A method, device, electronic equipment, medium and product for audio quality detection
By collecting and slicing audio data and combining it with a multi-model cascaded decision-making mechanism, the problems of low efficiency and poor consistency in audio quality detection are solved, achieving efficient and accurate audio quality detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- UNISOC CHONGQING TECH CO LTD
- Filing Date
- 2026-02-10
- Publication Date
- 2026-06-09
AI Technical Summary
Existing technologies suffer from low efficiency and poor consistency in audio quality detection, while subjective scoring by human listeners also suffers from low efficiency and poor consistency.
By collecting initial audio data, identifying effective signal segments and performing targeted slicing, and combining a multi-model cascaded decision-making mechanism, audio quality detection is performed using a trained single-tone detection model and an audio scoring model, replacing manual subjective scoring.
It improves the utilization and standardization of audio data, enhances the efficiency and consistency of audio quality testing, and adapts to the efficiency and accuracy requirements of smart hardware and real-time communication scenarios.
Smart Images
Figure CN122177157A_ABST
Abstract
Description
Technical Field
[0001] This disclosure generally relates to the field of data detection technology, and more particularly to a method, apparatus, electronic device, medium, and product for audio quality detection. Background Technology
[0002] With the rapid development of smart hardware (such as TWS earphones and smart speakers) and real-time communication technologies (such as VoIP and video conferencing), audio quality testing has become a core element in ensuring user experience.
[0003] In related technologies, subjective scoring by human listeners (e.g., according to the ITU-T P.800 standard) is usually used, which has problems such as low efficiency and poor consistency (the difference in MOS scores between different evaluators can be as high as ±0.8). Summary of the Invention
[0004] This disclosure addresses some of the deficiencies mentioned in the background art by providing a method, apparatus, electronic device, medium, and product for audio quality detection.
[0005] In a first aspect, embodiments of this disclosure provide a method for audio quality detection, comprising:
[0006] Acquire initial audio data and determine the valid signal segments of the initial audio data; The initial audio data is sliced based on the effective signal segment to obtain a target audio slice that meets the preset slicing conditions; wherein, the preset slicing conditions are that the effective signal segment will not be cut and assigned to different target audio slices and the audio slice meets the preset audio slice duration. The target audio slice is processed by a multi-model cascaded decision mechanism to obtain the quality detection result of the initial audio data; wherein, the multi-model cascaded decision mechanism includes a trained single-tone detection model and a trained audio scoring model.
[0007] In one embodiment of the first aspect, the multi-model cascaded decision mechanism is trained in the following manner: Obtain initial test audio data; Multiple types of noise are injected into the initial test audio data using multimodal parametric noise injection technology to obtain the target test audio data; Noise annotation is performed on the target test audio data to obtain a training sample set; The single-tone detection model and the audio scoring model are trained using the training sample set to obtain the trained single-tone detection model and the trained audio scoring model. The trained single-tone detection model and the trained audio scoring model are identified as the multi-model cascade decision mechanism.
[0008] In one embodiment of the first aspect, the step of noise annotation based on the target test audio data to obtain a training sample set includes: Determine the noise information in each of the target test audios; wherein the noise information includes noise intensity, frequency of noise occurrence, and type of noise; The target test audio data is labeled based on the noise information to obtain the training sample set.
[0009] In one embodiment of the first aspect, the step of slicing the initial audio data based on the effective signal segment to obtain a target audio slice that meets preset slicing conditions includes: The starting position of the effective signal segment of the initial audio data is determined as the starting point of the slicing process; A preset audio slice duration is determined, and the initial audio data is sliced based on the preset audio slice duration and the starting point to obtain a target audio slice that meets the preset slicing conditions.
[0010] In one embodiment of the first aspect, determining a preset target audio slice duration and performing slice processing on the initial audio data based on the preset target audio slice duration and the starting point to obtain a target audio slice that meets the preset slice conditions includes: Determine the preset target audio slice duration, and slice the initial audio data based on the preset target audio slice duration and the starting point to obtain the initial audio slice; Determine the slice boundaries of the initial audio slice; If the slice boundary is located within the silent signal segment of the initial audio data, the slice boundary is adjusted to the start point of the silent signal segment, or the end point of the silent signal segment, to obtain the target audio slice.
