Audio quality automatic scoring method and system combining MFCC and time domain statistical characteristics

Abstract

The invention discloses an automatic audio quality scoring method and system combining MFCC and time domain statistical characteristics, and the method comprises the steps: carrying out the segmentation according to seconds on the basis of reference and to-be-detected signal resampling and time alignment, and further extracting cepstrum and time domain statistical characteristics according to 100 ms sub-segments; the method comprises the following steps: forming 170-dimensional second-level features by performing hierarchical aggregation of mean value + maximum value on sub-segment features, splicing the 170-dimensional second-level features with corresponding reference segments to form 340-dimensional joint features, and inputting the 340-dimensional joint features into a pre-trained support vector machine model to output second-by-second discrete scores from 0 to 5, thereby realizing high-time-resolution automatic quality evaluation and anomaly positioning under the conditions of light calculation power and small samples.

Description

Technical Field

[0001] This invention relates to the technical field of audio signal processing and automated detection, and in particular to an automatic audio quality scoring method and system that combines MFCC and time-domain statistical features. Background Technology

[0002] With the rapid iteration of smart audio terminals (mobile phones, headphones, speakers, wearable devices, etc.), audio quality inspection in the manufacturing process has become a key process determining product consistency and user perception. Current common methods include: First, manual listening inspection, where trained quality inspectors identify defects such as distortion, noise, and popping sounds based on experience, but suffer from low efficiency, strong subjectivity, and auditory fatigue; Second, subjective evaluation (MOS), which organizes multiple listeners to score according to the ITU-T P.800 series, and the average value is used as a quality indicator, but it is difficult to automate and has insufficient temporal resolution; Third, objective evaluation (PESQ, ITU-T P.862), which gives a single score based on the perceived difference between the reference and tested audio, but the algorithm is complex, lacks sensitivity to short-term transient anomalies, and has limited adaptability to music and complex sound fields; Fourth, end-to-end scoring based on deep learning (such as CNN / RNN), while having potential accuracy advantages, relies on large-scale labeled data and high computing power, making it difficult to deploy stably in the industrial control computer environment of the production line.

[0003] However, for large-scale production, enterprises have higher requirements for testing: to achieve high temporal resolution scoring with millisecond-level processing latency, to accurately locate the time and type of anomalies, to ensure consistency across batches, and to maintain availability and maintainability under small sample conditions and on general industrial control platforms; in general, existing technical solutions are difficult to balance automation, high efficiency, lightweight and robustness to small samples.

[0004] CN201911250828.1 discloses an audio quality evaluation method that determines the overall quality level by inputting the transmitted audio into a quality evaluation model and combining it with a confidence threshold. It has a certain degree of robustness and scalability, but it mainly outputs the overall level with a coarse temporal granularity, making it difficult to locate transient defects such as pops and clipping second by second. Furthermore, the deep model is not friendly to the production line environment in terms of sample size and computational power dependence.

[0005] CN201710498188.0 discloses a method for objective evaluation of no-reference speech quality based on deep learning speech enhancement. It uses a deep belief network for initial speech enhancement, and then uses the difference in MFCC before and after enhancement to regress the objective score through a BP network. It has good correlation with subjective scores in no-reference conditions and speech scenarios. However, its training and inference have high requirements for data and computing power. It does not optimize for reference-test pair alignment, hierarchical segment aggregation and second-level discrete scoring, and does not solve the problem of interpretable localization of short-term anomalies in music and complex sound source scenarios.

[0006] In summary, existing technologies still have significant shortcomings in terms of temporal resolution, anomaly localization, computational power threshold, and small sample availability. They are still far from meeting the comprehensive requirements of production lines for speed, accuracy, stability, and feasibility. Therefore, existing audio quality evaluation technologies still suffer from problems such as coarse temporal granularity, easy omission of transient anomalies, strong dependence on computational power and samples, and difficulty in lightweight deployment on industrial control computers. Summary of the Invention

[0007] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section, as well as in the abstract and title of the present application, to avoid obscuring the purpose of this section, the abstract and title of the invention. Such simplifications or omissions shall not be used to limit the scope of the present invention.

[0008] In view of the aforementioned existing problems, the present invention is proposed.

