Real-time noise scene classification system and system and artificial cochlea equipment

Abstract

The invention discloses a real-time noise scene classification system and system and an artificial cochlea device, and the method comprises the steps: for a collected audio signal, extracting a feature vector of a mixed time-frequency domain feature, the vector is composed of a frequency domain feature based on a critical band ratio (CBR), a short-time energy (STE) feature and a short-time zero-crossing rate (ZCR) feature, the information dimension of the features is effectively expanded, and the multi-level characterization capability of a complex auditory scene is enhanced. Inputting the feature vectors into a trained support vector machine (SVM) classification model to obtain a scene classification result of the current audio segment; and performing frame-by-frame feature extraction and segmented average accumulation processing on continuously input long audio signals, and outputting a final scene classification probability through frequency statistics and confidence coefficient weighting of multi-segment classification results. Through cooperation of feature combination and a segmentation probability weighting mechanism, real-time classification of quiet, normal and noise scenes is realized.

Description

Technical Field

[0001] This application relates to the field of cochlear implant signal processing, and more specifically, to a real-time noise scene classification system, a system, and a cochlear implant device. Background Technology

[0002] Current cochlear implant scene recognition algorithms are primarily based on extracting features from environmental sounds and using machine learning models for classification and recognition. Traditional methods typically rely on single acoustic features such as Mel-Frequency Cepstral Coefficients (MFCCs) or spectral centroids, which struggle to comprehensively capture the multidimensional information in complex auditory scenes. Consequently, these models are prone to overfitting under limited training data, leading to a significant decrease in their generalization ability in real-world environments.

[0003] The reliance on a single feature limits the ability to represent dynamic sound scenes, and model overfitting affects recognition stability, failing to meet the auditory needs of cochlear implant users for real-time scene switching. These technical shortcomings collectively restrict the practical performance of existing systems. Summary of the Invention

[0004] To address the aforementioned technical problems, this application discloses a real-time noise scene classification system, a system, and a cochlear implant device. This method constructs a training dataset using multi-scene audio data collected in real-world scenarios and applies gain processing to transitional scene data to enhance category boundaries. Based on the synergy of feature combination and segmented probability weighting mechanisms, it achieves real-time classification of quiet, normal, and noisy scenes. Specifically, the technical solution of this application is as follows: Firstly, this application discloses a real-time noise scene classification method, including: For the acquired audio signal, a specified number of mixed time-frequency domain features are extracted as classification criteria; the mixed time-frequency domain features consist of multi-dimensional critical band ratio features, as well as short-time energy features and short-time zero-crossing rate features; The feature vectors of the hybrid time-frequency domain features are input into the trained support vector machine classification model to obtain the classification results of the audio segments. For the long audio signals input in real time, frame-by-frame feature extraction and segmented average accumulation processing are performed, and frequency statistics and confidence weighting are performed on the classification results of multiple segments to obtain the final classification results.

[0005] In some implementations, extracting a specified number of mixed time-frequency domain features of a given dimension as a classification criterion for the acquired audio signal specifically includes: After pre-emphasis and windowing of the input audio signal, it is divided into frames with fixed frame length and frame shift. The power spectrum of each frame signal is calculated by applying Fast Fourier Transform and the positive frequency part is retained. Divide the power spectrum into N critical bands on the frequency axis, calculate the bandwidth of each critical band, and sum the power spectrum energies within the bands to obtain the critical band energy. Calculate the ratio of each critical band energy to the total energy of each frame, i.e., the critical band ratio, to obtain the N-dimensional critical band ratio feature; The short-time energy of each frame is calculated as the sum of squares of the intra-frame signal, and the short-time energy characteristics are log-normalized. Calculate the short-time zero-crossing rate for each frame, where the short-time zero-crossing rate is the ratio of the number of symbol changes within the frame to the frame length; The hybrid time-frequency domain feature is obtained by concatenating the N-dimensional critical band ratio feature, the short-time energy feature, and the short-time zero-crossing rate feature.

[0006] In some implementations, the critical band energy is calculated using the following formula; ; ; in, Let m be the calculated bandwidth of the b-th critical band; b is the critical band index, and m is the frame index. The center frequency of the b-th critical band; Let be the total energy of the m-th frame signal within the b-th critical band. Let k be the physical frequency of the frequency point; The lower limit frequency; The upper limit frequency; Let k be the power spectral density value of the m-th frame signal at frequency k. The critical band ratio is calculated using the following formula; ; in, The critical band ratio of the b-th critical band in the m-th frame; The total energy of the signal in the m-th frame; To prevent division by zero of extremely small constants.