[0011] In one embodiment of the first aspect, the step of determining a preset target audio slice duration and performing slice processing on the initial audio data based on the preset target audio slice duration and the starting point to obtain a target audio slice that meets the preset slice conditions further includes: If the slice boundary is within the valid signal segment of the initial audio data, the valid signal segment corresponding to the slice boundary is extracted to obtain the target valid signal segment; Zero-padding is performed on the target valid signal segment to obtain a target audio slice that meets the preset slicing conditions.
[0012] In a second aspect, embodiments of this disclosure provide an audio quality detection apparatus, comprising: The acquisition unit is used to acquire initial audio data and determine the valid signal segments of the initial audio data; A slicing unit is used to slice the initial audio data based on the effective signal segment to obtain a target audio slice that meets preset slicing conditions; wherein, the preset slicing conditions are that the effective signal segment will not be cut and assigned to different target audio slices and the audio slice meets the preset audio slice duration. The processing unit is used to process the target audio slice through a multi-model cascaded decision mechanism to obtain the quality detection result of the initial audio data; wherein, the multi-model cascaded decision mechanism includes a trained single-tone detection model and a trained audio scoring model.
[0013] In a third aspect, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement a method for audio quality detection.
[0014] In a fourth aspect, a computer-readable storage medium is provided having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of a method for audio quality detection.
[0015] In a fifth aspect, a computer program product is provided, including a computer program / instructions that, when executed by a processor, implement the steps of a method for audio quality detection.
[0016] As will be described in detail below, an audio quality detection method, apparatus, electronic device, medium, and product according to embodiments of this disclosure are disclosed. By acquiring initial audio data and identifying valid signal segments, combined with directional slicing logic (ensuring that valid signal segments are not split and meet preset durations), the omission or missegmentation of valid audio features in manual processing is avoided, thereby improving the utilization rate and standardization of audio data.
[0017] By leveraging a multi-model cascaded decision-making mechanism, a trained single-tone detection model is first used to quickly screen for audio anomalies, and then a trained audio scoring model is used to perform quality assessment, replacing the method of manual subjective scoring. This not only solves the problem of low efficiency of manual scoring (it can process audio data in batches, with a processing speed far exceeding that of manual scoring), but also eliminates the subjective differences between different evaluators, thereby improving the consistency of quality detection results. It effectively meets the requirements of high efficiency and accuracy for audio quality detection in smart hardware and real-time communication scenarios. Attached Figure Description
[0018] Figure 1 A flowchart of an audio quality detection method provided in this disclosure embodiment; Figure 2A framework diagram of an audio quality detection system provided in this disclosure embodiment; Figure 3 A schematic diagram of an audio quality detection device provided in an embodiment of this disclosure; Figure 4 This is a schematic diagram of an electronic device provided in an embodiment of the present disclosure. Detailed Implementation
[0019] The present disclosure will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present disclosure and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the present disclosure are shown in the drawings, not the entire structure.
[0020] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.
[0021] In this document, the term "and / or" merely describes a relationship, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.
[0022] Research has found that with the rapid development of smart hardware (such as TWS earphones and smart speakers) and real-time communication technologies (such as VoIP and video conferencing), audio quality testing has become a core element in ensuring user experience.
[0023] In related technologies, subjective scoring by human listeners (e.g., according to the ITU-T P.800 standard) is usually used, which has problems such as low efficiency and poor consistency (the difference in MOS scores between different evaluators can be as high as ±0.8).
[0024] Based on the above research, this disclosure provides an audio quality detection method. By collecting initial audio data and identifying effective signal segments, combined with directional slicing logic (ensuring that effective signal segments are not split and meet the preset duration), the method avoids the omission or missegmentation of effective audio features in manual processing, thereby improving the utilization rate and standardization of audio data.
[0025] By leveraging a multi-model cascaded decision-making mechanism, a trained single-tone detection model is first used to quickly screen for audio anomalies, and then a trained audio scoring model is used to perform quality assessment, replacing the method of manual subjective scoring. This not only solves the problem of low efficiency of manual scoring (it can process audio data in batches, with a processing speed far exceeding that of manual scoring), but also eliminates the subjective differences between different evaluators, thereby improving the consistency of quality detection results. It effectively meets the requirements of high efficiency and accuracy for audio quality detection in smart hardware and real-time communication scenarios.
[0026] To facilitate understanding of this embodiment, a method for audio quality detection disclosed in this disclosure will first be described in detail. The execution entity of the audio quality detection method provided in this disclosure is generally a computer device with certain computing capabilities, such as a terminal device, a server, or other processing devices. In some possible implementations, the audio quality detection method can be implemented by a processor calling computer-readable instructions stored in memory.
