Sound spectrum mode recognizer
By combining audio acquisition and spectrogram conversion with a neural network model, the problems of low efficiency and low accuracy in automated identification during acoustic product testing have been solved, achieving efficient and accurate abnormal sound identification and automated judgment.
Patent Information
- Application Number
- CN202511622446.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-07
- Publication Date
- 2026-02-03
AI Technical Summary
Existing technologies for acoustic product testing suffer from problems such as low testing efficiency, inconsistent subjective judgment, high labor costs, high false negative rate, and inability to achieve automated production. Furthermore, existing algorithms have low recognition accuracy.
By employing an audio acquisition module, a spectrogram module, and a neural network module, audio data is converted into spectrograms, and features are extracted and classified using a neural network. A training set is constructed and the neural network model is optimized to achieve automated abnormal sound recognition.
It improves the accuracy and efficiency of abnormal sound recognition, reduces manual costs, achieves automated judgment, and reduces the false judgment rate.
Smart Images
Figure CN121462964A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of acoustic testing, and in particular to a spectral pattern recognizer. Background Technology
[0002] When acoustic products such as loudspeakers and microphones undergo factory testing, they typically require testing with specific test signals. These test signals are either played through a loudspeaker and received by a reference microphone in a test fixture, or played through a test reference loudspeaker and received by a device on a microphone or sensor. It is usually necessary to manually determine whether the sound of the reference microphone is abnormal, and thus determine whether the tested device is functioning properly.
[0003] Current technologies rely on manual judgment of abnormal sounds during testing, which suffers from low testing efficiency, inconsistent subjective judgment standards, high labor costs, missed detections or substandard quality, and the inability to automate production. Algorithm-based methods for identifying abnormal sounds mainly fall into three categories: template matching, probabilistic statistical models, and discriminative models. Template matching methods extract feature parameters that characterize sound differences from training sample sounds as reference templates. During recognition, the feature vector of the sound is first extracted and then compared with the features of the reference template. Classic methods include Dynamic Time Warping (DTW) and Vector Quantization (VQ). Probabilistic statistical model-based methods assume that sound signals are stable within a frame after reasonable framing. Parameters in statistical probability methods, such as variance, mean, and probability density function, can be used to reflect the statistical regularity of the sound. Probability is used to model and judge based on the probability distribution. Classic methods include Hidden Markov Models (HMMs) and Gaussian Mixture Models (GMMs). Discriminative model-based methods include K-means, Support Vector Machines (SVMs), and neural networks. Among them, neural networks have made significant progress in recent years. Due to their powerful adaptive, self-learning, and self-organizing capabilities, they can classify and identify abnormal sound features. Neural networks take normal and abnormal sound characteristics as input, use labels to guide the training process, and utilize the powerful classification capabilities of neural networks to obtain the identification results at the network output.
[0004] The above methods can all recognize audio, but they are not well adapted to industrial inspection and recognition applications and have problems with low recognition accuracy. Summary of the Invention
[0005] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a spectrogram pattern recognizer with high accuracy and high detection efficiency in identifying abnormal sound signals.
[0006] The technical solution adopted in this invention is as follows: This invention includes an audio acquisition module, a spectrogram module, a neural network module, and a result output module; The audio acquisition module is used to acquire sound and convert it into an audio signal; The spectrogram module converts the audio signal acquired by the audio acquisition module into a spectrogram; The neural network module extracts features from the spectrogram, thereby extracting features of normal and abnormal sounds. The specific extraction process includes classifying and labeling the spectrogram in sequence, setting parameters to generate a training set, selecting a neural network model, setting hyperparameters, inputting the training set into the neural network model to start training, and saving the network model. The result output module retrieves the network model stored in the neural network module to compare and judge the spectrum of the target image converted by the spectrogram module, and outputs the judgment result.
[0007] As can be seen from the above scheme, the audio acquisition module converts the sound output by the product under test during operation into audio data, and then the spectrogram module converts the audio data into a spectrogram. A neural network model extracts features and performs classification and annotation on the spectrogram, thereby analyzing the acoustic information of the product under test. Abnormal signals are identified by comparing the spectrogram with a normal one, and the result output module outputs the judgment result. By converting audio data into a spectrogram for annotation and training, the accuracy of abnormal sound recognition is effectively improved. Furthermore, it enables accurate testing and reduces interference when testing electronic devices. In addition, computer-based judgment avoids the drawbacks of manual judgment, reduces labor costs, and improves judgment accuracy.