[0009] To solve the above-mentioned technical problems, the present invention provides the following technical solution: receiving a reference audio signal and a test audio signal, resampling the two audio signals to 16000Hz, and performing time alignment; The time-aligned audio is segmented by second, and each segment is further divided into 10 sub-segments; Extract features from each 100-millisecond segment to obtain the segment feature vector; For each 1-second audio segment, the mean and maximum value of the 10 sub-segments feature vectors are calculated for each dimension, resulting in 85-dimensional mean features and 85-dimensional maximum features, which are then concatenated to form a 170-dimensional second-level feature vector. The second-level feature vectors of the reference audio and the audio to be tested for the corresponding time period are concatenated to form a 340-dimensional joint feature vector; The joint feature vector is input into a pre-trained support vector machine (SVM) model, which outputs an integer score in the range of 0 to 5 to represent the degree of distortion of the audio at that second.

[0010] As a preferred embodiment of the automatic audio quality scoring method combining MFCC and temporal statistical features described in this invention, the feature extraction for each 100-millisecond segment includes: Extract the 40-dimensional frequency cepstral coefficients (MFCCs) and calculate their mean and standard deviation to obtain 80-dimensional frequency domain features; Calculate five time-domain statistical characteristics: root mean square energy (RMS), zero-crossing rate (ZCR), maximum amplitude, skewness, and kurtosis. The frequency domain features and the time domain statistical features are concatenated to form an 85-dimensional sub-segment feature vector.

[0011] As a preferred embodiment of the automatic audio quality scoring method combining MFCC and time-domain statistical features described in this invention, the specific process of extracting the Mel frequency cepstral coefficient (MFCC) features includes: Pre-emphasis processing of audio signals; Hamming windows are used for framing, with a frame length of 25 milliseconds and a frame shift of 10 milliseconds. Perform a Fast Fourier Transform on each frame; The Mel spectrum was calculated using a 40-triangular Mel filter bank; After taking the logarithm, a discrete cosine transform is performed to obtain the 40-dimensional MFCC coefficients.

[0012] As a preferred embodiment of the automatic audio quality scoring method combining MFCC and temporal statistical features described in this invention, when calculating the temporal statistical features, RMS, ZCR, maximum amplitude, kurtosis and skewness are multiplied by coefficients of 5, 3, 3, 3 and 3 respectively to enhance the features.

[0013] As a preferred embodiment of the automatic audio quality scoring method combining MFCC and temporal statistical features described in this invention, the SVM model employs a radial basis function (RBF) kernel, with the kernel function taking the form: K(x,y)=exp(-γ||xy|| 2 ) Where γ > 0 is the kernel function parameter, used to control the rate at which similarity decays with distance. A larger γ results in a narrower kernel distribution, making the model more sensitive to local differences, leading to a more complex decision boundary and increased risk of overfitting. A smaller γ results in a wider kernel distribution, slower similarity decay, and a smoother decision boundary. Soft-margin SVM is used during model training, with a penalty coefficient of C = 10, where C is the regularization parameter. K(x,y) is the output of the kernel function, representing the similarity between samples x and y in the implicit high-dimensional feature space. exp(⋅) is an exponential function with the natural constant e as its base, used to map the distance metric to the similarity in the interval (0,1). x and y are two sample vectors to be compared; ||xy|| 2 The squared Euclidean distance between the samples is given.

[0014] As a preferred embodiment of the automatic audio quality scoring method combining MFCC and temporal statistical features described in this invention, the SVM model needs to be trained, specifically including: Collect reference audio, audio to be tested, and corresponding manually annotated scores; Perform a joint feature vector generation process on the training data to obtain the initial joint feature vector; The initial joint feature vector was standardized using StandardScaler; The SVM classification model is trained using the standardized initial joint feature vector and labeled scores. Save the trained model parameters and standardized parameters.

[0015] As a preferred embodiment of the automatic audio quality scoring method combining MFCC and temporal statistical features described in this invention, the scoring level is defined as follows: 0 points, no difference, the reference audio and the audio to be tested are almost identical; 1 point, slight difference, with a brief electrical hum or weak noise present; 2 points, slightly distorted, but acceptable; 3 points, clearly distorted, but still acceptable; 4 points, severely distorted, music is obviously distorted or speech is incomprehensible; 5 points, extremely poor, signal completely lost or extremely poor sound quality.