[0027] See Figure 1 The diagram shows a flowchart of an audio quality detection method provided in this embodiment of the present disclosure. The method includes steps S101 to S103, wherein: S101. Acquire initial audio data and determine the valid signal segment of the initial audio data.
[0028] In the embodiments of this disclosure, initial audio data can be obtained through collaborative acquisition by multiple devices. Examples include an artificial head, a directional external microphone, a sound card, and a microphone from a mobile terminal device.
[0029] Among them, the artificial head can simulate human hearing and collect audio output from headphones / speakers to obtain initial audio data; the directional external microphone can collect sound emitted by devices in the environment to obtain initial audio data; the sound card can directly collect audio from headphones / Bluetooth channels to obtain initial audio data; and the microphone of the mobile terminal device can assist in collecting scene audio to obtain initial audio data.
[0030] Here, audio data can be collected during voice calls (acquiring microphone input + speaker output), music playback (acquiring headphone / Bluetooth output), and recording playback (acquiring playback audio after recording by the device) to obtain initial audio data.
[0031] Here, you can control the audio playback device and the audio acquisition device to start synchronously to obtain the initial audio data.
[0032] Here, the output volume of the target device can be automatically detected before acquisition: if the volume is too low (i.e., amplitude < 10% of full scale), the acquisition gain is automatically increased; if the volume is too high (amplitude > 90% of full scale), the gain is automatically reduced to avoid audio clipping distortion.
[0033] After acquiring the initial audio data, the initial audio data acquired by each device can be parsed to determine the encoding type of the initial audio data.
[0034] Then, the initial audio data of different encoding types can be converted into a unified format based on the encoding type.
[0035] Then, the initial audio data, which has been converted to a uniform format, can be normalized by sampling rate to obtain intermediate audio data.
[0036] Here, a combination of linear interpolation and low-pass filtering algorithms can be used to standardize the sampling rate of the initial audio data in a uniform format to obtain intermediate audio data. Specifically, the initial audio data in a uniform format can first be low-pass filtered to remove audio signals exceeding the target sampling rate; then, a linear interpolation algorithm is used to supplement sampling points, converting the original sampling rate to the target sampling rate to obtain the intermediate audio data.
[0037] Next, the intermediate audio data can be processed to unify the channels, thus obtaining the target audio data. Here, the channel number parameter of each intermediate audio data can be determined. For example, a channel number parameter of 1 indicates that the intermediate audio data is mono, and a channel number parameter of 2 indicates that the intermediate audio data is stereo.
[0038] Here, when the intermediate audio data is determined to be stereo based on the number of channels, the stereo (i.e., two-channel) audio is converted to mono by using the left and right channel amplitude mean fusion technology. For example, mono amplitude = (left channel amplitude + right channel amplitude) / 2.
[0039] After obtaining the target audio data, the energy corresponding to each frame of the target audio data can be determined. Then, based on the energy of each frame of the target audio data, the mean and standard deviation of the energy can be calculated.
[0040] Then, a dynamic threshold can be obtained, and the dynamic threshold can be adjusted based on the above energy mean and standard deviation to obtain the target threshold.
[0041] After determining the target threshold, frames in the target audio data that are greater than the target threshold can be identified as valid signals, and frames in the target audio data that are less than or equal to the target threshold can be identified as silence signals.
[0042] For example, if the average energy value of the target audio data from 0 to 3 seconds is greater than the target threshold, then 0 to 3 seconds is determined as a valid signal segment; if the average energy value of the target audio data from 5 to 7 seconds is less than or equal to the target threshold, then 5 to 7 seconds is determined as a silent signal segment.
[0043] S102. Based on the effective signal segments, the initial audio data is sliced to obtain target audio slices that meet the preset slicing conditions; wherein, the preset slicing conditions are that the effective signal segments will not be cut and assigned to different target audio slices and that the audio slices meet the preset audio slice duration.
[0044] In the embodiments of this disclosure, the target audio data can be sliced based on the effective signal segment and the silence signal segment to obtain a target audio slice that meets the preset slicing conditions.
[0045] Each target audio slice has the same length, which is the preset audio slice duration.
[0046] First, silent segments can be removed from the target audio data. Then, the target audio data from which silent segments have been removed is sliced.
[0047] Among them, the same valid signal segments are in the same target audio slice. For example, with a preset audio slice duration of 10 seconds, the slice is started from the beginning of the valid signal segment. If the end of the slice is still within the valid signal segment, it is cut directly; if the end is close to the end of the valid signal segment (remaining length < 3 seconds), the "tail priority strategy" is triggered to obtain the target audio slice.
