into a pre-trained audio model to obtain the pre-trained audio features output by the pre-trained audio model.

[0006] Optionally, the step of extracting time-frequency features from the target audio signal to obtain the depth audio features includes: A time-frequency representation matrix is ​​generated based on the target audio signal, the time-frequency representation matrix including a time dimension and a frequency dimension; The time-frequency representation matrix is ​​subjected to deep feature extraction using a convolutional neural network (CNN) to obtain the first time-frequency feature; The depth audio features are generated based on the first time-frequency features.

[0007] Optionally, generating the depth audio features based on the first time-frequency features includes: Based on the first time-frequency feature, a time attention weight and / or a frequency attention weight are determined. The time attention weight is used to characterize the probability that a time frame in the target audio signal contains noise, and the frequency attention weight is used to characterize the energy intensity of noise in the frequency band of the target audio signal. Based on the time attention weight and / or the frequency attention weight, the first time-frequency feature is weighted to obtain the second time-frequency feature; The second time-frequency feature is pooled to obtain the depth audio feature.

[0008] Optionally, determining the audio quality detection result based on the handcrafted features, the deep audio features, and the pre-trained audio features includes: The handcrafted features, the deep audio features, and the pre-trained audio features are fused to obtain the first audio feature; The audio quality detection result is determined based on the first audio feature, and the audio quality detection result includes one or more of the following: noise detection result and noise severity.

[0009] Optionally, fusing the handcrafted features, the deep audio features, and the pre-trained audio features to obtain the first audio feature includes: Based on the handcrafted features, the deep audio features, and the pre-trained audio features, a gating mechanism is used to determine the first weight corresponding to the handcrafted features, the second weight corresponding to the deep audio features, and the third weight corresponding to the pre-trained audio features. Based on the first weight, the second weight, and the third weight, the handcrafted features, the deep audio features, and the pre-trained audio features are weighted and fused to obtain the first audio feature.

[0010] Optionally, fusing the handcrafted features, the deep audio features, and the pre-trained audio features to obtain the first audio feature includes: The handcrafted features, the deep audio features, and the pre-trained audio features are concatenated to obtain the second audio feature; Based on the second audio feature, the fourth weight corresponding to the handcrafted feature, the fifth weight corresponding to the deep audio feature, and the sixth weight corresponding to the pre-trained audio feature are determined through an attention mechanism; Based on the fourth weight, the fifth weight, and the sixth weight, the handcrafted features, the deep audio features, and the pre-trained audio features are weighted and fused to obtain the first audio feature.

[0011] Secondly, an audio quality detection device is provided, the device comprising: The acquisition module is used to acquire the target audio signal; The feature extraction module is used to extract features from the target audio signal to obtain handcrafted features, deep audio features, and pre-trained audio features; The determination module is used to determine the audio quality detection result based on the handcrafted features, the deep audio features, and the pre-trained audio features.

[0012] Thirdly, a computer device is provided, the computer device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the audio quality detection method described in the first aspect.

[0013] Fourthly, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in the first aspect.

[0014] Fifthly, a computer program product is provided, which, when run on a computer device, causes the computer device to perform the audio quality detection method described in the first aspect.

[0015] It is understood that the beneficial effects of the second, third, fourth, and fifth aspects mentioned above can be found in the relevant descriptions in the first aspect above, and will not be repeated here. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the structure of an audio quality detection system provided in an embodiment of this application; Figure 2This is a flowchart of an audio quality detection method provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of an audio quality detection device provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0017] In the following description, specific details such as particular system architectures and technologies are set forth for illustrative purposes and not for limiting purposes, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details.

[0018] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or collections thereof. The terms "comprising," "including," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0019] It should be understood that "one or more" as used in this application refers to one, two, or more, and "multiple" as used in this application refers to two or more. In the description of this application, unless otherwise stated, " / " means "or," for example, A / B can mean A or B. "And / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone.

[0020] To facilitate a clear description of the technical solutions of this application, the terms "first" and "second" are used to distinguish identical or similar items with essentially the same function and effect. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and that the terms "first" and "second" do not necessarily imply that they are different.

[0021] The terms "one embodiment" or "some embodiments" used in this application mean that one or more embodiments of this application include the specific features, structures, or characteristics described in that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this application do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized.

[0022] The application scenarios involved in the embodiments of this application are described below.

[0023] As a core component of audio equipment, the sound quality of loudspeakers directly affects the user experience. During the loudspeaker manufacturing process, factors such as manufacturing defects, material defects, and assembly errors can cause various noises (such as distortion, howling, and mechanical vibration noise) to be generated when the loudspeaker amplifies sound, which seriously affects product quality.

[0024] In related technologies, the following methods are commonly used to detect the audio quality played by a speaker.

[0025] The first method: Manual listening test, in which a person listens to the audio played by the speaker and evaluates the sound quality accordingly.

[0026] The second method involves using a machine learning classifier to extract features from the original audio signal played by the speaker, and then using a pre-trained classifier to detect whether there is airflow noise in the original audio signal, thereby achieving speaker noise detection.

[0027] The third approach is to use a multi-scale convolutional feature fusion method to perform speaker noise detection, then determine the noise threshold through a pre-trained classification model, and use this to determine whether the audio signal contains noise.