[0007] In some implementations, the short-time energy is calculated using the following formula; ;

[0008] in, Let m be the short-time energy of the m-th frame; The short-time energy feature of the m-th frame after logarithmic normalization is used for feature vector concatenation; m is the frame index; L is the frame length; n is the index of the sample within the frame. This represents the nth sample value in the m-th frame of the signal; The short-time zero-crossing rate is calculated using the following formula; ; in, Let be the short-time zero-crossing rate of the m-th frame; This is the (n-1)th sample value in the m-th frame of the signal.

[0009] In some implementations, for the acquired audio signal, a specified number of mixed time-frequency domain features are extracted as a classification criterion. This also includes: The preprocessing operation for the acquired audio signal specifically includes: A pre-emphasis filter is applied to compensate for the attenuation of high-frequency components in the audio signal, thereby improving signal clarity; The processed continuous signal is divided into short time frames to capture the time-varying characteristics of the signal; a window function is applied to the short time frames to reduce spectral leakage.

[0010] In some implementations, the real-time noise scene classification method further includes: collecting audio samples from multiple environments in real-world scenarios to construct an initial dataset; performing gain processing on audio samples from transitional scenarios to expand the category boundaries; and dividing the dataset into a training set, a validation set, and a test set according to a set ratio. Feature extraction is performed on audio samples in the training and validation sets respectively to obtain corresponding feature matrices and add label vectors for training and hyperparameter optimization of the support vector machine model; The same features are extracted from the audio samples in the test set for model performance evaluation.

[0011] In some implementations, the real-time noise scene classification method further includes: training a support vector machine model using a radial basis function kernel function and a multi-class one-to-one encoding strategy; specifically including the following steps: Based on the feature matrix and the label vector, a one-to-one binary classifier is constructed for each scenario; The hyperparameters are automatically adjusted using Bayesian optimization; the optimization objective is to minimize the classification error; and finally, a final model that can be used for classification in real-world scenarios is obtained by training based on the optimal parameter combination.

[0012] In some embodiments, the real-time noise scene classification method further includes: adding a real-time frame segmentation processing framework to the trained support vector machine model; For long audio signals input in real time, frame segmentation is performed with fixed frame length and frame shift; The hybrid time-frequency domain features are extracted frame by frame, and the average of the frame number corresponding to 1-2 seconds is taken to obtain the multi-dimensional features of the segment. The features are then input into the scene classification model for classification, and the intermediate scene results and confidence scores are output. Process each segment until the long audio signal ends, count the frequency of each scene and calculate the weighted probability: ; Where i represents the index of the three scene types. Let j be the frequency of segment j. The confidence level of this segment; based on The maximum value is output as the final scene classification result.

[0013] Secondly, this application also discloses a real-time noise scene classification system, which is used to implement the steps of a real-time noise scene classification method described in any of the above embodiments; Includes: a signal acquisition module for acquiring audio signals; The feature extraction module is used to extract a specified number of mixed time-frequency domain features as a classification criterion for the audio signal; the mixed time-frequency domain features consist of multi-dimensional critical band ratio features, as well as short-time energy features and short-time zero-crossing rate features; The model classification module is used to input the feature vector of the hybrid time-frequency domain features into the trained support vector machine classification model to obtain the classification result of the audio segment; for the long audio signal input in real time, it performs frame-by-frame feature extraction and segment average accumulation processing, and performs frequency statistics and confidence weighting on the classification results of multiple segments to obtain the final classification result.

[0014] Thirdly, this application also discloses a cochlear implant device, which includes a real-time noise scene classification system as described in the above embodiments.

[0015] Compared with the prior art, this application has at least one of the following beneficial effects: 1. In the feature extraction stage, this application adopts critical band ratio frequency domain features combined with short-time energy and short-time zero-crossing rate to form a more robust multi-dimensional feature vector; by extracting complementary acoustic features of multiple dimensions from the time domain and frequency domain, the information dimension of the features is effectively expanded, and the ability to represent complex auditory scenes at multiple levels is enhanced.

[0016] 2. This application introduces real-world multi-scene recording and gain adjustment during the data acquisition stage to enhance dataset diversity; and selects support vector machines during model training and classification to optimize decision boundaries and improve generalization ability.

[0017] 3. The real-time processing architecture of this application achieves dynamic classification of continuous signals by adding frame accumulation and probability weighting mechanisms; finally, it is ported to a digital signal processing platform in the hardware deployment unit to ensure the feasibility of practical applications. Thus, a good balance is achieved between algorithm robustness and real-time performance. Attached Figure Description

[0018] The preferred embodiments will now be described in a clear and easy-to-understand manner, in conjunction with the accompanying drawings, to further explain the above-mentioned characteristics, technical features, advantages, and implementation methods of this application.

[0019] Figure 1 This is a flowchart illustrating the steps of an embodiment of a real-time noise scene classification method according to this application; Figure 2 This is a flowchart illustrating the steps of another embodiment of a real-time noise scene classification method according to this application. Detailed Implementation

[0020] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application can also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0021] It should be understood that, when used in this specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or sets.

