Voice recognition method and system based on artificial intelligence

Abstract

The invention provides a voice recognition method and system based on artificial intelligence, and the method employs the artificial intelligence technology, improves the voice recognition effect of a model through the continuous training of a recognition model, achieves the accurate recognition of a voice through the trained recognition model, and achieves the recognition of the voice through tone filtering after the voice is recognized. The target sound is accurately searched from the plurality of sound sources, so that the problem of background noise interference in the recording process is solved. According to the invention, the target sound can be reserved, and other noises can be filtered out, so that the recognition accuracy is improved.

Description

Technical Field

[0001] This invention relates to the field of voice recognition technology, and more specifically, to a voice recognition method and system based on artificial intelligence. Background Technology

[0002] Product testing is an essential part of the headphone manufacturing process. During testing, if defective headphones are found, such as a clicking sound or other abnormal audio feedback when pressing the non-force area, the testing department needs to record this abnormal audio feedback and analyze its waveform, peaks, amplitude, duration, and other information to determine the cause of the defect.

[0003] However, due to the lack of professional and high-precision noise reduction equipment, the recording environment is often filled with various external noises, such as traffic noise and human voices. These noises can seriously interfere with the quality and accuracy of sound recording, making sound recognition and analysis difficult. Most existing sound recognition systems cannot effectively cope with sound interference in multi-source environments.

[0004] Therefore, a new type of voice recognition system is needed that can accurately identify and filter in such an environment. Summary of the Invention

[0005] In view of the above problems, the purpose of this invention is to provide a voice recognition method and system based on artificial intelligence to achieve the above.

[0006] According to one aspect of the present invention, an artificial intelligence-based voice recognition method is provided, applied to an electronic device, comprising:

[0007] The pre-recorded audio data is subjected to noise reduction preprocessing to reduce noise in the audio data;

[0008] The preset CNN model is trained based on the noise-reduced preprocessed audio data. The recognition ability of the CNN model is optimized by continuously adjusting the parameters of the CNN model.

[0009] The trained CNN model is used to process the real-time collected sound data and output preliminary recognition results.

[0010] The initial recognition results are subjected to timbre filtering based on a preset threshold to obtain the target sound.

[0011] In addition, an optional technical solution is that the data format of the pre-recorded audio data includes WAV, AIPP, MP3, and AQF.

[0012] In addition, an optional technical solution is that the noise reduction preprocessing includes:

[0013] Normalization is used to scale the amplitude of the original audio signal to a preset range.

[0014] Trimming or padding is used to trim the original audio signal, which has been scaled to a preset range, to a preset length; or, to add extra audio signal to the trimmed audio signal by padding to prevent audio loss caused by trimming.

[0015] Downsampling is used to reduce the sampling rate of an audio signal that has been cropped or padded to a preset length in order to reduce the MIPS of the signal processing.

[0016] In addition, an optional technical solution is to include a feature extraction step before training the CNN model, which is used to convert the noise-reduced sound data into a feature representation suitable for inputting into the CNN model, and to obtain the Mel-spectral coefficients, temporal features and spectrograms in the sound data.

[0017] In addition, an optional technical solution is to further perform data augmentation processing on the extracted audio data after the feature extraction step, and then train a preset CNN model based on the augmented audio data; wherein, the data augmentation processing includes:

[0018] Random cropping is used to randomly crop audio segments of a fixed length from the noise-reduced preprocessed audio data.

[0019] Noise addition processing is used to add random noise to the noise-reduced preprocessed audio data;

[0020] Time warp processing is used to randomly change the playback speed of the pre-processed noise-reduced audio data;

[0021] Frequency distortion processing is used to randomly change the frequency of the pre-processed noise-reducing audio data.

[0022] In addition, an optional technical solution is to further include, after data augmentation, extracting features from the augmented audio data and converting the extracted features into a tensor; and during the training of the CNN model, stacking the features of multiple audio files into batches and inputting them into the CNN model.

[0023] In addition, an optional technical solution is to continuously adjust the parameters of the CNN model using the following adjustment scheme:

[0024] Hyperparameter tuning is used to adjust the parameters of the CNN model by adjusting the learning rate, batch size, number of layers, and number of neurons per layer;

[0025] Regularization is used to reduce overfitting by adding regularization terms or using Dropout layers to adjust the parameters of the CNN model;

[0026] Model structure adjustment is used to adjust the parameters of the CNN model by adding or removing convolutional layers and fully connected layers.