[0048] S103. The target audio slice is processed through a multi-model cascaded decision mechanism to obtain the quality detection results of the initial audio data; wherein, the multi-model cascaded decision mechanism includes a post-trained single-tone detection model and a post-trained audio scoring model.
[0049] In the embodiments of this disclosure, features of the target audio slice can be extracted to obtain target audio features.
[0050] The target audio features can then be input into a multi-model cascaded decision mechanism for processing to obtain the quality detection results of the initial audio data.
[0051] After determining the quality test results of the initial audio data, a visual report can be generated based on these results.
[0052] The visualization report includes overall statistics: anomaly detection rate, average MOS score, and percentage of each quality level; detailed data: slice ID, anomaly information, MOS score, and rating for each audio file; and visualization charts: bar chart of noise type distribution, histogram of MOS score distribution, and cross-device detection accuracy comparison chart.
[0053] In the embodiments of this disclosure, firstly, initial audio data is collected, and valid signal segments of the initial audio data are determined; secondly, the initial audio data is sliced based on the valid signal segments to obtain target audio slices that meet preset slicing conditions; wherein, the preset slicing conditions are that the valid signal segments will not be cut and assigned to different target audio slices and that the audio slices meet the preset audio slice duration; finally, the target audio slices are processed through a multi-model cascaded decision mechanism to obtain the quality detection result of the initial audio data; wherein, the multi-model cascaded decision mechanism includes a trained single-tone detection model and a trained audio scoring model.
[0054] In the above embodiments, by collecting initial audio data and identifying valid signal segments, and combining directional slicing logic (ensuring that valid signal segments are not split and meet the preset duration), the omission or missegmentation of valid audio features in manual processing is avoided, thereby improving the utilization rate and standardization of audio data.
[0055] By leveraging a multi-model cascaded decision-making mechanism, a trained single-tone detection model is first used to quickly screen for audio anomalies, and then a trained audio scoring model is used to perform quality assessment, replacing the method of manual subjective scoring. This not only solves the problem of low efficiency of manual scoring (it can process audio data in batches, with a processing speed far exceeding that of manual scoring), but also eliminates the subjective differences between different evaluators, thereby improving the consistency of quality detection results. It effectively meets the requirements of high efficiency and accuracy for audio quality detection in smart hardware and real-time communication scenarios.
[0056] In one optional implementation, the multi-model cascaded decision-making mechanism is trained as follows: First, obtain the initial test audio data; Secondly, various noises are injected into the initial test audio data using multimodal parametric noise injection technology to obtain the target test audio data; Secondly, noise is labeled based on the target test audio data to obtain the training sample set; Secondly, the single-tone detection model and the audio scoring model are trained using the training sample set to obtain the trained single-tone detection model and the trained audio scoring model. Finally, the trained single-tone detection model and the trained audio scoring model were determined to be a multi-model cascade decision-making mechanism.
[0057] In the embodiments of this disclosure, the initial test audio data can cover clean audio in multiple scenarios such as voice calls, music playback, and recording playback, adapt to different sampling rates (44.1kHz, 48kHz, etc.) and bit depths (16bit, 24bit, etc.), cover the typical output audio characteristics of smart hardware (such as headphones, smart speakers) and real-time communication devices, ensure the comprehensiveness of the data in terms of scenarios and device compatibility, and provide diversified basic data support for subsequent model training.
[0058] Here, multimodal parametric noise injection technology can be used to inject various types of noise into the initial test audio data to obtain the target test audio data.
[0059] Among them, the multimodal parametric noise injection technology supports the precise implantation of noise according to preset types, covering common audio anomaly types such as pop sounds, current sounds, stuttering, background noise, and howling.
[0060] Here, you can configure the noise intensity to suit different noise levels, and you can also set the range of noise occurrence frequency to achieve diversified simulation of noise distribution.
[0061] In the injection process, a time-frequency domain hybrid enhancement algorithm can be used. First, the initial test audio and the target noise are subjected to short-time Fourier transform to convert them into time-frequency domain signals. Then, the noise time-frequency domain amplitude is adjusted according to the preset intensity parameters, and the fusion rules are optimized by combining the physical characteristics of different noise types. Finally, the inverse short-time Fourier transform is used to restore the time-domain audio, realizing the physical-level fusion of noise and original audio signals to obtain the target test audio data.