[0022] To keep the drawings concise, each figure only schematically shows the parts relevant to the invention, and these do not represent the actual structure of the product. Furthermore, to facilitate understanding, in some figures, only one of components with the same structure or function is schematically depicted, or only one is labeled. In this document, "one" not only means "only one," but can also mean "more than one."

[0023] It should also be further understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0024] Furthermore, in the description of this application, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0025] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the specific implementation methods of this application will be described below with reference to the accompanying drawings. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings and other implementation methods can be obtained based on these drawings without creative effort.

[0026] In existing technologies, cochlear implant sound processors typically employ noise classification systems based on fixed thresholds or simple spectral analysis. For example, they distinguish between quiet, normal, and noisy scenes by calculating the power spectral density of the signal or using pre-defined noise template matching. These methods primarily rely on static feature extraction, such as Mel-frequency cepstral coefficients or zero-crossing rate combined with energy thresholding, and use basic decision trees or k-nearest neighbor algorithms for scene recognition during the classification stage. For instance, in a typical sound processor, the input audio signal is first converted to the frequency domain via Fourier transform, then MFCC features are extracted, and scenes are classified as quiet (high signal-to-noise ratio), normal (moderate noise), and noisy (low signal-to-noise ratio) using predefined thresholds, thereby triggering corresponding noise reduction strategies, such as spectral subtraction or Wiener filtering.

[0027] However, these existing solutions have significant drawbacks: First, feature extraction relies too heavily on a single dimension, either the time or frequency domain, leading to low discrimination accuracy in complex noisy environments. This is especially true when scene transitions (e.g., an office changing from quiet to moderately noisy), where classification accuracy drops by 20%-30%. This is because features like MFCC are not sensitive enough to dynamic changes and cannot effectively capture the spectral distribution and time-varying characteristics of noise. Second, classification models (such as decision trees) have weak generalization ability, easily overfitting or underfitting in noisy scenarios. Particularly in real-time processing, they ignore signal continuity, resulting in unstable classification results and a misclassification rate as high as 15%-25%. Third, existing real-time processing architectures are mostly batch-based, unable to adapt to the continuous input requirements of cochlear implants. For example, they lack segmented accumulation and probability weighting mechanisms for long audio signals, causing delays or biases in final scene judgment and affecting the immediate application of noise reduction algorithms.

[0028] Furthermore, these methods do not consider hardware constraints in real-world deployments, such as the limited computing resources of DSP platforms. This results in limited real-time performance of the algorithms in noisy environments, with processing delays reaching hundreds of milliseconds, thereby reducing users' speech perception capabilities in noisy settings. Research shows that in noisy environments with Mandarin speakers, the tone recognition rate of existing technologies drops by up to 40% compared to quiet environments, significantly limiting the actual effectiveness of cochlear implants.

[0029] To address the aforementioned algorithmic shortcomings, this application provides a processing method for real-time noise scene classification. By extracting complementary acoustic features across multiple dimensions from the time and frequency domains, the information dimensionality of the features is effectively expanded, enhancing the multi-level representation capability of complex auditory scenes. Based on this, model training is performed using a large-scale, multi-scene dataset, and a probability-weighted real-time processing mechanism is introduced, significantly improving the system's classification accuracy.

[0030] Reference manual attached Figure 1As shown, an embodiment of a real-time noise scene classification method of this application specifically includes the following steps: S100: For the acquired audio signal, extract a specified number of mixed time-frequency domain features as a classification criterion. The mixed time-frequency domain features consist of multi-dimensional critical band ratio features, as well as short-time energy features and short-time zero-crossing rate features.

[0031] S200, the feature vector of the hybrid time-frequency domain features is input into the trained support vector machine classification model to obtain the classification result of the audio segment; for the long audio signal input in real time, frame-by-frame feature extraction and segment-by-segment averaging and accumulation processing are performed, and frequency statistics and confidence weighting are performed on the classification results of multiple segments to obtain the final classification result.

[0032] Specifically, the hybrid time-frequency domain features in this application include at least time-domain features and frequency-domain features. In the following embodiments, the time-domain features include at least: one-dimensional short-time energy (STE) features and one-dimensional short-time zero-crossing rate (ZCR) features. The frequency-domain features include at least N-dimensional critical band ratio (CBR) features. The combination of these features forms a hybrid time-frequency domain feature with N+1+1 dimensions.

[0033] In other embodiments, the number of dimensions of the critical band ratio feature used in this application is based on performance evaluations of feature dimensions by those skilled in the art. Specifically, by comparing model performance with different numbers of dimensions on a validation set, the number of dimensions with the best overall performance is selected as the final configuration. Optionally, the frequency domain features include at least 24 dimensions of critical band ratio features. This is combined with one-dimensional short-time energy features and one-dimensional short-time zero-crossing rate features. The above feature combination forms a hybrid time-frequency domain feature with 26 dimensions.