[0028] However, all three methods described above have certain problems to varying degrees. The first method, relying on subjective judgment, suffers from low accuracy, high cost, and low efficiency. The second and third methods, firstly, have limitations in accurately representing problems in the audio played by the speaker due to their single-dimensional features; they lack sensitivity in detecting certain types of noise and have limited feature representation capabilities. Secondly, they lack targeted training for different speaker models and audio content, resulting in poor detection performance. Furthermore, they lack generalization design for cross-source and cross-device scenarios, limiting their practical application scope.

[0029] Therefore, this application provides an audio quality detection method. For example, this method can be applied to the audio quality detection system described below. In this method, a target audio signal is acquired, and features are extracted from the target audio signal to obtain handcrafted features, deep audio features, and pre-trained audio features. These handcrafted features, deep audio features, and pre-trained audio features are then fused, and the audio quality detection result is determined based on the fused features. Since handcrafted features have a clear physical orientation towards nonlinear distortion and structural anomalies of loudspeakers, deep audio features can automatically capture local anomaly patterns in the time-frequency domain, and pre-trained features provide rich general audio representations, the model can perform multi-dimensional, micro-to-macro audio quality detection based on these three features, thereby significantly improving the accuracy of the audio quality detection results.

[0030] The audio quality detection system provided in the embodiments of this application will be described below.

[0031] Figure 1 This is a schematic diagram of the structure of an audio quality detection system provided in an embodiment of this application. See also... Figure 1 The audio quality detection system 10 may include an audio acquisition module 101, a preprocessing module 102, a multimodal feature extraction module 103, a feature fusion module 104, and a classification decision module 105.

[0032] For example, the audio quality inspection system 10 can be applied to scenarios such as speaker production quality control and audio equipment anomaly detection.

[0033] The audio acquisition module 101 is used to acquire audio signals and send them to the preprocessing module 102. For example, the audio acquisition module 101 may include a microphone. For example, the audio acquisition module 101 may acquire audio signals played by a speaker using the microphone.

[0034] The preprocessing module 102 receives the audio signal sent by the audio acquisition module 101, preprocesses the audio signal, and inputs the preprocessed audio signal into the multimodal feature extraction module 103. For example, the preprocessing may include noise reduction, normalization, etc., but this embodiment does not limit the scope of the application.

[0035] The multimodal feature extraction module 103 receives the audio signal input from the preprocessing module 102, extracts features from the audio signal to obtain the corresponding handcrafted features, deep audio features (also known as CNN deep features), and pre-trained audio features, and inputs the handcrafted features, the deep audio features (also known as CNN deep features), and the pre-trained audio features into the feature fusion module 104. For example, the multimodal feature extraction module 103 can extract the handcrafted features and deep audio features in parallel.

[0036] Optionally, the multimodal feature extraction module 103 may include a deep audio feature extraction unit, a pre-trained audio feature extraction unit, and a hand-crafted feature extraction unit.

[0037] The deep audio feature extraction unit is used to extract deep audio features from the audio signal. Specifically, this unit performs a time-frequency transformation on the audio signal to obtain a time-frequency representation matrix. This time-frequency representation matrix is ​​then expanded into a tensor suitable for input to a two-dimensional convolutional network. A CNN feature extraction network then extracts feature information from this tensor level by level. This feature information may include local time-frequency texture, harmonic structure, transient anomalies, and higher-level acoustic modes. Finally, this feature information is aggregated to obtain the deep audio features (also known as deep feature vectors).

[0038] This pre-trained audio feature extraction unit is used to extract pre-trained audio features that characterize the acoustic event categories and acoustic properties of audio signals. For example, this pre-trained audio feature extraction unit can input an audio signal into a pre-trained model to obtain the pre-trained audio features of the audio signal. Specifically, it can take the feature map before the last pooling layer of the pre-trained model, perform average pooling to obtain a fixed-dimensional embedding vector e, and then perform dimensionality reduction processing on this embedding vector to obtain e' (i.e., the pre-trained audio feature). For example, the pre-trained model can be a PANNs model or a YAMNet model, etc., and this application embodiment does not limit this.

[0039] In some cases, backpropagation in a pre-trained model can be disabled to avoid overfitting and other problems. Since the amount of actual labeled data for loudspeakers is usually small, fine-tuning the pre-trained model based on the labeled data parameters during training can easily introduce overfitting. Therefore, backpropagation in the pre-trained model can be disabled (e.g., setting all parameters to require_grad=False) so that the pre-trained model does not participate in backpropagation and is used only as a feature encoder.

[0040] This handcrafted feature extraction unit is used to extract handcrafted features from audio signals. Specifically, this handcrafted feature extraction unit can calculate one or more acoustic data (also known as audio engineering features) based on the audio signal, perform multilayer perceptron (MLP) compression on the one or more acoustic data, and obtain the handcrafted features of the audio signal.

[0041] The feature fusion module 104 is used to receive handcrafted features, deep audio features and pre-trained audio features input from the multimodal feature extraction module 103, fuse the handcrafted features, the deep audio features and the pre-trained audio features to obtain the first audio feature, and input the first audio feature into the classification decision module 105.

[0042] The classification decision module 105 is used to receive the first audio feature input by the feature fusion module 104, classify the first audio feature, and obtain the audio quality detection result.

[0043] Optionally, the classification decision module 105 may include a classifier and a regressor. The classifier is obtained by training and learning the mapping relationship between fusion features (i.e., first audio features) and noise categories, and the classifier can perform binary classification of normal audio and noise audio. The regressor is obtained by learning the mapping relationship between fusion features and noise severity (such as expert subjective rating), and the regressor can predict the noise severity.