[0062] Here, noise information can be extracted and determined for each segment of target test audio data. This noise information includes the noise type, noise intensity, frequency of occurrence, and the time and location of the noise occurrence.
[0063] Then, based on this complete noise information, a standardized labeling format can be used to add corresponding labels to each target test audio data, forming a training sample set containing the mapping relationship between "audio data and noise labels".
[0064] Here, the single-tone detection model adopts a 5-layer gated recurrent unit (GRU) network architecture, using the Mel frequency cepstral coefficients (MFCC) and the first-order difference of MFCC (ΔMFCC) of the audio as input features.
[0065] One approach is to use labeled data from the training sample set to train the single-tone detection model to recognize various types of noise, optimize the model parameters to improve the accuracy of noise type judgment and location positioning, and obtain the trained single-tone detection model.
[0066] Here, the audio scoring model is based on a CNN-BiLSTM hybrid architecture, which is optimized on the basis of the traditional NISQA model. It incorporates features such as noise type and device type, and is fine-tuned using the training sample set to improve the audio scoring model's ability to evaluate audio quality in different scenarios and on different devices, ensuring that the output MOS score is correlated with the actual audio quality.
[0067] Here, the collaborative working logic of the two models (i.e., the trained single-tone detection model and the trained audio scoring model) can be clearly defined. The single-tone detection model prioritizes anomaly screening of the audio, while the audio scoring model scores the quality of the results output by the single-tone detection model.
[0068] The above implementation forms a cascaded processing flow of "anomaly detection - quality assessment". At the same time, it configures a weight allocation rule based on the confidence level of noise type to ensure that the detection results output by multiple models are more accurate and reliable, providing core algorithm support for subsequent audio quality detection.
[0069] In one optional implementation, noise is labeled based on the target test audio data to obtain a training sample set, specifically including the following steps: First, determine the noise information in each target test audio; the noise information includes noise intensity, frequency of noise occurrence, and type of noise. Then, the target test audio data is labeled based on noise information to obtain the training sample set.
[0070] In the embodiments of this disclosure, the noise type can be identified by combining the preset configuration of multimodal parameterized noise injection with audio signal analysis, the noise intensity can be quantified by the signal-to-noise ratio (SNR), the number of times the noise occurs can be counted, and its time location can be located.
[0071] Here, a unique identifier can be assigned to each target test audio segment, integrating noise type, SNR value, frequency of occurrence, and time location into a structured label, establishing a one-to-one mapping between "audio data - noise label". After labeling, the sample set is categorized and organized according to noise type and intensity level to obtain the training sample set.
[0072] In an optional implementation, the initial audio data is sliced based on the effective signal segments to obtain target audio slices that meet preset slicing conditions, specifically including the following steps: First, the starting position of the effective signal segment of the initial audio data is determined as the starting point of the slicing process; Then, the preset audio slice duration is determined, and the initial audio data is sliced based on the preset audio slice duration and starting point to obtain the target audio slice that meets the preset slicing conditions.
[0073] In the embodiments of this disclosure, effective signal segments can be identified by a dynamic energy threshold algorithm, that is, by calculating the energy of each frame of audio, statistically analyzing the energy distribution, and setting an adaptive threshold, continuous signal segments (such as speech, music, abnormal noise, etc.) with energy higher than the threshold are selected.
[0074] Then, by using the starting position of the effective signal segment as the starting point of the slice, we can ensure that each slice starts from the effective signal, avoid the slice starting from the silent segment and generating invalid data, and provide a high-quality signal foundation for subsequent model analysis.
[0075] Here, starting from the beginning of the effective signal segment, the initial audio data is sliced sequentially according to the preset audio slice duration. This ensures that the same effective signal segment is not cut and assigned to different target audio slices, and that each target audio slice meets the preset duration requirement, thus guaranteeing the consistency and effectiveness of subsequent model analysis.
[0076] In one optional implementation, a preset target audio slice duration is determined, and the initial audio data is sliced based on the preset target audio slice duration and the starting point to obtain a target audio slice that meets the preset slicing conditions. Specifically, this includes the following steps: First, determine the preset target audio slice duration, and then slice the initial audio data based on the preset target audio slice duration and starting point to obtain the initial audio slice; Then, determine the slice boundaries of the initial audio slice; Finally, if the slice boundary is located at the silent signal segment of the initial audio data, the slice boundary is adjusted to the start point or the end point of the silent signal segment to obtain the target audio slice.
[0077] In the embodiments of this disclosure, the start and end times of each initial audio slice can be determined by parsing the time-series segmentation results after slice processing, and the end time is the slice boundary.