[0016] As a preferred embodiment of the automatic audio quality scoring system combining MFCC and temporal statistical features described in this invention, it includes: An audio input module is used to receive reference audio and the audio to be tested; The preprocessing module is used for audio resampling and time alignment; The segmentation module is used to segment audio by second and further subdivide it into sub-segments; The feature extraction module is used to extract MFCC features and temporal statistical features; The feature aggregation module is used to calculate the mean and maximum value of the features of the sub-segments; The model inference module is used to load the SVM model and output a score. The results output module is used to generate second-by-second scoring sequences and anomaly location information.

[0017] As a preferred embodiment of the automatic audio quality scoring system combining MFCC and time-domain statistical features described in this invention, the system further includes one or more processors; The memory stores operable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, including the flow of an automatic audio quality scoring method combining MFCC and time-domain statistical features as described above.

[0018] As a preferred embodiment of the computer-readable medium for storing software according to the present invention, the software includes instructions executable by one or more computers, the instructions causing the one or more computers to perform operations, the operations including the flow of an automatic audio quality scoring method combining MFCC and time-domain statistical features as described above.

[0019] The beneficial effects of this invention are as follows: Based on the resampling and time alignment of the reference and test signals, our invention segments the signal by second and further extracts cepstrum and temporal statistical features in 100 ms segments. By performing hierarchical aggregation of the segment features using the mean and maximum values, a 170-dimensional second-level feature is formed, which is then concatenated with the corresponding reference segment to form a 340-dimensional joint feature. This feature is then input into a pre-trained support vector machine model to output a discrete score of 0 to 5 per second, thereby achieving automated quality assessment and anomaly localization with high temporal resolution under conditions of light computing power and small sample size. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart illustrating the automatic audio quality scoring method combining MFCC and temporal statistical features as shown in this invention. Figure 2a This is a schematic diagram of the test results for production line sample a shown in this invention; Figure 2b This is a schematic diagram of the test results for production line sample b as shown in this invention; Figure 2c This is a schematic diagram of the test results for production line sample c as shown in this invention; Figure 2d This is a schematic diagram of the test results for production line sample d as shown in this invention; Figure 3 This is a schematic diagram illustrating the module functions of the automatic audio quality scoring system combining MFCC and time-domain statistical features as shown in this invention. Figure 4 This is a schematic diagram of the module structure distribution of the automatic audio quality scoring system that combines MFCC and time-domain statistical features as shown in this invention. Detailed Implementation

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0022] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort should fall within the scope of protection of this invention.

[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0024] According to an embodiment of the present invention, in combination Figure 1 The flowchart shown illustrates an automatic audio quality scoring method that combines MFCC and temporal statistical features, specifically including the following steps: S1. Receive the reference audio signal and the audio signal to be tested, resample the two audio signals to 16000Hz, and perform time alignment; S2. Divide the time-aligned audio into segments by second, and divide each segment into 10 sub-segments to obtain the original signal sub-segment sampling data: ; S3. For each 100-millisecond segment, calculate the 40-dimensional MFCC mean and 40-dimensional standard deviation, and calculate five time-domain statistics: RMS, zero-crossing rate (ZCR), maximum amplitude, skewness, and kurtosis. These statistics form the 85-dimensional feature vector of the m-th segment in the t-th second. :

[0025] Where t represents the t-th second, and m represents the m-th 100ms segment within that second. The mean of MFCC in each dimension The standard deviation of MFCC in each dimension, with 5 time-domain characteristics as follows: Root mean square energy: Zero crossing rate: Maximum amplitude: Skewness: , The third moment centered, Standard deviation; Kurtosis: , The fourth moment centered, Standard deviation; S4. For the feature vectors of the 10 sub-segments in each 1-second audio segment, calculate the mean and maximum value of each dimension to obtain the 85-dimensional mean feature vectors. and 85-dimensional sequential maximum features These are concatenated to form a 170-dimensional feature vector that operates at a speed of seconds. ; S5. Concatenate the second-level feature vectors of the corresponding time periods of the reference audio A and the audio B to be tested to form a 340-dimensional joint feature vector. ; S6. Input the joint feature vector into the pre-trained support vector machine (SVM) model and output an integer score in the range of 0 to 5 to represent the degree of distortion of the audio in that second.