[0034] Reference manual attached Figure 2 As shown, this application provides another embodiment of a real-time noise scene classification method. Based on any embodiment of the above system, for the acquired audio signal, a specified number of mixed time-frequency domain features are extracted as the classification criterion. Prior to this, the method further includes preprocessing the acquired audio signal. Specifically, it includes the following steps: S010, a pre-emphasis filter is applied to compensate for the attenuation of high-frequency components in the audio signal, thereby improving signal clarity.

[0035] In some implementations, the following formula (1) is used for pre-emphasis filtering to enhance high-frequency components and improve signal clarity.

[0036]

[0037] in, This is the data after pre-emphasis; The data is before weighting; 'a' is the filtering coefficient, with 'a' being the preferred value of 0.97.

[0038] S020, the processed continuous signal is divided into short time frames to capture the time-varying characteristics of the signal. A window function is applied to the short time frames to reduce spectral leakage.

[0039] In some implementations, the pre-emphasized data is framed and windowed. Frames are divided into segments with a frame length of 256 points and a frame shift of 128 points (i.e., 50% overlap), and window functions are used to reduce spectral leakage.

[0040] In other implementations, window functions are used specifically, Hamming windows, to reduce spectral leakage. During waveform stitching, a window function is applied to the overlapping portions of each waveform segment. The window function gradually reduces the amplitude of the waveforms within the overlapping region to create a smooth transition. Thus, even if the amplitudes of two waveforms are not perfectly matched, the window function reduces the amplitude difference, thereby decreasing the likelihood of abrupt changes in sound.

[0041] This application provides another embodiment of a real-time noise scene classification method. Based on any embodiment of the above system, step S100, for the acquired audio signal, extracts a specified number of mixed time-frequency domain features as a classification criterion, including two parts: calculating frequency domain features and calculating time domain features, specifically including the following sub-steps: S110 pre-emphasizes and windows the input audio signal, then divides it into frames with a fixed frame length and frame shift. It applies a fast Fourier transform to each frame to calculate the power spectrum and retains the positive frequency portion.

[0042] Specifically, a Fast Fourier Transform (FFT) is applied to each frame of signal to calculate the power spectrum, retaining only the positive frequency portion, that is, the portion from 0 Hz to the Nyquist frequency.

[0043] S120, the power spectrum is divided into N critical bands on the frequency axis, the bandwidth of each critical band is calculated, and the power spectrum energy within the band is summed to obtain the critical band energy.

[0044] Specifically, 24 critical bands are divided on the Bark scale, with center frequencies ranging from 20 Hz to the Nyquist frequency (approximately 8 kHz). The bandwidth of each critical band is calculated according to equation (2), and the critical band energy (CBE) is obtained by summing the power spectra within the band, as shown in equation (3).

[0045]

[0046] in, Let m be the calculated bandwidth of the b-th critical band; b is the critical band index, and m is the frame index. The center frequency of the b-th critical band; Let be the total energy of the m-th frame signal within the b-th critical band. Let k be the physical frequency of the frequency point; The lower limit frequency; The upper limit frequency; Let be the power spectral density value of the m-th frame signal at frequency k.

[0047] S130, calculate the ratio of each critical band energy to the total energy of each frame, i.e., the critical band ratio, to obtain the N-dimensional critical band ratio feature.

[0048] Specifically, CBR, based on the ratio calculation of critical band energy, simulates the Barker-scale nonlinear frequency lobes of the human auditory system. In cochlear implant applications, noisy scenarios often involve complex variations in spectral distribution. While traditional MFCC features capture the overall spectral shape, they are insufficiently sensitive to the local energy distribution of noise. CBR, through its ratio calculation, emphasizes relative energy distribution, prioritizing areas sensitive to human hearing, making it suitable for distinguishing scenarios with significant differences in noise levels.

[0049] In some implementations, the CBE is normalized by dividing the total energy per frame to obtain the critical band ratio (CBR), as shown in equation (4).

[0050]

[0051] in, The critical band ratio of the b-th critical band in the m-th frame; The total energy of the signal in the m-th frame; To prevent division by zero of extremely small constants.

[0052] S140, calculate the short-time energy of each frame as the sum of squares of the intra-frame signals, and log-normalize the short-time energy characteristics.

[0053] S150, calculate the short-time zero-crossing rate for each frame, where the short-time zero-crossing rate is the ratio of the number of symbol changes within a frame to the frame length.

[0054] Specifically, Short-Time Energy (STE) characterizes the intensity of a signal within a short time window, and its value is calculated by formula (6) and normalized to log10. At the physical level, STE reflects the amplitude dynamics of the signal. In applications, this characteristic enables it to effectively highlight transient high-energy events such as noise bursts in noisy environments.