[0044] In some cases, data augmentation strategies can be used during the model training phase to improve the model's generalization ability to different speaker types and different audio content.

[0045] For example, different types of audio materials (such as speech, music, ambient sounds, etc.) are used as excitation signals during the training phase to cover diverse sound source scenarios in real-world applications and enhance the diversity of sound sources.

[0046] For example, the frequency response characteristics of different loudspeaker models can be simulated using acoustic modeling or data-driven methods to generate diverse training samples for model training.

[0047] For example, enhancement operations such as time stretching, frequency shifting, and adding background noise can be performed on audio signals to improve the model's robustness to signal changes.

[0048] Targeted data augmentation can improve the model's generalization ability to different speaker types and audio content, enabling the system to adapt to diverse application scenarios and improving the accuracy and recall of identifying different types and levels of speaker noise.

[0049] The audio quality detection method provided in the embodiments of this application will be explained in detail below.

[0050] Figure 2 This is a flowchart of an audio quality detection method provided in an embodiment of this application. See also... Figure 2 The method may include the following steps: Step 201: The computer device acquires the target audio signal.

[0051] For example, the target audio signal can be the audio signal played by the speaker that needs to be detected.

[0052] For example, the target audio signal can be historically acquired, or it can be acquired in real time; this application embodiment does not limit this.

[0053] Acquiring the target audio signal provides a data basis for detecting whether there is noise in the audio played by the speaker.

[0054] For example, after acquiring the target audio signal, the computer device can perform one or more of the following on the target audio signal: noise reduction, normalization, etc., to improve the accuracy of the target audio signal.

[0055] Step 202: The computer device extracts features from the target audio signal to obtain handcrafted features, deep audio features, and pre-trained audio features.

[0056] Handcrafted features are low-dimensional statistical quantities with clear physical meanings, designed and calculated manually based on audio signal processing theory and acoustic physics principles. They are extracted from the original audio waveform or spectrum using fixed formulas and can explicitly characterize the nonlinear distortion, spectral anomalies, and time-frequency characteristics of a loudspeaker.

[0057] Deep audio features refer to high-dimensional semantic feature vectors obtained by automatically learning the time-frequency representation of audio signals through convolutional neural networks, and progressively abstracting them from the data. These deep audio features can represent the local time-frequency texture, harmonic structure, transient anomalies, and higher-level acoustic modes of audio signals. For example, deep audio features can be captured by multiple convolutional blocks through end-to-end training, capturing hierarchical acoustic information from local textures to global patterns in audio signals. Each of these convolutional blocks can include convolutional layers, normalization layers, nonlinear activation layers, and pooling layers.

[0058] Pre-trained audio features refer to general audio embedding vectors extracted from input audio by using a deep neural network model pre-trained on a large-scale general audio dataset as a fixed feature extractor with frozen parameters.

[0059] Because handcrafted features provide physical priors and interpretability, deep audio features are sensitive to distortions and transient noises, and offer local time-frequency sensitivity, while pre-trained audio features provide general audio priors and semantic information, these three complement each other. They utilize the priors from massive amounts of data, retain the determinism of physical rules, and possess sensitivity to local anomalies. This improves the accuracy of subsequent audio quality detection results.

[0060] In some implementations, the computer device may perform feature extraction on the target audio signal to obtain handcrafted features, deep audio features and pre-trained audio features, including the following steps (1) to (3).

[0061] Step (1): The computer device determines one or more acoustic data based on the target audio signal, including one or more of the following: total harmonic distortion data, spectral flatness, spectral centroid, spectral roll-off point, and zero-crossing rate; and generates the handcrafted feature based on the one or more acoustic data.

[0062] Total harmonic distortion (THD) data is used to characterize the degree of nonlinear distortion in a loudspeaker. For example, THD data may include one or more of total harmonic distortion (THD), total harmonic distortion plus noise (THD+N), and individual harmonic distortion components (such as 2nd harmonic distortion, 3rd harmonic distortion, etc.), and this application does not limit this. For example, the energy ratio of the fundamental frequency to each harmonic in the target audio signal can be calculated to obtain the THD data.

[0063] Spectral flatness measures the flatness of a frequency spectrum and identifies spectral anomalies (such as formants, non-uniform frequency response, etc.). For example, spectral flatness can be obtained by calculating the ratio of the geometric mean to the arithmetic mean of the spectrum of a target audio signal.

[0064] Spectral centroid, spectral rolloff, and zero-crossing rate are used to characterize the time-frequency characteristics of a target audio signal in multiple dimensions. The spectral centroid measures the center frequency position of the audio signal's power spectrum, reflecting the brightness or timbre of the sound. It can identify whether the spectral energy distribution is biased towards high or low frequencies, thereby detecting timbre abnormalities (such as muffled, harsh, or missing high frequencies). A high spectral centroid indicates excessive high-frequency energy, potentially corresponding to distortion, feedback, or other noises; a low spectral centroid indicates excessive low-frequency energy, potentially corresponding to abnormal diaphragm damping or a muffled timbre. For example, the power spectrum of each frame of the target audio signal is calculated, the product of the frequency value and the corresponding power is summed, and then divided by the total power to obtain the spectral centroid of that frame.