[0008] In a preferred embodiment, the audio acquisition module includes a signal output unit, a reference speaker unit, a reference microphone unit, and a signal acquisition unit. The signal output unit outputs a test audio signal to the speaker under test or the reference speaker unit, and the signal acquisition unit acquires analog signals from the microphone under test or the reference microphone unit and converts and outputs them.
[0009] The acoustic spectrum pattern recognizer also includes the following specific working steps: Step S1: Connect the standard product and the abnormal product to the audio acquisition module respectively, and collect audio data through the audio acquisition module to obtain multiple sets of normal and abnormal audio data; Step S2: The spectrogram module converts the audio data in terms of dynamic range, frequency, and time resolution to obtain a spectrogram; Step S3: The neural network module performs classification and annotation operations on the obtained spectrograms and obtains a training dataset; Step S4: The neural network module selects the neural network model type and constructs the neural network model; Step S5: Input the training dataset obtained in step S3 into the neural network model constructed in step S4, set the hyperparameters, train the neural network model, compare the prediction results output by the neural network model with the labels, update the parameters of each layer using the gradient descent algorithm, and optimize the model; after training is complete, save the trained neural network model to obtain the audio anomaly recognition model. Step S6: During testing, the audio acquisition module collects audio data, and the spectrogram module converts the audio data to be identified into a spectrogram. The spectrogram is then input into the audio anomaly recognition model. The result output module determines whether the audio segment is abnormal. If there is an anomaly, the result output module provides the location of the anomaly based on the spectrogram comparison.
[0010] A further preferred embodiment is that the hyperparameters in step S5 include batch-size, epoch, and learning rate.
[0011] A further preferred embodiment is that, in step S6, the spectrogram of the current test is added to the training dataset at each test, thereby increasing the quantity and diversity of the training data.
[0012] A further preferred embodiment is that the neural network model type in step S4 is semantic segmentation or semantic classification. Attached Figure Description
[0013] Figure 1 This is a system flowchart of the present invention. Detailed Implementation
[0014] like Figure 1As shown, in this embodiment, the present invention includes an audio acquisition module 1, a spectrogram module 2, a neural network module 3, and a result output module 4. The audio acquisition module 1 is used to acquire sound and convert it into an audio signal. The audio acquisition module 1 includes a signal output unit 11, a reference speaker unit 12, a reference microphone unit 13, and a signal acquisition unit 14. The signal output unit 11 outputs a test audio signal to the speaker under test or the reference speaker unit 12. The signal acquisition unit 14 acquires analog signals from the microphone under test or the reference microphone unit 13 and converts them for output. When the product under test is a speaker, sound is acquired through the reference microphone unit 13. When the product under test is a microphone, audio is played through the reference speaker unit 12 and acquired by the microphone under test. The analog signals acquired by the reference microphone unit 13 or the microphone under test are converted into digital signals by the signal acquisition unit 14. The spectrogram module 2 converts the digital audio signals acquired by the audio acquisition module 1 into a spectrogram, thereby converting the audio signals into a spectrum that can be analyzed and from which feature information can be extracted. The neural network module 3 extracts features from the spectrogram by recording normal and abnormal sounds. The specific extraction process includes classifying and labeling the spectrogram sequentially, setting parameters to generate a training set, selecting a neural network model, setting hyperparameters, inputting the training set into the neural network model to start training, and saving the network model. The trained neural network model enables automated comparison. The result output module 4 retrieves the network model saved by the neural network module 3 to compare and judge the spectrogram to be tested converted by the spectrogram module 2, and outputs the judgment result.
[0015] The present invention also includes the following specific working steps: Step S1: Connect the standard product and the abnormal product to the audio acquisition module 1 respectively, and collect audio data through the audio acquisition module 1 to obtain multiple sets of normal and abnormal audio data; Step S2: The audio data is dynamically adjusted and its frequency and time resolution are converted using the spectrogram module 2 to obtain a spectrogram; Step S3: The neural network module 3 performs classification and annotation operations on the obtained spectrograms and obtains a training dataset; Step S4: The neural network module 3 selects a neural network model type and constructs a neural network model; in this embodiment, the neural network model type is semantic segmentation; Step S5: Input the training dataset obtained in step S3 into the neural network model constructed in step S4, set the hyperparameters, train the neural network model, compare the prediction results output by the neural network model with the labels, update the parameters of each layer using the gradient descent algorithm, and optimize the model; after training is complete, save the trained neural network model to obtain the audio anomaly recognition model. Step S6: During testing, the audio acquisition module 1 connects to the product under test and acquires audio data. Then, the spectrogram module 2 converts the audio data to be identified into a spectrogram. The result output module 4 calls the trained neural network model as the audio anomaly recognition model and inputs the spectrogram to be tested into the audio anomaly recognition model. The result output module 4 determines whether the audio segment is abnormal through the audio anomaly recognition model. If it is normal, the result outputs feedback. If there is an anomaly, the result output module 4 compares the spectrograms through the audio anomaly recognition model and provides the location of the anomaly in the audio segment.