[0078] For example, starting with 0 seconds, the slice boundary of the first initial audio slice is 10 seconds, the second slice is 20 seconds, and so on, to ensure the timing accuracy of the slice boundaries.
[0079] Here, silent signal segments are identified using a dynamic energy threshold algorithm, which identifies continuous signal segments with energy below an adaptive threshold. During adjustment, the principle of "reducing silence redundancy" is prioritized. If a slice boundary falls within a silent signal segment, it is adjusted to the start or end point of that segment to prevent the silent segment from being split into two target audio slices.
[0080] Here, the silent signal segment after the slice boundary can also be discarded. For example, if the slice boundary is 10s and the silent signal segment is 11-15s, the silent signal segment can be discarded directly, and 16s can be used as the starting point of the next initial audio slice.
[0081] In an optional implementation, a preset target audio slice duration is determined, and the initial audio data is sliced based on the preset target audio slice duration and the starting point to obtain a target audio slice that meets the preset slicing conditions. The implementation also includes the following steps: First, based on the fact that the slice boundary is within the valid signal segment of the initial audio data, the valid signal segment corresponding to the slice boundary is extracted to obtain the target valid signal segment; Then, zero-padding is performed on the target valid signal segment to obtain the target audio slice that meets the preset slicing conditions.
[0082] In the embodiments of this disclosure, effective signal segments can be identified by a dynamic energy threshold algorithm, that is, continuous signal segments with energy higher than an adaptive threshold, which contain valuable audio information such as speech, music, and abnormal noise.
[0083] In cases where the slice boundary falls within the effective signal segment, it indicates that the effective signal segment may be split by the slice boundary. In this case, the split effective signal segment must be extracted completely to ensure that its timing characteristics and abnormal noise (such as pop tones and current tones) are not truncated. The extracted complete effective signal segment is the target effective signal segment.
[0084] Here, the preset slicing conditions require the target audio slice to meet the preset audio slice duration. If the duration of the extracted target valid signal segment does not meet the preset audio slice duration, zero padding is performed at the end of the target valid signal segment. The zero-padding part is a silent signal and does not change the original characteristics of the target valid signal segment.
[0085] Here, zero padding can be used to adapt the target effective signal segment to the preset duration requirement, forming a standardized target audio slice, ensuring that the subsequent multi-model cascaded decision-making mechanism can be input normally and analyzed accurately.
[0086] Based on the same inventive concept, this disclosure also provides an audio quality detection system corresponding to the audio quality detection method. Since the principle of the system in this disclosure is similar to the audio quality detection method described above, the implementation of the system can refer to the implementation of the method, and the repeated parts will not be described again.
[0087] Reference Figure 2 The diagram shown is a framework diagram of an audio quality detection system provided in this embodiment of the present disclosure, including: an audio acquisition subsystem and an audio detection engine, wherein: The audio acquisition subsystem includes a real-time monitoring module and an offline monitoring module; the real-time monitoring module is used to acquire initial audio data in real time, and the offline monitoring module is used to acquire initial audio data for each channel.
[0088] Here, the offline monitoring module corresponds to pathways including headphone pathway, speaker pathway, Bluetooth pathway, etc.
[0089] The audio detection engine includes: an algorithm module, an audio data preprocessing module, a multi-model joint analysis module, and an abnormal audio generation module.
[0090] The algorithm module provides algorithmic support for the multi-model joint analysis module. This module includes various machine learning and deep learning algorithms.
[0091] The audio data preprocessing module is used to process the acquired initial audio to obtain the target audio slice, and then input the target audio slice into the multi-model joint analysis module.
[0092] The multi-model joint analysis module includes an audio anomaly detection submodule and a speech quality assessment submodule. First, the target audio slice is input into the audio anomaly detection submodule. Then, the trained single-tone detection model set up in the audio anomaly detection submodule processes the target audio slice to obtain the anomaly detection results of the initial audio data.
[0093] Then, the target audio slice is processed after training using the speech quality assessment submodule to obtain the quality assessment result; subsequently, the anomaly detection result and the quality assessment result are determined as the quality detection result of the initial audio data.
[0094] Here, the abnormal audio generation module is used to inject various noises into the initial test audio data to obtain the target test audio data. Among them, the target test audio data corresponding to music playback and audio peripherals can be used as training samples for the single-tone detection model; the target test audio data corresponding to recording playback and voice calls can be used as training samples for the audio scoring model.