[0065] The roll-off point describes the frequency value at which the accumulated energy in an audio signal reaches a preset percentage, reflecting the extent of the high-frequency energy extension. The roll-off point can identify abnormal high-frequency energy distribution, such as clipping distortion, high-frequency resonance, or abrupt changes in the frequency response curve. A significant rightward shift (i.e., a higher frequency) indicates an abnormal increase in high-frequency energy, possibly caused by clipping, high-frequency resonance, or noise; a leftward shift indicates excessively rapid high-frequency energy decay, possibly due to insufficient high-frequency response of the speaker. For example, the power spectrum of each frame of the target audio signal can be accumulated to find the minimum frequency at which the accumulated energy first exceeds a preset percentage of the total energy, thus obtaining the roll-off point.

[0066] Zero-crossing rate measures the number of times an audio signal waveform crosses the zero level per unit time, roughly reflecting the frequency and volatility of the information. It can identify anomalies such as transient impacts, clicks, or sudden noises, which cause the waveform to frequently cross zero points within a very short time. When the zero-crossing rate is higher than a preset value for normal audio, it indicates that the frame may contain noise such as clicks, pops, or transient impacts; conversely, a rate that is too low may correspond to information saturation. For example, the zero-crossing rate is calculated by counting the number of sign changes between adjacent sampling points for a frame of the target audio signal, and then dividing by the difference between the number of sampling points N in that frame and 1.

[0067] Since acoustic data is a physical indicator of audio quality, a known audio signal can be quickly identified by determining one or more acoustic data points. Based on these one or more acoustic data points, handcrafted features can be generated, providing an accurate and reliable basis for subsequent analysis.

[0068] In some implementations, the computer device may generate the handcrafted feature based on the one or more acoustic data by performing MLP compression on the one or more acoustic data to obtain the handcrafted feature (such as the encoded vector h) of the audio signal.

[0069] Step (2): The computer device extracts the time-frequency features of the target audio signal to obtain the depth audio features.

[0070] For example, deep audio features can be obtained by extracting time-frequency features from the target audio signal using a deep neural network.

[0071] Time-frequency features can represent the time evolution and frequency distribution of audio signals. By extracting deep audio features, complex noise patterns hidden in the data that are difficult for the human ear to detect (such as periodic noise and transient impact sounds in specific frequency bands) can be automatically learned, thereby extracting nonlinear changes in the target audio signal more accurately.

[0072] In some implementations, the computer device performs time-frequency feature extraction on the target audio signal to obtain the deep audio feature. This operation can be as follows: generating a time-frequency representation matrix based on the target audio signal, the time-frequency representation matrix including a time dimension and a frequency dimension; performing convolutional neural network (CNN) feature extraction on the time-frequency representation matrix to obtain a first time-frequency feature; and generating the deep audio feature based on the first time-frequency feature.

[0073] The time-frequency representation matrix is ​​a matrix that can represent both the time and frequency domain dimensions, obtained by performing a transformation on the target audio signal. It can also be called a spectrum. For example, algorithms such as short-time Fourier transform or Mel filter bank mapping can be used to perform a time-frequency transformation on the target audio signal to obtain the time-frequency representation matrix.

[0074] By generating a time-frequency representation matrix, the advantages of CNN in two-dimensional spatial pattern recognition can be well utilized, thereby efficiently and accurately capturing local and complex time-frequency abnormal patterns in audio signals. This provides a high-quality underlying visual feature foundation for subsequent noise detection.

[0075] In some implementations, the operation of the computer device to generate the deep audio feature based on the first time-frequency feature may be as follows: determining a time attention weight and / or a frequency attention weight based on the first time-frequency feature, wherein the time attention weight is used to characterize the probability that a time frame in the target audio signal contains noise, and the frequency attention weight is used to characterize the energy intensity of noise in the frequency band of the target audio signal; performing weighted processing on the first time-frequency feature based on the time attention weight and / or the frequency attention weight to obtain a second time-frequency feature; and performing pooling processing on the second time-frequency feature to obtain the deep audio feature.

[0076] For example, the computer device can average the first time-frequency feature along its frequency dimension to obtain a temporal attention weight, which is a weight vector that varies over time. This temporal attention weight can represent which time points across the entire frequency band contain the most information. For instance, if the target audio signal distorts at frame 50, then the value of the temporal attention weight at frame 50 will be significantly higher than at other frames.

[0077] For example, the computer device can average the first time-frequency feature over its time dimension to obtain a frequency attention weight, which is a weight vector that varies with frequency. This frequency attention weight can represent which frequency bands have the highest energy throughout the entire time process. For instance, if noise is mainly concentrated in the 2kHz-3kHz range, then the frequency attention weight will have a high response at the corresponding 2kHz-3kHz position.

[0078] In this way, the model can intelligently focus on key time frames and key frequency bands where noise may exist, which can greatly enhance the ability of deep audio features to represent noise and resist interference, and is a key technical link to improve detection accuracy.

[0079] For ease of understanding, the following example illustrates the extraction of time-frequency features from the target audio signal by the computer device, using steps A to D.

[0080] Step A: Input the target audio signal into the computer device.

[0081] For example, the sampling rate, truncation duration, and number of sampling points of the target audio signal can be set according to the speaker detection scenario, device acquisition conditions, and model configuration.