[0055] The short-time zero-crossing rate (ZCR) is calculated by counting the number of positive and negative polarity changes between adjacent sampling points within a signal frame. Since this rate of change is directly related to the instantaneous frequency and amplitude of the signal, ZCR can effectively characterize the joint frequency and amplitude dynamics of the signal within a short time window.

[0056] These time-domain features are chosen because they are computationally simple, do not require full FFT spectrum, are suitable for low-power DSP deployments, and complement CBR features in the frequency domain: CBR provides static frequency distribution, while STE / ZCR supplements dynamic time-domain variation, avoiding ambiguity of single features in transitional scenarios (such as from quiet scenes to moderate noise).

[0057]

[0058] in, Let m be the short-time energy of the m-th frame; The short-time energy feature of the m-th frame after logarithmic normalization is used for feature vector concatenation; m is the frame index; L is the frame length; n is the index of the sample within the frame. This represents the nth sample value in the m-th frame of the signal; Let be the short-time zero-crossing rate of the m-th frame; This is the (n-1)th sample value in the m-th frame of the signal.

[0059] S160, the N-dimensional critical band ratio feature, the short-time energy feature, and the short-time zero-crossing rate feature are concatenated to obtain the hybrid time-frequency domain feature.

[0060] In this embodiment, 24-dimensional CBR features, 1-dimensional short-time energy, and 1-dimensional short-time zero-crossing rate are concatenated into a 26-dimensional feature vector. Specifically, to construct an effective classification model, the 26-dimensional feature vector is designed as a CBR-dominated spectrum, combined with STE and ZCR to enhance time-varying characteristics, thereby achieving dimensional balance among different attribute information. This combined feature is normalized before being input into the support vector machine to adapt to the nonlinear mapping mechanism of its radial basis function (RBF) kernel, thus ensuring that the model has stronger expressive power and robustness when facing multi-dimensional mixed inputs.

[0061] This application discloses another embodiment of a real-time noise scene classification method. Based on the above embodiment, before putting the classification model into practical application, the following step needs to be performed first: S300, constructing a dataset for each scene, and using the dataset to train and optimize the scene classification model to improve its recognition performance and generalization ability.

[0062] S310 involves constructing the dataset. Specifically, this includes the following sub-steps: S311, collecting audio samples from various environments in real-world scenarios to construct an initial dataset; performing gain processing on audio samples from transitional scenarios to expand category boundaries; and dividing the dataset proportionally into a training set, a validation set, and a test set.

[0063] Specifically, to construct a training dataset for cochlear implant noise scenarios, this application collected audio data from various typical environments, including a meeting room (defined as a quiet scenario with a high signal-to-noise ratio), a cafeteria (a normal scenario with moderate noise), a computer room (a noisy scenario with high noise), and an office (a transitional scenario). The data collection equipment used a professional recorder (sampling rate of 16129Hz), with each scenario recorded for at least 10 minutes to ensure coverage of natural variations.

[0064] In the specific processing, the office audio data was divided into two parts: one half was directly assigned to the "quiet" scene, and the other half, after energy comparison with the cafeteria scene and application of gain to simulate moderate noise enhancement, was assigned to the "normal" scene. This expanded the boundary coverage of the dataset, avoiding the overfitting problem caused by a single dataset in existing technologies, and further increasing the robustness of the model. Subsequently, all data were divided into training, validation, and test sets in a 6:2:2 ratio. The training set was used for model feature parameter learning and optimization, the validation set was used for hyperparameter tuning and preventing overfitting, and the test set was used for final classification performance evaluation. This method ensures the realism and balance of the dataset, and compared with the simulated data or public datasets of existing technologies, it is more consistent with the actual use scenario of cochlear implants.

[0065] S312, feature extraction is performed on the audio samples in the training set and validation set respectively to obtain the corresponding feature matrix and add label vectors for training and hyperparameter optimization of the support vector machine model.

[0066] S313, Extract the same features from the audio samples in the test set for model performance evaluation.

[0067] Specifically, features are extracted from the training set labeled with three scene types: normal, quiet, and noise, to form a feature matrix. (Each row represents a sample, 26 dimensions) and label vector (Category index: 1=normal, 2=quiet, 3=noise).

[0068] This application discloses another embodiment of a real-time noise scene classification method, which further includes: S320, SVM model training: using radial basis function (RBF) kernel function, the SVM model is trained through a multi-class one-vs-one encoding strategy.

[0069] In some implementations, the scene classification model of this application is a Support Vector Machine (SVM) model, which uses SVM as the classifier to train the aforementioned feature vectors. Specifically, SVM is a supervised learning model based on statistical learning theory, which achieves classification by constructing a maximum margin hyperplane and has good generalization ability. The scene classification model is trained through the following sub-steps.

[0070] S321, Based on the feature matrix and the label vector, construct a one-to-one binary classifier for each scenario.