[0095] Based on the same inventive concept, this disclosure also provides an audio quality detection apparatus corresponding to the audio quality detection method. Since the principle of the apparatus in this disclosure for solving the problem is similar to the audio quality detection method described above in this disclosure, the implementation of the apparatus can refer to the implementation of the method, and the repeated parts will not be described again.
[0096] Reference Figure 3 The diagram shown is a schematic representation of an audio quality detection device provided in an embodiment of this disclosure, comprising: a data acquisition unit 31, a slicing unit 32, and a processing unit 33; wherein: The acquisition unit is used to acquire initial audio data and determine the valid signal segments of the initial audio data; A slicing unit is used to slice the initial audio data based on the effective signal segment to obtain a target audio slice that meets preset slicing conditions; wherein, the preset slicing conditions are that the effective signal segment will not be cut and assigned to different target audio slices and the audio slice meets the preset audio slice duration. The processing unit is used to process the target audio slice through a multi-model cascaded decision mechanism to obtain the quality detection result of the initial audio data; wherein, the multi-model cascaded decision mechanism includes a trained single-tone detection model and a trained audio scoring model.
[0097] This embodiment collects initial audio data and identifies valid signal segments. Combined with directional slicing logic (ensuring that valid signal segments are not split and meet the preset duration), it avoids the omission or missegmentation of valid audio features in manual processing, thereby improving the utilization rate and standardization of audio data.
[0098] By leveraging a multi-model cascaded decision-making mechanism, a trained single-tone detection model is first used to quickly screen for audio anomalies, and then a trained audio scoring model is used to perform quality assessment, replacing the method of manual subjective scoring. This not only solves the problem of low efficiency of manual scoring (it can process audio data in batches, with a processing speed far exceeding that of manual scoring), but also eliminates the subjective differences between different evaluators, thereby improving the consistency of quality detection results. It effectively meets the requirements of high efficiency and accuracy for audio quality detection in smart hardware and real-time communication scenarios.
[0099] The processing flow of each module in the device and the interaction flow between each module can be referred to the relevant descriptions in the above method embodiments, and will not be detailed here.
[0100] Corresponding to Figure 1 In addition to the data processing methods described in this disclosure, an electronic device 400 is also provided, such as... Figure 4 The diagram shown is a structural schematic of an electronic device 400 provided in an embodiment of this disclosure, including: The system includes a processor 41, a memory 42, and a bus 43. The memory 42 stores execution instructions and includes main memory 421 and external memory 422. The main memory 421, also called internal memory, temporarily stores the computational data in the processor 41, as well as data exchanged with external memory such as a hard disk. The processor 41 exchanges data with the external memory 422 through the main memory 421. When the electronic device 400 is running, the processor 41 communicates with the memory 42 through the bus 43, causing the processor 41 to execute the following instructions: Acquire initial audio data and determine the valid signal segments of the initial audio data; The initial audio data is sliced based on the effective signal segment to obtain a target audio slice that meets the preset slicing conditions; wherein, the preset slicing conditions are that the effective signal segment will not be cut and assigned to different target audio slices and the audio slice meets the preset audio slice duration. The target audio slice is processed by a multi-model cascaded decision mechanism to obtain the quality detection result of the initial audio data; wherein, the multi-model cascaded decision mechanism includes a trained single-tone detection model and a trained audio scoring model.
[0101] The basic principles of this disclosure have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this disclosure are merely examples and not limitations, and should not be considered as essential features of each embodiment of this disclosure. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the scope of this disclosure to the necessity of employing the aforementioned specific details for implementation.
[0102] The block diagrams of devices, apparatuses, devices, and systems disclosed herein are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as “comprising,” “including,” “having,” etc., are open-ended terms meaning “including but not limited to,” and are used interchangeably with them. The terms “or” and “and” as used herein refer to the terms “and / or,” and are used interchangeably with them unless the context clearly indicates otherwise. The term “such as” as used herein refers to the phrase “such as but not limited to,” and is used interchangeably with it.
[0103] Additionally, as used herein, the "or" used in a list of items beginning with "at least one" indicates a separate list, such that a list of, for example, "at least one of A, B, or C" means A or B or C, or AB or AC or BC, or ABC (i.e., A and B and C). Furthermore, the word "exemplary" does not imply that the described example is preferred or better than other examples.
[0104] It should also be noted that in the systems and methods of this disclosure, the components or steps can be decomposed and / or recombined. These decompositions and / or recombinations should be considered as equivalent solutions to this disclosure.