[0082] Step B: The computer device performs time-frequency transformation on the target audio signal to obtain a two-dimensional time-frequency representation matrix M∈R^{F×T}, where F represents the frequency dimension, such as the number of Mel frequency bands; T represents the number of time frames, which is determined by the length of the target audio signal segment, the analysis window length, and the frame shift.

[0083] For example, algorithms such as short-time Fourier transform and Mel filter bank mapping can be used to perform time-frequency transformation on audio signals.

[0084] Step C: The computer device expands the time-frequency representation matrix M into a tensor form (B, 1, F, T) suitable for input to a two-dimensional convolutional network, where B is the batch size, 1 represents a single-channel time-frequency plot, F represents the frequency dimension, and T represents the time dimension. This tensor is then input into a CNN feature extraction network to obtain a deep feature vector.

[0085] For example, with an audio sampling rate of 48 kilohertz (kHz), a 3-second truncation length, 64 Mel frequency bands, and 4 convolutional blocks in the CNN, feature extraction is performed on the target audio signal, and the number of channels is progressively expanded to 16, 32, 64, and 128, ultimately resulting in a 128-dimensional deep feature vector.

[0086] Step D: The computer device performs aggregation on the depth feature vector to obtain a fixed-length representation of the depth audio features.

[0087] Step (3): The computer device inputs the target audio signal into the pre-trained audio model and obtains the pre-trained audio features output by the pre-trained audio model.

[0088] The pre-trained model already possesses powerful general audio understanding capabilities. Using it as a feature extractor can yield advanced semantic features, which can help the model understand the content of the audio (such as whether it is human voice or instrumental sound), thereby better judging the noise present in the target audio signal.

[0089] Step 203: The computer device fuses the handcrafted features, the deep audio features, and the pre-trained audio features, and determines the audio quality detection result based on the fused features.

[0090] The audio quality test results can indicate the quality of the speaker playing the audio signal.

[0091] Since a single feature among the handcrafted feature, the deep audio feature, and the pre-trained audio feature is unlikely to fully represent the noise-related features in the target audio feature, fusing the handcrafted feature, the deep audio feature, and the pre-trained audio feature can detect audio quality in a multi-dimensional, multi-angle, and complementary manner. This can overcome the limitations of a single feature in describing audio quality, comprehensively and accurately represent the quality state of the audio signal, and improve the accuracy of audio quality detection.

[0092] In some implementations, step 203 may involve: fusing the handcrafted feature, the deep audio feature, and the pre-trained audio feature to obtain a first audio feature; and determining the audio quality detection result based on the first audio feature, wherein the audio quality detection result includes one or more of noise detection result and noise severity.

[0093] The first audio feature is an audio feature that can comprehensively represent prior knowledge, physical laws, and local anomalies.

[0094] The noise detection result is used to indicate whether noise exists in the target audio signal. Optionally, the computer device can input the first audio features into a classifier to obtain the noise detection result output by the classifier. This classifier can be trained to learn the mapping relationship between fused features and noise categories, enabling binary classification of normal audio and noisy audio.

[0095] Noise severity is used to indicate the degree of noise in the target audio signal. Optionally, the computer device can input the first audio features into a regressor to obtain the noise severity output by the regressor. This regressor can be obtained by learning the mapping relationship between the fused features and the noise severity (i.e., expert subjective rating), thus enabling the prediction of noise severity. In some cases, the noise severity detection operation is performed only after the noise detection result indicates the presence of noise in the target audio signal, thereby reducing computational overhead.

[0096] By fusing these three audio features, a comprehensive description from micro to macro is achieved. Firstly, handcrafted features focus on physical quantities, deep audio features on local texture, and pre-trained features on overall semantics. This three-layered information overlay provides a complete picture, achieving complementary data dependencies. Secondly, handcrafted features require no training data, deep audio features require a small amount of data, and pre-trained audio features are derived from massive datasets. Combining these three features leverages the prior knowledge of vast amounts of data, preserves the determinism of physical rules, and possesses sensitivity to local anomalies. Therefore, based on the first audio feature, audio quality detection of the target audio signal can be performed comprehensively from micro to macro, improving the accuracy of audio quality detection. Furthermore, handcrafted features offer interpretability based on physical laws, while deep audio features and pre-trained audio features provide detection accuracy. The first audio feature enables both accurate audio quality detection and interpretation of the detection results.

[0097] In some implementations, the computer device may fuse the handcrafted feature, the deep audio feature, and the pre-trained audio feature in three ways to obtain a first audio feature.

[0098] The first method involves the computer device concatenating the handcrafted feature, the deep audio feature, and the pre-trained audio feature to obtain the first audio feature.

[0099] Specifically, the handcrafted feature h, deep audio feature c, and pre-trained audio feature e′ can be concatenated along the feature dimension. In PyTorch, this can be achieved using torch.cat([h, c, e′], dim=1), concatenating along the feature dimension to obtain a joint vector, forming a joint feature representation. Here, dim=1 indicates that the handcrafted feature h, deep audio feature c, and pre-trained audio feature e′ are horizontally connected.

[0100] The second method: The computer device determines the first weight corresponding to the handcrafted feature, the second weight corresponding to the deep audio feature, and the third weight corresponding to the pre-trained audio feature through a gating mechanism based on the handcrafted feature, the deep audio feature, and the pre-trained audio feature; based on the first weight, the second weight, and the third weight, the handcrafted feature, the deep audio feature, and the pre-trained audio feature are weighted and fused to obtain the first audio feature.