[0071] Specifically, the feature matrix X and label vector Y are input into the SVM classifier to construct three sets of one-to-one sub-binary classifiers for the three scenarios.

[0072] S322 uses Bayesian optimization to automatically adjust hyperparameters, with the optimization objective being to minimize classification error. Finally, a final model suitable for classification in real-world scenarios is obtained by training based on the optimal parameter combination.

[0073] Specifically, Bayesian optimization is used to automatically adjust the hyperparameters (BoxConstraint, controlling the soft margin) and kernel scale parameter (KernelScale, controlling the RBF kernel width), with optimization ranges of [0.1, 100] and [0.01, 10], respectively. The optimization process employs 5-fold cross-validation, aiming to minimize the classification error, with an iteration limit of 50. The "expected-improvement-plus" acquisition function is used to balance exploration and exploitation. After optimization, the optimal parameter combination is selected to generate the final model.

[0074] In some implementations, this embodiment further includes the step: S323, obtaining the output result of the classifier based on the trained training model.

[0075] Specifically, in this embodiment, the outputs of the three sub-classifiers are integrated through a voting mechanism to determine the final classification result. Compared with traditional decision trees or k-nearest neighbor algorithms, this method improves classification performance through high-dimensional feature space mapping and maximum margin hyperplane optimization.

[0076] In other implementations, the step S330 is also included: classifying and predicting the newly input signal based on the trained training model.

[0077] Specifically, for the feature vector of a new input signal, the SVM model maps it to a high-dimensional feature space and classifies it based on the optimized hyperplane, outputting the scene category (normal, quiet, or noise).

[0078] In other implementations, the step S340, performance evaluation, is also included.

[0079] Specifically, precision, recall, and F1 score are calculated for each scenario using a confusion matrix to ensure balanced classification performance. The calculation formula is as follows:

[0080] in, For accuracy; Recall rate; The score is the F1 score. For a real example; This is a false positive example. This is a false negative.

[0081] Specifically, precision measures the model's accuracy in classifying a specific scene, recall reflects its completeness in identifying that scene, and the F1 score serves as the harmonic mean of the two to comprehensively evaluate the recognition performance across different categories. By analyzing the distribution of these metrics across different scene categories, it is possible to effectively assess whether the model exhibits classification bias or structural bias.

[0082] Based on any of the above embodiments, this application discloses another embodiment of a real-time noise scene classification method, which further includes the following steps: S400, adding a real-time frame processing framework on the basis of the SVM model.

[0083] Specifically, to simulate the continuous signal input of a cochlear implant, this application adds a real-time frame-segmentation processing framework to the trained SVM model. (Reference) Figure 2 As shown, the specific steps include the following: The S410 performs frame segmentation on long audio signals input in real time, with fixed frame length and frame shift.

[0084] Specifically, long audio signals input in real time (such as those captured by a microphone) are framed with a frame length of 256 points and a frame shift of 128 points (50% overlap).

[0085] S420, extract the hybrid time-frequency domain features frame by frame, and average the number of frames corresponding to 1-2 seconds for this segment to obtain the multidimensional features of this segment. Input the multidimensional features into the scene classification model for classification, and output the intermediate scene results and confidence scores.

[0086] Specifically, with a classification interval of 1 second or 2 seconds (corresponding to 125 or 251 frames, based on a sampling rate of 16129Hz), when the cumulative number of frames reaches the required number, the 26-dimensional feature vector of that segment is extracted, and the mean of the features of that segment is calculated. The vector is then input into the SVM model for classification, and the intermediate scene result (normal, quiet, or noise) is output.

[0087] Traditional methods require complete signal processing, while cochlear implants are sensitive to signal delays (requiring real-time processing). This architecture classifies signals segment by segment, adapting to the continuous flow of the cochlear implant. Simultaneously, the feature accumulation mechanism buffers variations, avoiding the potential impact of single-frame jitter. At the hardware deployment level, the two-second 26-dimensional feature accumulation buffer also effectively saves memory, conserves hardware resources, and improves operational efficiency.

[0088] S430, process segment by segment until the long audio signal ends, count the frequency of occurrence of various scenarios and calculate the weighted probability.

[0089] Specifically, the signal is processed segment by segment until it ends, and the frequency of occurrence of various scenarios is counted. (i=1,2,3 correspond to three scenarios), calculate the weighted probability as shown in equation (8), where the weights are... Adjust the confidence level (e.g., posterior probability) based on the SVM output.

[0090]

[0091] Where i represents the index of the three scene types. Let j be the frequency of segment j. The confidence level of this segment; based on The maximum value is output as the final scene classification result.

[0092] The final scene classification result is output based on weighted probability, ensuring the continuity and stability of long signal processing.