[0105] Various changes, substitutions, and modifications can be made to the technology described herein without departing from the teachings defined by the appended claims. Furthermore, the scope of the claims of this disclosure is not limited to the specific aspects of the processes, machines, manufactures, events, means, methods, and actions described above. Currently existing or later-developed processes, machines, manufactures, events, means, methods, or actions that perform substantially the same function or achieve substantially the same result as the corresponding aspects described herein can be utilized. Therefore, the appended claims include such processes, machines, manufactures, events, means, methods, or actions within their scope.
[0106] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this disclosure. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of this disclosure. Therefore, this disclosure is not intended to be limited to the aspects shown herein, but rather to be carried out within the widest scope consistent with the principles and novel features disclosed herein.
[0107] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this disclosure to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations therein.
Claims
1. A method for audio quality detection, characterized in that, include: Acquire initial audio data and determine the valid signal segments of the initial audio data; The initial audio data is sliced based on the effective signal segment to obtain a target audio slice that meets the preset slicing conditions; wherein, the preset slicing conditions are that the effective signal segment will not be cut and assigned to different target audio slices and the audio slice meets the preset audio slice duration. The target audio slice is processed by a multi-model cascaded decision mechanism to obtain the quality detection result of the initial audio data; wherein, the multi-model cascaded decision mechanism includes a trained single-tone detection model and a trained audio scoring model.
2. The method as described in claim 1, characterized in that, The multi-model cascaded decision-making mechanism is trained in the following manner: Obtain initial test audio data; Multiple types of noise are injected into the initial test audio data using multimodal parametric noise injection technology to obtain the target test audio data; Noise annotation is performed on the target test audio data to obtain a training sample set; The single-tone detection model and the audio scoring model are trained using the training sample set to obtain the trained single-tone detection model and the trained audio scoring model. The trained single-tone detection model and the trained audio scoring model are identified as the multi-model cascade decision mechanism.
3. The method as described in claim 2, characterized in that, The training sample set is obtained by noise annotation based on the target test audio data, including: Determine the noise information in each of the target test audios; wherein the noise information includes noise intensity, frequency of noise occurrence, and type of noise; The target test audio data is labeled based on the noise information to obtain the training sample set.
4. The method as described in claim 1, characterized in that, The step of slicing the initial audio data based on the effective signal segment to obtain a target audio slice that meets preset slicing conditions includes: The starting position of the effective signal segment of the initial audio data is determined as the starting point of the slicing process; A preset audio slice duration is determined, and the initial audio data is sliced based on the preset audio slice duration and the starting point to obtain a target audio slice that meets the preset slicing conditions.
5. The method as described in claim 4, characterized in that, The step of determining the preset target audio slice duration and performing slice processing on the initial audio data based on the preset target audio slice duration and the starting point to obtain a target audio slice that meets the preset slice conditions includes: Determine the preset target audio slice duration, and slice the initial audio data based on the preset target audio slice duration and the starting point to obtain the initial audio slice; Determine the slice boundaries of the initial audio slice; If the slice boundary is located within the silent signal segment of the initial audio data, the slice boundary is adjusted to the start point of the silent signal segment, or the end point of the silent signal segment, to obtain the target audio slice.
6. The method as described in claim 5, characterized in that, The step of determining a preset target audio slice duration and performing slice processing on the initial audio data based on the preset target audio slice duration and the starting point to obtain a target audio slice that meets the preset slice conditions further includes: If the slice boundary is within the valid signal segment of the initial audio data, the valid signal segment corresponding to the slice boundary is extracted to obtain the target valid signal segment; Zero-padding is performed on the target valid signal segment to obtain a target audio slice that meets the preset slicing conditions.
7. An audio quality detection device, characterized in that, include: The acquisition unit is used to acquire initial audio data and determine the valid signal segments of the initial audio data; A slicing unit is used to slice the initial audio data based on the effective signal segment to obtain a target audio slice that meets preset slicing conditions; wherein, the preset slicing conditions are that the effective signal segment will not be cut and assigned to different target audio slices and the audio slice meets the preset audio slice duration. The processing unit is used to process the target audio slice through a multi-model cascaded decision mechanism to obtain the quality detection result of the initial audio data; wherein, the multi-model cascaded decision mechanism includes a trained single-tone detection model and a trained audio scoring model.
8. An electronic device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method as described in any one of claims 1 to 6.
9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method as described in any one of claims 1 to 6.
10. A computer program product, comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method as described in any one of claims 1 to 6.