[0026] It should be noted that the method of the present invention achieves high temporal resolution audio detection through a two-layer segmentation strategy. The system receives reference audio and audio to be tested, and after preprocessing, it is segmented into 1-second segments. Each segment is further subdivided into 10 100-millisecond sub-segments. MFCC features and temporal statistical features are extracted for each sub-segment. Second-level features are obtained by aggregating the mean and maximum values. The features of the reference audio and the audio to be tested are concatenated and then input into the SVM model for scoring.

[0027] Preferably, compared with the prior art, the method of the present invention accurately locates transient anomalies of 10 milliseconds by combining second-by-second scoring with 100-millisecond segment division, achieving high temporal resolution; the single-second audio processing time is no more than 10 milliseconds, meeting the real-time detection requirements of the production line; at the same time, the SVM model file is only tens of MB, which can run on ordinary industrial control computers without GPU, achieving lightweight deployment, and a stable model is trained through small sample learning, improving the accuracy and stability of detection.

[0028] The following is combined with Figures 2a-2d The schematic diagrams shown, along with some preferred or optional examples of the present invention, more specifically describe the implementation process and / or effects of certain embodiments of the present invention.

[0029] (1) Audio loading and resampling First, the input audio file is read and parsed. The system supports stereo audio files in WAV and MP3 formats. The reading process consists of three steps: first, the audio file header information is parsed to extract metadata such as the original sampling rate, number of channels, and bit depth; then, the compressed audio data is decoded into a floating-point array to ensure the accuracy of subsequent processing; finally, the stereo signal is separated into independent left and right channels for easy processing and comparative analysis.

[0030] To ensure that audio files with different sampling rates can have a unified time scale, the sinc interpolation method is used to resample all audio to 16kHz. This standard sampling rate is chosen based on the Nyquist theorem, which can both preserve the main spectral information of the audio and reduce computational complexity.

[0031] (2) Audio segmentation The audio is first divided into 1-second segments, each containing 16,000 sampling points. This second-level segmentation captures the mid-range changes in the audio and facilitates real-time processing and scoring output. For segments shorter than 1 second, zero padding is used to ensure consistent processing.

[0032] Each main segment is further subdivided into 10 segments of 100 milliseconds each, with each segment containing 1600 sampling points. This millisecond-level fine-grained analysis can capture the transient characteristics and local anomalies of the audio. Through a two-level segmentation strategy, the system can simultaneously analyze the short-term and medium-term characteristics of the audio.

[0033] (3) Feature extraction Mel-frequency cepstral coefficients (MFCCs) are one of the most important features in audio signal processing, effectively characterizing the spectral envelope properties of audio. The steps for extracting MFCCs are as follows: Pre-emphasis processing: The signal is pre-emphasized to compensate for the attenuation of high-frequency components. Pre-emphasis is achieved through a first-order high-pass filter. This step enhances the high-frequency components and improves the signal-to-noise ratio. Framing and windowing: Each sub-segment is further divided into frames with a frame length of 25ms and a frame shift of 10ms, resulting in 75% overlap. A Hamming window function is applied to each frame to reduce spectral leakage. Spectrum analysis: Perform a Fast Fourier Transform (FFT) on the windowed signal to obtain its frequency domain representation, calculate the power spectral density, and provide input for subsequent Mel filtering; Mel filter bank: Employs 40 triangular filters uniformly distributed across the Mel frequency scale. The Mel scale simulates the nonlinear frequency perception characteristics of the human ear, offering high low-frequency resolution and low high-frequency resolution. The Mel frequency conversion formula is:

[0034] Logarithmic energy and DCT transform: Calculate the logarithmic energy of the output of each filter, and then perform a discrete cosine transform to obtain 40-dimensional MFCC coefficients; Statistical feature extraction: Calculate the mean and standard deviation of the MFCC coefficients for all frames within each sub-segment to form an 80-dimensional feature vector; Temporal feature extraction: RMS energy is sensitive to volume changes and energy anomalies, and is an important indicator for detecting audio distortion. The calculation formula is RMS = sqrt(Σx 2[n] / N), where x[n] represents the value of the audio signal at the nth sampling point, and N=1600 represents the total number of sampling points in the sub-segment; the frequency of the signal crossing zero points is counted to reflect the high-frequency component content of the signal. A high zero-crossing rate usually indicates the presence of noise or friction sound; the maximum absolute amplitude in the sub-segment is calculated to measure the intensity of the audio signal; the sharpness of the signal distribution is measured to reflect the impulse characteristics of the signal. A kurtosis value deviating from the normal distribution (kurtosis=3) indicates the presence of abnormal events or distortion; the symmetry of the signal distribution is measured, and the skewness value reflects the degree of asymmetric distortion of the signal; each 100-millisecond sub-segment generates an 85-dimensional feature vector, which contains: the mean and standard deviation of 40-dimensional MFCC (80 dimensions in total), RMS energy, zero-crossing rate, maximum amplitude, kurtosis and skewness (5 dimensions in total). These features are concatenated to form a feature vector containing time-domain and frequency-domain information.

[0035] (4) Feature calculation per second For each 1-second audio segment, the mean and maximum values ​​of the features of 10 sub-segments (each with 85 dimensions) are calculated to generate a 170-dimensional feature vector. Then, the 170-dimensional feature vector of the reference audio and the 170-dimensional feature vector of the corresponding degraded audio are concatenated to form a 340-dimensional combined feature vector, which is then fed into the model for training.

[0036] To ensure consistency in the dimensions of different features, StandardScaler is used for standardization. After standardization, the features have zero mean and unit variance, thus avoiding the influence of different feature scales on the classification results.

[0037] (5) Use Support Vector Machine (SVM) for model training. The specific steps are as follows: The radial basis function (RBF) kernel is used to map features to a high-dimensional space to achieve nonlinear classification. The RBF kernel function is defined as: K(x,y)=exp(-γ||xy|| 2 In this model, γ > 0 is the kernel function parameter, used to control the rate at which similarity decays with distance. A larger γ results in a narrower kernel distribution, making the model more sensitive to local differences, leading to more complex decision boundaries and a higher risk of overfitting. A smaller γ results in a wider kernel distribution, slower similarity decay, and smoother decision boundaries. Soft-margin SVM is used during model training, with a penalty coefficient of C = 10, where C is the regularization parameter. K(x,y) is the output of the kernel function, representing the similarity between samples x and y in the implicit high-dimensional feature space. exp(⋅) is an exponential function with the natural constant e as its base, used to map the distance metric to the similarity in the interval (0,1]. x and y are two sample vectors to be compared; ||xy|| 2 The squared Euclidean distance between samples; The optimal classification hyperplane is determined by solving the following quadratic programming problem:

[0038] Among them, J is the objective function to be optimized, w is the normal vector of the classification hyperplane, b is the bias term, is the slack variable of the i-th sample, and C is the regularization parameter; The audio quality is divided into 6 levels. Adopting the one-versus-one strategy, the number of classifiers is:

[0039] The grade definition of the score is: 0 points, no difference, the reference audio and the audio to be measured are almost the same; 1 point, slight difference, there is a short current sound or weak noise; 2 points, slight distortion, but acceptable; 3 points, obvious distortion, still acceptable; 4 points, serious distortion, the music is significantly distorted or the speech is incomprehensible; 5 points, extremely poor, the signal is completely lost or the sound quality is extremely poor.

[0040] To ensure the consistency of the training and test data distributions, in the test stage, the joint feature vectors of each audio sample to be measured are processed dimension by dimension for standardization, and the standardized feature vectors are obtained:

[0041] Among them, is the 340-dimensional joint feature vector obtained in step S5, is the mean and standard deviation statistically obtained dimension by dimension on the training set.

[0042] The audio quality level set of the present invention is , where 0 represents no difference and 5 represents a great difference. Adopting the one-versus-one strategy: for any pair of i < j, only these two types of samples are used to train a binary classifier <www. , in the SVM framework, the discriminant function of this classifier in the prediction stage is: ; Among them, represents the set of support vector indices of this binary classifier; represents the 340-dimensional joint feature of the k-th support vector; represents the joint feature vector of the audio sample to be measured; represents the dual coefficient corresponding to the k-th support vector; K is the kernel function, which is used to measure the similarity of two samples in the feature space; represents the label of the k-th support vector; represents the bias of this binary classifier.