[0101] For example, the values ​​of the first weight, the second weight, and the third weight are greater than 0 and less than 1.

[0102] Different types of audio quality problems have varying degrees of dependence on different features. For example, for harmonic distortion, handcrafted features should have high weights; for ambient background noise, deep audio features have higher weights; and for reverberation problems in music recordings, pre-trained audio features have higher weights. By using a gating mechanism, the model can adaptively determine the corresponding weights for each feature based on its own contextual information, thereby improving the model's generalization ability and adaptability in different audio scenarios.

[0103] For example, the computer device can determine the first weight based on this manual feature using the following formula.

[0104] g h = σ(W) h ·h + b h ); Among them, g h σ is the first weight (i.e., the gate value of this handcrafted feature); σ is the activation function (such as the Sigmoid activation function); W h For learnable weight vectors (matched to the handcrafted feature); b h It is a learnable bias scalar used to adjust the offset of the activation function.

[0105] For example, the computer device can determine the second weight based on this depth audio feature using the following formula.

[0106] g c = σ(W) c ·c + b c ); Among them, g c σ is the second weight (i.e., the gate value of this depth audio feature); W is the activation function; c b is a learnable weight vector (matched to the deep audio features); c It is a learnable bias scalar.

[0107] For example, the computer device can determine the third weight based on the pre-trained audio features using the following formula.

[0108] g e = σ(W) e ·e′+ b e ); Among them, g e σ is the third weight (i.e., the gating value of the pre-trained audio feature); W is the activation function; e b is a learnable weight vector (matched to the pre-trained audio features); e It is a learnable bias scalar.

[0109] In some cases, the gating parameter Wh b h W c b c W e b e As the entire network is trained end-to-end, the contribution of various features to the final task is automatically learned through backpropagation.

[0110] In some implementations, the computer device performs a weighted fusion of the handcrafted feature, the deep audio feature, and the pre-trained audio feature based on a first weight, a second weight, and a third weight to obtain the first audio feature. This operation can be: determining the sum of the product of the handcrafted feature and the first weight, the product of the deep audio feature and the second weight, and the product of the pre-trained audio feature and the third weight to obtain the first audio feature.

[0111] For example, the computer device can determine the first audio feature based on the first weight, the second weight, the third weight, the handcrafted feature, the deep audio feature, and the pre-trained audio feature using the following formula.

[0112] f = g h ·h proj + g c ·c proj + g e · e′ proj ; Among them, h proj Features of handcraftsmanship; c proj For depth audio features; e′ proj For example, the pre-trained audio features are vectors formed by mapping the three types of features to the same dimensional space through a linear or multi-layer perceptron.

[0113] The third method involves the computer device concatenating the handcrafted feature, the deep audio feature, and the pre-trained audio feature to obtain a second audio feature. Based on the second audio feature, an attention mechanism is used to determine the fourth weight corresponding to the handcrafted feature, the fifth weight corresponding to the deep audio feature, and the sixth weight corresponding to the pre-trained audio feature. Based on the fourth, fifth, and sixth weights, the handcrafted feature, the deep audio feature, and the pre-trained audio feature are weighted and fused to obtain a first audio feature.

[0114] For example, the computer device can project the handcrafted feature, the deep audio feature, and the pre-trained audio feature into the same dimensional space to obtain the handcrafted feature h in the same dimension. proj The depth audio feature c proj and the pre-trained audio feature e′ proj .

[0115] For example, the sum of the fourth, fifth, and sixth weights is 1.

[0116] By concatenating features before calculating attention, the model can capture more complex interactions and mutual influences between different features. This makes weight allocation depend not only on individual features but also on their combination patterns, allowing the model to learn more refined fusion strategies and further improve its performance ceiling.

[0117] Optionally, the computer device determines the fourth weight corresponding to the handcrafted feature, the fifth weight corresponding to the deep audio feature, and the sixth weight corresponding to the pre-trained audio feature based on the second audio feature through an attention mechanism. This can be achieved by inputting the second audio feature into the attention network softmax(W). a ·[h proj ;c proj e proj ] + b a ), to obtain the attention weight vector a output by the attention network, a = [a h a c a e ], and a h +a c +a e =1. Where softmax() is the normalization function; W a The weight matrix is ​​a learnable matrix; [h proj ;c proj ;e proj ] represents the second audio feature; b a This is a learnable bias vector.

[0118] For example, the computer device can determine the first audio feature based on the fourth, fifth, and sixth weights, the handcrafted feature, the deep audio feature, and the pre-trained audio using the following formula: f = a h ·h proj + a c ·c proj + a e ·e proj ; Where f is the first audio feature; a h It is the fourth weight; a c The fifth weight; a e The sixth weight; h proj Features of handcraftsmanship; c proj For depth audio features; e proj These are pre-trained audio features.

[0119] In this embodiment, the computer device acquires a target audio signal, extracts features from the target audio signal to obtain handcrafted features, deep audio features, and pre-trained audio features, fuses these features, and determines the audio quality detection result based on the fused features. Because handcrafted features have a clear physical orientation to the nonlinear distortion and structural anomalies of the speaker, deep audio features can automatically capture local anomaly patterns in the time-frequency domain, and pre-trained features provide rich general audio representations, the model can perform multi-dimensional, micro-to-macro audio quality detection based on these three features, thereby significantly improving the accuracy of the audio quality detection results.