[0016] In this embodiment, the hyperparameters mentioned in step S5 include batch-size, epoch, and learning rate.
[0017] In this embodiment, in step S6, the current spectrogram to be tested is added to the training dataset during each test. After the comparative analysis is completed, the results are obtained, and classification is performed based on the results. Then, training is performed to increase the quantity and diversity of training data.
[0018] In step S4, the neural network model type can also be a classification model, etc.
[0019] Although the embodiments of the present invention are described with reference to actual solutions, they do not constitute a limitation on the meaning of the present invention. Modifications to the embodiments and combinations with other solutions based on this specification will be obvious to those skilled in the art.
Claims
1. A spectrogram pattern recognizer, characterized in that, It includes an audio acquisition module (1), a spectrogram module (2), a neural network module (3), and a result output module (4). The audio acquisition module (1) is used to acquire sound and convert it into an audio signal; The spectrogram module (2) converts the audio signal acquired by the audio acquisition module (1) into a spectrogram; The neural network module (3) performs feature extraction on the spectrogram, and then extracts the features of normal and abnormal sounds. The specific extraction process includes performing classification labeling on the spectrogram in sequence, setting parameters to generate a training set, selecting a neural network model, setting hyperparameters, inputting the training set into the neural network model to start training, and saving the network model. The result output module (4) retrieves the network model saved in the neural network module (3) to compare and judge the spectrum of the target spectrum converted by the spectrogram module (2), and outputs the judgment result.
2. The acoustic spectrum pattern recognizer according to claim 1, characterized in that: The audio acquisition module (1) includes a signal output unit (11), a reference speaker unit (12), a reference microphone unit (13), and a signal acquisition unit (14). The signal output unit (11) outputs a test audio signal to the speaker under test or the reference speaker unit (12). The signal acquisition unit (14) acquires analog signals from the microphone under test or the reference microphone unit (13) and converts and outputs them.
3. The acoustic spectrum pattern recognizer according to claim 1, characterized in that, It also includes the following specific working steps: Step S1: Connect the standard product and the abnormal product to the audio acquisition module (1) respectively, and collect audio data through the audio acquisition module (1) to obtain multiple sets of normal and abnormal audio data; Step S2: The spectrogram module (2) converts the audio data in terms of dynamic range, frequency and time resolution to obtain a spectrogram; Step S3: The neural network module (3) performs classification and annotation operations on the obtained spectrograms and obtains a training dataset; Step S4: The neural network module (3) selects the neural network model type and constructs the neural network model; Step S5: Input the training dataset obtained in step S3 into the neural network model constructed in step S4, set the hyperparameters, train the neural network model, compare the prediction results output by the neural network model with the labels, update the parameters of each layer using the gradient descent algorithm, and optimize the model. After training is complete, the trained neural network model is saved to obtain the audio anomaly recognition model. Step S6: During the test, the audio acquisition module (1) acquires audio data, and the spectrogram module (2) converts the audio data to be identified into a spectrogram. The spectrogram is then input into the audio anomaly identification model. The result output module (4) determines whether the audio segment is abnormal. If there is an anomaly, the result output module (4) provides the abnormal location of the audio segment based on the spectrogram comparison.
4. The acoustic spectrum pattern recognizer according to claim 1, characterized in that: The hyperparameters mentioned in step S5 include batch size, epoch, and learning rate.
5. The acoustic spectrum pattern recognizer according to claim 3, characterized in that: In step S6, the spectrogram of the current test is added to the training dataset at each test, increasing the quantity and diversity of the training data.
6. The acoustic spectrum pattern recognizer according to claim 3, characterized in that: In step S4, the neural network model type is semantic segmentation or semantic classification.