[0043] Output a binary result for each binary classifier:

[0044] in, This is a function for outputting categories.

[0045] The final multi-class classification result is obtained by voting fusion, and the indicator function is denoted as [function name missing]. The value is 1 if the condition is true, and 0 otherwise. The total number of votes is calculated for each category.

[0046] in, This represents the cumulative number of votes for category c. The final prediction category is:

[0047] in, , This indicates maximizing the objective.

[0048] For each test audio, the model outputs its second-by-second predicted label and evaluates the prediction accuracy by comparing it with the actual label. At the same time, it generates a second-by-second score sequence and a visualization curve to facilitate the rapid location of abnormal periods.

[0049] To verify its performance in mass production, a scoring system was deployed at the production terminal of the headphone production line. The test audio was an 80-second standard music audio, and the test sample consisted of 200 headphone samples. All samples were first manually listened to and marked with a score per second, which included several abnormal samples to compare and verify the accuracy of the system's detection.

[0050] The artificial ear used in this test was a KEMAR, and the microphone was a Bruel & Kjaer 4138.

[0051] The testing steps are as follows: Step 1: Connect the headphones to the artificial ear and assemble it. Correctly place the headphones under test into the ear canal of the artificial ear. Step 2: Play audio signals. Input a standardized audio test signal into the headphones. Step 3: Audio acquisition. Place the microphone firmly inside the ear canal of the artificial ear to record the audio signal emitted by the headphones. Connect the microphone to an audio analyzer to analyze and save the audio signal in real time. Step 4: Scoring system. Use the input standard audio and the acquired headphone output audio as input to the scoring system. Score the audio per second and output a scoring curve. Step 5: Comparison and verification. Compare the software output by the scoring system with the data manually labeled, especially analyzing the time periods when the score is non-zero, and statistically analyze the accuracy of the scoring.

[0052] Test results show that the system can complete the analysis within 2 seconds for each 80-second test audio. In the batch test of 200 products, the system successfully detected all manually labeled abnormal samples without false alarms.

[0053] Figures 2a-2d Four scoring cases are presented, with blue waveforms representing the left channel and red waveforms representing the right channel. Each set of graphs, from top to bottom, consists of: the original reference audio waveform Original_signal.wav, the damaged audio waveform Corrupted_signal.wav, the manually labeled abnormal location labeled_signal.wav, and the model's output score curve Score.wav. The horizontal axis represents time, and the vertical axis represents the model's output score. The scores have been normalized, mapping 0-5 points to 0-1 points proportionally. This allows for a clear view of the algorithm's sensitive response to distortion, consistent with manual labeling, demonstrating excellent distortion localization capabilities.

[0054] in, Figure 2a The simulation added brief knocking signals: one knocking sound in the 3-4 second interval, two knocking sounds in the 4-5 second interval, and three knocking sounds in the 6-7 second interval. The knocking sounds were short in duration, and the distortion could not be located by relying solely on MFCC features. After adding multiple instantaneous features, the time periods in which the knocking signals appeared were accurately identified.

[0055] Figure 2b In the audio file being tested, there was a problem of random packet loss. Judging from the scoring results, the time periods when the packet loss occurred were accurately identified. Furthermore, the tiny plosive sounds in the B file were also accurately identified in the 19-20 second and 30-31 second intervals.

[0056] Figure 2c Similarly, the simulation added a tapping signal. It can be clearly seen that the more times the tapping signal is added within one second, the greater the difference and the higher the score. Although there is a difference in amplitude of the time domain waveform in the second half of the AB file, the human ear cannot hear the difference, so the score is 0, and it is considered that there is no difference between the two.

[0057] Figure 2d The model incorporates human voice noise, allowing it to easily identify differences and achieve high scores.

[0058] The aforementioned feature extraction methods can be implemented using existing technologies and techniques, and will not be elaborated upon in this example.