[0120] Figure 3 This is a schematic diagram of an audio quality detection device provided in an embodiment of this application. The device can be implemented as part or all of a computer device by software, hardware, or a combination of both, and this computer device can be described below. Figure 4 The computer equipment shown. See also Figure 3 The device includes: an acquisition module 301, a feature extraction module 302, and a determination module 303.

[0121] Acquisition module 301 is used to acquire the target audio signal; Feature extraction module 302 is used to extract features from the target audio signal to obtain handcrafted features, deep audio features, and pre-trained audio features; The determination module 303 is used to fuse the handcrafted features, the deep audio features, and the pre-trained audio features, and determine the audio quality detection result based on the fused features.

[0122] Optionally, the feature extraction module 302 is used for: Based on the target audio signal, one or more acoustic data are determined, including one or more of the following: total harmonic distortion data, spectral flatness, spectral centroid, spectral roll-off point, and zero-crossing rate; the handcrafted features are generated based on the one or more acoustic data. The target audio signal is subjected to time-frequency feature extraction to obtain the depth audio features; The target audio signal is input into a pre-trained audio model to obtain the pre-trained audio features output by the pre-trained audio model.

[0123] Optionally, the feature extraction module 302 is used for: A time-frequency representation matrix is ​​generated based on the target audio signal, the time-frequency representation matrix including a time dimension and a frequency dimension; The time-frequency representation matrix is ​​subjected to convolutional neural network (CNN) feature extraction to obtain the first time-frequency feature; The depth audio features are generated based on the first time-frequency features.

[0124] Optionally, the feature extraction module 302 is used for: Based on the first time-frequency feature, a time attention weight and / or a frequency attention weight are determined. The time attention weight is used to characterize the probability that a time frame in the target audio signal contains noise, and the frequency attention weight is used to characterize the energy intensity of noise in the frequency band of the target audio signal. Based on the time attention weight and / or the frequency attention weight, the first time-frequency feature is weighted to obtain the second time-frequency feature; The second time-frequency feature is pooled to obtain the depth audio feature.

[0125] Optionally, the determining module 303 is used for: The handcrafted features, the deep audio features, and the pre-trained audio features are fused to obtain the first audio feature; The audio quality detection result is determined based on the first audio feature, and the audio quality detection result includes one or more of the following: noise detection result and noise severity.

[0126] Optionally, the determining module 303 is used for: Based on the handcrafted features, the deep audio features, and the pre-trained audio features, a gating mechanism is used to determine the first weight corresponding to the handcrafted features, the second weight corresponding to the deep audio features, and the third weight corresponding to the pre-trained audio features. Based on the first weight, the second weight, and the third weight, the handcrafted features, the deep audio features, and the pre-trained audio features are weighted and fused to obtain the first audio feature.

[0127] Optionally, the determining module 303 is used for: The handcrafted features, the deep audio features, and the pre-trained audio features are concatenated to obtain the second audio feature; Based on the second audio feature, the fourth weight corresponding to the handcrafted feature, the fifth weight corresponding to the deep audio feature, and the sixth weight corresponding to the pre-trained audio feature are determined through an attention mechanism; Based on the fourth weight, the fifth weight, and the sixth weight, the handcrafted features, the deep audio features, and the pre-trained audio features are weighted and fused to obtain the first audio feature.

[0128] In this embodiment, a target audio signal is acquired, and features are extracted from the target audio signal to obtain handcrafted features, deep audio features, and pre-trained audio features. These handcrafted features, deep audio features, and pre-trained audio features are then fused, and the audio quality detection result is determined based on the fused features. Because handcrafted features have a clear physical orientation to the nonlinear distortion and structural anomalies of the speaker, deep audio features can automatically capture local anomaly patterns in the time-frequency domain, and pre-trained features provide rich and general audio representations, the model can perform multi-dimensional, micro-to-macro audio quality detection based on these three features, thereby significantly improving the accuracy of the audio quality detection results.

[0129] It should be noted that the audio quality detection device provided in the above embodiments is only illustrated by the division of the above functional modules when detecting audio quality. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.

[0130] The functional modules in the above embodiments can be integrated into one processing unit, or each functional module can exist as a separate physical processing unit, or two or more functional modules can be integrated into one processing unit. The processing unit can be implemented in hardware or software. Furthermore, the specific names of the functional modules are only for easy differentiation and are not intended to limit the scope of protection of the embodiments of this application.

[0131] The audio quality detection device and audio quality detection method provided in the above embodiments belong to the same concept. The specific working process and technical effects of the functional modules in the above embodiments can be found in the method embodiments section, and will not be repeated here.

[0132] Figure 4 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Figure 4 As shown, the computer device 40 includes a processor 400, a memory 401, and a computer program 402 stored in the memory 401 and executable on the processor 400. When the processor 400 executes the computer program 402, it implements the steps in the audio quality detection method in the above embodiments.

[0133] Computer device 40 can be a general-purpose computer device or a special-purpose computer device. In specific implementations, computer device 40 can be a desktop computer, portable computer, network server, handheld computer, mobile phone, tablet computer, wireless terminal device, communication device, or embedded device. This application embodiment does not limit the type of computer device 40. Those skilled in the art will understand that... Figure 4The computer device 40 is merely an example and does not constitute a limitation on the computer device 40. It may include more or fewer components than shown, or combine certain components, or different components, such as input / output devices, network access devices, etc.