[0093] Based on the same concept, this application also discloses a real-time noise scene classification system. The system is used to implement the real-time noise scene classification method described in any of the above embodiments. Specifically, one embodiment of the real-time noise scene classification system of this application includes: The signal acquisition module is used to acquire audio signals.

[0094] The feature extraction module is used to extract a specified number of mixed time-frequency domain features as a classification criterion for the audio signal; the mixed time-frequency domain features consist of multi-dimensional critical band ratio features, as well as short-time energy features and short-time zero-crossing rate features; The model classification module is used to input the feature vector of the hybrid time-frequency domain features into the trained support vector machine classification model to obtain the classification result of the audio segment; for the long audio signal input in real time, it performs frame-by-frame feature extraction and segment average accumulation processing, and performs frequency statistics and confidence weighting on the classification results of multiple segments to obtain the final classification result.

[0095] Specifically, the real-time noise scene classification system of this application is applicable to the sound processor of a cochlear implant.

[0096] In some implementations, the feature extraction module specifically includes: The power spectrum calculation submodule is used to pre-emphasize and window the input audio signal, divide it into frames with a fixed frame length and frame shift, apply a fast Fourier transform to each frame to calculate the power spectrum, and retain the positive frequency part.

[0097] The critical band energy calculation submodule is used to divide the power spectrum into N critical bands on the frequency axis, calculate the bandwidth of each critical band, and sum the power spectrum energies within the band to obtain the critical band energy.

[0098] The critical band ratio calculation submodule is used to calculate the ratio of the energy of each critical band to the total energy of each frame, i.e., the critical band ratio, to obtain the N-dimensional critical band ratio feature.

[0099] The temporal feature extraction submodule is used to calculate the short-time energy of each frame as the sum of squares of the intra-frame signals and to perform logarithmic normalization on the short-time energy features. It is also used to calculate the short-time zero-crossing rate of each frame, which is the ratio of the number of symbol changes within the frame to the frame length.

[0100] The feature fusion submodule is used to concatenate the N-dimensional critical band ratio feature, the short-time energy feature, and the short-time zero-crossing rate feature to obtain the hybrid time-frequency domain feature.

[0101] In other embodiments, the feature extraction module further includes: a preprocessing module, used to perform preprocessing operations on the acquired audio signal, specifically including: The pre-emphasis processing unit is used to apply a pre-emphasis filter to compensate for the attenuation of high-frequency components in the audio signal, thereby improving signal clarity.

[0102] The framing and windowing unit is used to divide the processed continuous signal into short time frames to capture the time-varying characteristics of the signal. A window function is applied to the short time frames to reduce spectral leakage.

[0103] Based on the above embodiments, this application discloses another embodiment of a real-time noise scene classification system, which further includes modules for model training and classification, real-time processing architecture, and hardware deployment. The steps performed by the above modules are the same as those of any of the above real-time noise scene classification methods. This embodiment will not repeat the description.

[0104] Based on the same technical concept, this application also discloses a cochlear implant device, wherein the cochlear implant includes a real-time noise scene classification system as described in any of the above embodiments.

[0105] The real-time noise scene classification system, system, and cochlear implant device of this application have the same technical concept, and the technical details of the embodiments of the two are applicable to each other. In order to reduce repetition, they will not be repeated here.

[0106] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of program modules is merely an example. In practical applications, the above functions can be assigned to different program modules as needed, that is, the internal structure of the device can be divided into different program units or modules to complete all or part of the functions described above. The program modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one processing unit. The integrated unit can be implemented in hardware or as a software program unit. Furthermore, the specific names of the program modules are only for easy differentiation and are not intended to limit the scope of protection of this application.

[0107] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

Claims

1. A real-time noise scene classification method, characterized in that, include: For the acquired audio signal, a specified number of mixed time-frequency domain features are extracted as classification criteria; the mixed time-frequency domain features consist of multi-dimensional critical band ratio features, as well as short-time energy features and short-time zero-crossing rate features; The feature vectors of the hybrid time-frequency domain features are input into the trained support vector machine classification model to obtain the classification results of the audio segments. For the long audio signals input in real time, frame-by-frame feature extraction and segmented average accumulation processing are performed, and frequency statistics and confidence weighting are performed on the classification results of multiple segments to obtain the final classification results.

2. The real-time noise scene classification method as described in claim 1, characterized in that, The aforementioned method of extracting a specified number of mixed time-frequency domain features as a classification criterion for the acquired audio signal specifically includes: After pre-emphasis and windowing of the input audio signal, it is divided into frames with fixed frame length and frame shift. The power spectrum of each frame signal is calculated by applying Fast Fourier Transform and the positive frequency part is retained. Divide the power spectrum into N critical bands on the frequency axis, calculate the bandwidth of each critical band, and sum the power spectrum energies within the bands to obtain the critical band energy. Calculate the ratio of each critical band energy to the total energy of each frame, i.e., the critical band ratio, to obtain the N-dimensional critical band ratio feature; The short-time energy of each frame is calculated as the sum of squares of the intra-frame signal, and the short-time energy characteristics are log-normalized. Calculate the short-time zero-crossing rate for each frame, where the short-time zero-crossing rate is the ratio of the number of symbol changes within the frame to the frame length; The hybrid time-frequency domain feature is obtained by concatenating the N-dimensional critical band ratio feature, the short-time energy feature, and the short-time zero-crossing rate feature.