[0027] In addition, an optional technical solution is that the timbre filtering process includes:

[0028] Based on the Mel-spectral cepstral coefficients of the preliminary identification result, the timbre features of the preliminary identification result are extracted; and the sound spectral characteristics of the preliminary identification result are captured, and the spectral center of the preliminary identification result is determined based on the spectral centroid of the sound spectral characteristics.

[0029] Based on the timbre characteristics and the spectral center, the preliminary identification results are subjected to timbre filtering processing to obtain the target sound.

[0030] On the other hand, the present invention also provides an artificial intelligence-based voice recognition system, which performs voice recognition using the artificial intelligence-based voice recognition method described above; including:

[0031] A noise reduction preprocessing unit is used to perform noise reduction preprocessing on pre-recorded audio data to reduce noise in the audio data;

[0032] The model training unit is used to train a preset CNN model based on the noise-reduced preprocessed audio data, and to optimize the recognition ability of the CNN model by continuously adjusting the parameters of the CNN model.

[0033] The recognition and processing unit is used to process the real-time acquired sound data through a trained CNN model and output preliminary recognition results.

[0034] The timbre filtering unit is used to perform timbre filtering processing on the preliminary recognition results according to a preset threshold to obtain the target sound.

[0035] In addition, an optional technical solution includes a feature extraction unit and a data augmentation unit; wherein, the feature extraction unit is used to convert the noise-reduced audio data into a feature representation suitable for input to a CNN model, and to obtain Mel-frequency coefficients, temporal features, and spectrograms from the audio data; the augmentation unit is used to perform data augmentation processing on the audio data extracted by the feature extraction unit; wherein, the data augmentation processing includes:

[0036] Random cropping is used to randomly crop audio segments of a fixed length from the noise-reduced preprocessed audio data.

[0037] Noise addition processing is used to add random noise to the noise-reduced preprocessed audio data;

[0038] Time warp processing is used to randomly change the playback speed of the pre-processed noise-reduced audio data;

[0039] Frequency distortion processing is used to randomly change the frequency of the pre-processed noise-reducing audio data.

[0040] The aforementioned AI-based voice recognition method and system utilize artificial intelligence technology to continuously train the recognition model to improve its voice recognition performance. The trained model then accurately identifies sounds and, after recognition, filters the sounds by timbre to precisely locate the target sound from multiple sound sources, thus solving the problem of background noise interference during recording. This invention preserves the target sound while filtering out other noise, thereby improving recognition accuracy.

[0041] To achieve the foregoing and related objectives, one or more aspects of the invention include the features that will be described in detail below. The following description and accompanying drawings illustrate certain exemplary aspects of the invention. However, these aspects indicate only a few of the various ways in which the principles of the invention can be used. Furthermore, the invention is intended to encompass all such aspects and their equivalents. Attached Figure Description

[0042] Other objects and results of the invention will become more apparent and readily understood with reference to the following description taken in conjunction with the accompanying drawings. In the drawings:

[0043] Figure 1 This is a flowchart illustrating an artificial intelligence-based voice recognition method according to an embodiment of the present invention.

[0044] Figure 2 This is a logic block diagram of an artificial intelligence-based voice recognition system according to an embodiment of the present invention.

[0045] In all the accompanying drawings, the same reference numerals indicate similar or corresponding features or functions. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] In the following description, numerous specific details are set forth for illustrative purposes and to provide a thorough understanding of one or more embodiments. However, it will be apparent that these embodiments may also be implemented without these specific details. In other instances, well-known structures and devices are shown in block diagram form for ease of description of one or more embodiments.

[0048] To better illustrate the technical solution of the present invention, some of the technical terms involved in the present invention will be briefly explained below.

[0049] CNN (Convolutional Neural Network) is a class of feedforward neural networks that incorporate convolutional computations and have a deep structure. It is one of the representative algorithms of deep learning. Convolutional neural networks have representation learning capabilities, enabling them to perform shift-invariant classification of input information according to their hierarchical structure.