[0134] Processor 400 can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor.

[0135] In some embodiments, memory 401 may be an internal storage unit of computer device 40, such as a hard disk or memory of computer device 40. In other embodiments, memory 401 may be an external storage device of computer device 40, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., provided on computer device 40. Furthermore, memory 401 may include both internal and external storage units of computer device 40. Memory 401 is used to store operating system, applications, boot loader, data, and other programs. Memory 401 may also be used to temporarily store data that has been output or will be output.

[0136] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0137] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0138] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the steps in the various method embodiments described above.

[0139] This application provides a computer program product that, when run on a computer, causes the computer to perform the steps described in the various method embodiments above.

[0140] 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, all or part of the processes in the above method embodiments of this application can be implemented by a computer program. This computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or some intermediate form. The computer-readable storage medium can include at least: any entity or device capable of carrying computer program code to a computer device, recording media, computer memory, read-only memory (ROM), random access memory (RAM), compact disc read-only memory (CD-ROM), magnetic tape, floppy disk, and optical data storage devices. The computer-readable storage medium mentioned in this application can be a non-volatile storage medium; in other words, it can be a non-transient storage medium.

[0141] It should be understood that all or part of the steps of the above embodiments can be implemented by software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented in whole or in part as a computer program product. The computer program product includes one or more computer instructions. The computer instructions can be stored in the above-described computer-readable storage medium.

[0142] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0143] In the embodiments provided in this application, it should be understood that the disclosed apparatus / computer devices and methods can be implemented in other ways. For example, the apparatus / computer device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms. Units described as separate components may or may not be physically separate, and components shown as units may or may not be physical units, i.e., 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 application according to actual needs.

[0144] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0145] The embodiments described above are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. An audio quality detection method, characterized in that, The method includes: Acquire the target audio signal; Feature extraction is performed on the target audio signal to obtain handcrafted features, deep audio features, and pre-trained audio features; The handcrafted features, the deep audio features, and the pre-trained audio features are fused together, and the audio quality detection result is determined based on the fused features.

2. The method as described in claim 1, characterized in that, The feature extraction process for the target audio signal, yielding handcrafted features, deep audio features, and pre-trained audio features, includes: Based on the target audio signal, one or more acoustic data are determined, including one or more of total harmonic distortion data, spectral flatness, spectral centroid, spectral roll-off point, and zero-crossing rate; the handcrafted features are generated based on the one or more acoustic data. The target audio signal is subjected to time-frequency feature extraction to obtain the depth audio features; The target audio signal is input into a pre-trained audio model to obtain the pre-trained audio features output by the pre-trained audio model.

3. The method as described in claim 2, characterized in that, The step of extracting time-frequency features from the target audio signal to obtain the depth audio features includes: A time-frequency representation matrix is ​​generated based on the target audio signal, the time-frequency representation matrix including a time dimension and a frequency dimension; The time-frequency representation matrix is ​​subjected to convolutional neural network (CNN) feature extraction to obtain the first time-frequency feature; The depth audio features are generated based on the first time-frequency features.

4. The method as described in claim 3, characterized in that, The generation of the depth audio features based on the first time-frequency features includes: Based on the first time-frequency feature, a time attention weight and / or a frequency attention weight are determined. The time attention weight is used to characterize the probability that a time frame in the target audio signal contains noise, and the frequency attention weight is used to characterize the energy intensity of noise in the frequency band of the target audio signal. Based on the time attention weight and / or the frequency attention weight, the first time-frequency feature is weighted to obtain the second time-frequency feature; The second time-frequency feature is pooled to obtain the depth audio feature.

5. The method according to any one of claims 1 to 4, characterized in that, The process of fusing the handcrafted features, the deep audio features, and the pre-trained audio features, and determining the audio quality detection result based on the fused features, includes: The handcrafted features, the deep audio features, and the pre-trained audio features are fused to obtain the first audio feature; The audio quality detection result is determined based on the first audio feature, and the audio quality detection result includes one or more of the following: noise detection result and noise severity.

6. The method as described in claim 5, characterized in that, The process of fusing the handcrafted features, the deep audio features, and the pre-trained audio features to obtain the first audio feature includes: Based on the handcrafted features, the deep audio features, and the pre-trained audio features, a gating mechanism is used to determine the first weight corresponding to the handcrafted features, the second weight corresponding to the deep audio features, and the third weight corresponding to the pre-trained audio features. Based on the first weight, the second weight, and the third weight, the handcrafted features, the deep audio features, and the pre-trained audio features are weighted and fused to obtain the first audio feature.

7. The method as described in claim 5, characterized in that, The process of fusing the handcrafted features, the deep audio features, and the pre-trained audio features to obtain the first audio feature includes: The handcrafted features, the deep audio features, and the pre-trained audio features are concatenated to obtain the second audio feature; Based on the second audio feature, the fourth weight corresponding to the handcrafted feature, the fifth weight corresponding to the deep audio feature, and the sixth weight corresponding to the pre-trained audio feature are determined through an attention mechanism; Based on the fourth weight, the fifth weight, and the sixth weight, the handcrafted features, the deep audio features, and the pre-trained audio features are weighted and fused to obtain the first audio feature.

8. A computer device, characterized in that, The computer device includes a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the computer program, when executed by the processor, implements the method as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, When the computer program product is run on a computer device, the computer device performs the method as described in any one of claims 1 to 7.