3. The real-time noise scene classification method as described in claim 2, characterized in that: The critical band energy is calculated using the following formula; ； ； in, Let m be the calculated bandwidth of the b-th critical band; b is the critical band index, and m is the frame index. The center frequency of the b-th critical band; Let be the total energy of the m-th frame signal within the b-th critical band. Let k be the physical frequency of the frequency point; The lower limit frequency; The upper limit frequency; Let k be the power spectral density value of the m-th frame signal at frequency k. Calculated using the following formula Critical band ratio; ； in, The critical band ratio of the b-th critical band in the m-th frame; The total energy of the signal in the m-th frame; To prevent division by zero of extremely small constants.

4. A real-time noise scene classification method as described in claim 1 or 2, characterized in that: The short-time energy is calculated using the following formula; ； ； in, Let m be the short-time energy of the m-th frame. The short-time energy feature of the m-th frame after log-normalization is used for feature vector concatenation; m is the frame index; L is the frame length; n is the index of the sample within the frame; This represents the nth sample value in the m-th frame of the signal; The short-time zero-crossing rate is calculated using the following formula; ； in, Let be the short-time zero-crossing rate of the m-th frame; This is the (n-1)th sample value in the m-th frame of the signal.

5. The real-time noise scene classification method as described in claim 1, characterized in that, For the acquired audio signals, a specified number of mixed time-frequency domain features are extracted as the classification criterion. This previously included: The preprocessing operation for the acquired audio signal specifically includes: A pre-emphasis filter is applied to compensate for the attenuation of high-frequency components in the audio signal, thereby improving signal clarity; The processed continuous signal is divided into short time frames to capture the time-varying characteristics of the signal; a window function is applied to the short time frames to reduce spectral leakage.

6. The real-time noise scene classification method as described in claim 1, characterized in that, Also includes: Audio samples from various environments were collected in real-world scenarios to construct an initial dataset. Gain processing was applied to the audio samples from transitional scenarios to expand the category boundaries. The dataset was then divided into training, validation, and test sets according to a set ratio. Feature extraction is performed on audio samples in the training and validation sets respectively to obtain corresponding feature matrices and add label vectors for training and hyperparameter optimization of the support vector machine model; The same features are extracted from the audio samples in the test set for model performance evaluation.

7. The real-time noise scene classification method as described in claim 6, characterized in that, Also includes: A support vector machine model is trained using a multi-class one-to-one encoding strategy with radial basis function kernel function. Specifically, the following steps are included: Based on the feature matrix and the label vector, a one-to-one binary classifier is constructed for each scenario; The hyperparameters are automatically adjusted using Bayesian optimization; the optimization objective is to minimize the classification error; and finally, a final model that can be used for classification in real-world scenarios is obtained by training based on the optimal parameter combination.

8. A real-time noise scene classification method as described in claim 6 or 7, characterized in that, Also includes: A real-time frame-segmentation framework is added to the trained support vector machine model. For long audio signals input in real time, frame segmentation is performed with fixed frame length and frame shift; The hybrid time-frequency domain features are extracted frame by frame, and the average of the frame number corresponding to 1-2 seconds is taken to obtain the multi-dimensional features of the segment. The features are then input into the scene classification model for classification, and the intermediate scene results and confidence scores are output. Process each segment until the long audio signal ends, count the frequency of each scene and calculate the weighted probability: ； Where i represents the index of the three scene types. Let j be the frequency of segment j. The confidence level of this segment; based on The maximum value is output as the final scene classification result.

9. A real-time noise scene classification system, characterized in that, The system is used to implement the steps of the real-time noise scene classification method according to any one of claims 1-8; Includes: a signal acquisition module for acquiring audio signals; The feature extraction module is used to extract a specified number of mixed time-frequency domain features as a classification criterion for the audio signal; the mixed time-frequency domain features consist of multi-dimensional critical band ratio features, as well as short-time energy features and short-time zero-crossing rate features; The model classification module is used to input the feature vector of the hybrid time-frequency domain features into the trained support vector machine classification model to obtain the classification result of the audio segment; for the long audio signal input in real time, it performs frame-by-frame feature extraction and segment average accumulation processing, and performs frequency statistics and confidence weighting on the classification results of multiple segments to obtain the final classification result.

10. A cochlear implant device, characterized in that, The cochlear implant device includes a real-time noise scene classification system as described in claim 9.

Images