[0050] CNNs include input layers and hidden layers. The hidden layers of convolutional neural networks include three common types of architectures: convolutional layers, pooling layers, and fully connected layers.

[0051] To address the aforementioned problems in existing technologies, this invention provides a sound recognition method and system based on artificial intelligence. By utilizing artificial intelligence technology, the recognition model is continuously trained to improve its sound recognition performance. The trained sound recognition model is then used to accurately identify sounds. After the sound is identified, timbre filtering is applied to precisely locate the target sound from multiple sound sources, thereby solving the problem of background noise interference during recording.

[0052] To describe in detail the artificial intelligence-based voice recognition method and system of the present invention, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0053] Figure 1 The flowchart of an artificial intelligence-based voice recognition method according to an embodiment of the present invention is shown. Figure 2 A logic block diagram of an artificial intelligence-based voice recognition system according to an embodiment of the present invention is shown.

[0054] The AI-based voice recognition method provided in this embodiment is applied to electronic devices, which can be any headphones, headsets, or wearable devices capable of playing audio data, such as VR (Virtual Reality) devices, AR (Augmented Reality) devices, MR (Mixed Reality) devices, XR (Extended Reality) devices, ER (Emulated Reality) devices, or headsets or wearable devices with audio players or multimedia players, such as watches with various functions, headphones, etc.

[0055] like Figure 1 As shown, the voice recognition method based on artificial intelligence mainly includes the following steps:

[0056] S120: Perform noise reduction preprocessing on the pre-recorded audio data to reduce noise in the audio data.

[0057] The pre-recorded audio data, as the raw audio signal, can be read from a file system or dataset. The audio data format can be WAV, AIFF, MP3, AQF, etc. Because Python has a large number of mathematical and method libraries and excellent performance in data processing, in one specific embodiment of this invention, the librosa library of Python is used to read and process the audio data to complete this task.

[0058] The AI-based sound recognition method provided in this embodiment can effectively cope with sound interference in multi-source environments by processing noise reduction and highlighting the target sound source. After obtaining the pre-recorded sound data, the original audio signal first needs to be pre-processed for noise reduction to reduce the noise in the original audio signal.

[0059] Specifically, as an example, in order to obtain a more accurate noise reduction effect, in a specific embodiment of the present invention, the following three noise reduction processing methods are used in sequence for noise reduction preprocessing: normalization processing, cropping or padding processing, and downsampling processing.

[0060] The normalization process is used to scale the amplitude of the original audio signal to a preset range in order to eliminate the dimensional differences between features.

[0061] Trimming or padding is used to trim the original audio signal, which has been scaled to a preset range, to a preset length to increase the amount of data. Furthermore, additional audio signals can be added to the trimmed audio signal by padding to prevent audio loss caused by trimming.

[0062] Downsampling is used to reduce the sampling rate of audio signals that have been cropped or padded to a preset length, thereby reducing the MIPS (Million Instructions Per Second) of signal processing.

[0063] For example, the amplitude of the original audio signal can be scaled to the range of [-1, 1], as shown in the following formula:

[0064] x norm [n] = x[n] / max(∣x[n]∣)

[0065] This formula represents the normalization of the signal x[n]. Specifically, first, the maximum value max(|x[n]|) among all elements of the signal x[n] is calculated. Then, for each time n, x[n] is divided by this maximum value max(|x[n]|) to obtain the normalized audio signal x. norm [n].

[0066] This normalization process limits the amplitude range of the original audio signal to [-1, 1], thereby standardizing the amplitude of the original audio signal to facilitate subsequent processing and analysis.

[0067] Then, the scaled audio signal is trimmed to a specified length L (for example, removing parts exceeding L, or padding audio signals shorter than L to length L). If the original audio signal length N > L, the trimming formula is:

[0068]

[0069] Where x[n] represents the nth sample of the original audio signal, x clip [n] represents the cropped signal sample value, N represents the length of the original audio signal, and L represents the length of the cropped or padded audio signal.

[0070] Finally, the audio signal, trimmed or padded to a preset length, is downsampled to reduce the MIPS of the signal processing. Specifically, as an example, during the processing, the audio signal trimmed or padded to a preset length is first filtered to avoid signal aliasing; then, the filtered audio signal is decimated at a preset sampling frequency to obtain the noise-reduced preprocessed audio data.