[0059] Reference Figure 3 and Figure 4 Other aspects disclosed in the embodiments of the present invention also propose an automatic audio quality scoring system that combines MFCC and temporal statistical features, including: An audio input module is used to receive reference audio and the audio to be tested; The preprocessing module is used for audio resampling and time alignment; The segmentation module is used to segment audio by second and further subdivide it into sub-segments; The feature extraction module is used to extract MFCC features and temporal statistical features; The feature aggregation module is used to calculate the mean and maximum value of the features of the sub-segments; The model inference module is used to load the SVM model and output a score. The results output module is used to generate second-by-second scoring sequences and anomaly location information.

[0060] The system further includes one or more processors and a memory; the memory is used to store operable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, including the flow of the audio quality automatic scoring method combining MFCC and temporal statistical features of the foregoing embodiments, especially... Figure 1 The flowchart of the method is shown.

[0061] Other aspects disclosed in the embodiments of the present invention also propose a computer-readable medium for storing software including instructions executable by one or more computers, which, upon execution, cause the one or more computers to perform operations including the flow of the audio quality automatic scoring method combining MFCC and temporal statistical features of the foregoing embodiments, particularly... Figure 1 The flowchart of the method is shown.

[0062] It should be recognized that embodiments of the present invention may be implemented or carried out by computer hardware, a combination of hardware and software, or by computer instructions stored in a non-transitory computer-readable storage medium.

[0063] The method can be implemented using standard programming techniques, including a non-transitory computer-readable storage medium configured with a computer program in the computer program, wherein the storage medium is configured such that the computer operates in a specific and predefined manner.

[0064] Each program can be implemented in a high-level procedural or object-oriented programming language to communicate with the computer system; however, if required, the program can be implemented in assembly or machine language.

[0065] In any case, the language can be either compiled or interpreted.

[0066] Furthermore, for this purpose, the program can run on programmed application-specific integrated circuits.

[0067] The processes described herein (or variations and / or combinations thereof) can be executed under the control of one or more computer systems configured with executable instructions, and can be implemented by hardware or a combination thereof as code (e.g., executable instructions, one or more computer programs, or one or more applications) that commonly executes on one or more processors. The computer program includes a plurality of instructions executable by one or more processors.

[0068] Furthermore, the method can be implemented in any suitable computing platform, including but not limited to personal computers, minicomputers, mainframes, workstations, networked or distributed computing environments, standalone or integrated computer platforms, or in communication with charged particle tools or other imaging devices.

[0069] Various aspects of the present invention can be implemented in machine-readable code stored on a non-transitory storage medium or device, whether portable or integrated into a computing platform, such as a hard disk, optical read and / or write storage medium, RAM, ROM, etc., such that it can be read by a programmable computer, and when the storage medium or device is read by the computer, it can be used to configure and operate the computer to perform the processes described herein.

[0070] Furthermore, machine-readable code, or parts thereof, can be transmitted via wired or wireless networks.

[0071] When such media includes instructions or programs that combine with a microprocessor or other data processor to implement the steps described above, the invention described herein includes these and other different types of non-transitory computer-readable storage media.

[0072] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An automatic audio quality scoring method combining MFCC and temporal statistical features, characterized in that, include: Receive the reference audio signal and the audio signal to be tested, resample the two audio signals to 16000Hz, and perform time alignment; The time-aligned audio is segmented by second, and each segment is further divided into 10 sub-segments; Extract features from each 100-millisecond segment to obtain the segment feature vector; For each 1-second audio segment, the mean and maximum value of the 10 sub-segments feature vectors are calculated for each dimension, resulting in 85-dimensional mean features and 85-dimensional maximum features, which are then concatenated to form a 170-dimensional second-level feature vector. The second-level feature vectors of the reference audio and the audio to be tested for the corresponding time period are concatenated to form a 340-dimensional joint feature vector; The joint feature vector is input into a pre-trained support vector machine (SVM) model, which outputs an integer score in the range of 0 to 5 to represent the degree of distortion of the audio at that second.

2. The automatic audio quality scoring method combining MFCC and temporal statistical features according to claim 1, characterized in that, The feature extraction for each 100-millisecond segment includes: Extract the 40-dimensional frequency cepstral coefficients (MFCCs) and calculate their mean and standard deviation to obtain 80-dimensional frequency domain features; Calculate five time-domain statistical characteristics: root mean square energy (RMS), zero-crossing rate (ZCR), maximum amplitude, skewness, and kurtosis. The frequency domain features and the time domain statistical features are concatenated to form an 85-dimensional sub-segment feature vector.

3. The automatic audio quality scoring method combining MFCC and time-domain statistical features according to claim 2, characterized in that, The specific process for extracting the Mel frequency cepstral coefficients (MFCC) features includes: Pre-emphasis processing of audio signals; Hamming windows are used for framing, with a frame length of 25 milliseconds and a frame shift of 10 milliseconds. Perform a Fast Fourier Transform on each frame; The Mel spectrum was calculated using a 40-triangular Mel filter bank; After taking the logarithm, a discrete cosine transform is performed to obtain the 40-dimensional MFCC coefficients.

4. The automatic audio quality scoring method combining MFCC and time-domain statistical features according to claim 2, characterized in that, When calculating the time-domain statistical features, RMS, ZCR, maximum amplitude, kurtosis, and skewness are multiplied by coefficients of 5, 3, 3, 3, and 3, respectively, to enhance the features.

5. The automatic audio quality scoring method combining MFCC and time-domain statistical features according to claim 1, characterized in that, The SVM model uses a radial basis function (RBF) kernel, and the kernel function is as follows: K(x,y)=exp(-γ||x-y|| 2 ) Where γ > 0 is the kernel function parameter, used to control the rate at which similarity decays with distance. A larger γ results in a narrower kernel distribution, making the model more sensitive to local differences, leading to a more complex decision boundary and increased risk of overfitting. A smaller γ results in a wider kernel distribution, slower similarity decay, and a smoother decision boundary. Soft-margin SVM is used during model training, with a penalty coefficient of C = 10, where C is the regularization parameter. K(x,y) is the output of the kernel function, representing the similarity between samples x and y in the implicit high-dimensional feature space. exp(⋅) is an exponential function with the natural constant e as its base, used to map the distance metric to the similarity in the interval (0,1). x and y are two sample vectors to be compared; ||xy|| 2 The squared Euclidean distance between the samples is given.

6. The automatic audio quality scoring method combining MFCC and temporal statistical features according to claim 1 or 5, characterized in that, The SVM model needs to be trained, specifically including: Collect reference audio, audio to be tested, and corresponding manually annotated scores; Perform a joint feature vector generation process on the training data to obtain the initial joint feature vector; The initial joint feature vector was standardized using StandardScaler; The SVM classification model is trained using the standardized initial joint feature vector and labeled scores. Save the trained model parameters and standardized parameters.

7. The automatic audio quality scoring method combining MFCC and time-domain statistical features according to claim 1, characterized in that, The rating levels are defined as follows: 0 points, no difference, the reference audio and the audio to be tested are almost identical; 1 point, slight difference, with a brief electrical hum or weak noise present; 2 points, slightly distorted, but acceptable; 3 points, clearly distorted, but still acceptable; 4 points, severely distorted, music is obviously distorted or speech is incomprehensible; 5 points, extremely poor, signal completely lost or extremely poor sound quality.

8. An automatic audio quality scoring system combining MFCC and time-domain statistical features, characterized in that, The system is implemented based on the automatic audio quality scoring method combining MFCC and temporal statistical features as described in any one of claims 1 to 7, including: An audio input module is used to receive reference audio and the audio to be tested; The preprocessing module is used for audio resampling and time alignment; The segmentation module is used to segment audio by second and further subdivide it into sub-segments; The feature extraction module is used to extract MFCC features and temporal statistical features; The feature aggregation module is used to calculate the mean and maximum value of the features of the sub-segments; The model inference module is used to load the SVM model and output a score. The results output module is used to generate second-by-second scoring sequences and anomaly location information.

9. The automatic audio quality scoring system combining MFCC and time-domain statistical features according to claim 8, characterized in that, The system also includes one or more processors; The memory stores operable instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, including the flow of the automatic audio quality scoring method combining MFCC and time-domain statistical features as described in any one of claims 1 to 7.

10. A computer-readable medium for storing software, characterized in that: The software includes instructions executable by one or more computers, which, upon execution, cause the one or more computers to perform operations including the flow of an automatic audio quality scoring method combining MFCC and temporal statistical features as described in any one of claims 1 to 7.

Images