[0071] S140: Train a preset CNN model based on the noise-reduced preprocessed sound data, and continuously adjust the parameters of the CNN model to optimize its recognition ability.

[0072] The noise-reduced audio data is divided into training data and testing data to train and test the preset CNN model respectively.

[0073] In another specific embodiment of the present invention, in order to improve the quality of training data, after noise reduction preprocessing, a feature extraction step is further included to convert the noise-reduced preprocessed sound data into a feature representation suitable for inputting into a CNN model, so as to obtain Mel spectral coefficients, temporal features and spectrograms in the sound data.

[0074] In addition, to improve the performance and generalization ability of the CNN model and highlight the target sound, after feature extraction, the extracted sound data is augmented to increase the diversity of the dataset, and then the preset CNN model is trained based on the augmented sound data.

[0075] In one specific embodiment of the present invention, the data augmentation process includes random cropping, noise addition, time warping, and frequency warping.

[0076] More specifically, during the data augmentation process, random cropping, noise addition, time warping, and frequency warping are performed one by one.

[0077] The random cropping process is used to randomly crop fixed-length audio segments from the noise-reduced preprocessed audio data. Specifically, as an example, a random starting point t is selected to begin cropping the original audio signal, where t = random integer in [0, NL]; then the original audio signal is cropped using the following formula:

[0078] x clip [n] = x[t+n] for 0≤n <L

[0079] Where x[n] represents the nth sample of the original audio signal, x clip [n] represents the cropped signal sample value, t is the cropping start point, x[t+n] represents the nth sample of the original audio signal after the cropping start point t, and L is the length of the cropped audio signal.

[0080] Noise addition processing is used to add random noise to the pre-processed audio data after noise reduction. Specifically, as an example, the following formula is used for noise addition processing:

[0081] x noisy [n] = x[n] + α·n[n]

[0082] in:

[0083] x[n] is the nth sample value of the audio signal after noise reduction preprocessing.

[0084] x noisy [n] represents the sample value of the audio signal after adding noise.

[0085] n[n] is the added random noise signal.

[0086] α is the noise amplitude (the control factor for noise intensity).

[0087] Time warping is used to randomly change the playback speed of the pre-processed, noise-reduced audio data; for example, by randomly slowing down or speeding up the playback speed through time stretching. In a specific embodiment of the invention, the time stretching factor is set to β (where β>1 for acceleration and β<1 for deceleration), then the new audio signal obtained by the time warping formula is:

[0088]

[0089] in:

[0090] x[n] is the nth sample value of the audio signal after noise reduction preprocessing. warp [n] is the time-warped audio signal sample value, and β is the time scaling factor (scaling ratio).

[0091] Frequency warping is used to randomly change the frequency (pitch) of the noise-reduced preprocessed audio data. Specifically, as an example, frequency warping is implemented using the following four steps:

[0092] S131: Fast Fourier Transform (FFT) transforms the noise-reduced preprocessed audio data from the time domain to the frequency domain by performing a Fast Fourier Transform on the audio data.

[0093] S132: Random frequency offset, generated by randomly selecting a frequency offset within a certain range (e.g., ±5% to 20% of the original frequency).

[0094] S133: Frequency modification, used to modify each frequency component in the frequency domain according to the frequency offset randomly generated in the aforementioned step S132.

[0095] S134: Inverse Fast Fourier Transform (IFFT), used to perform an inverse fast Fourier transform on frequency domain data after frequency modification, converting it back to the time domain to obtain frequency-distorted sound data.

[0096] The process of training a pre-defined CNN model using enhanced audio data includes:

[0097] S141: Extract features from the enhanced audio data to convert the audio waveform data into a feature representation suitable for input into the CNN model;

[0098] Specifically, as an example, a feature representation suitable for inputting a CNN model can be obtained by performing a Short-Time Fourier Transform (STFT) on the enhanced audio data.

[0099] In one specific embodiment of the present invention, the librosa library in Python is used to calculate the spectrogram of the enhanced audio data. This spectrogram is used as the feature representation input to the CNN model. The spectrogram is a two-dimensional matrix, where the horizontal axis represents time and the vertical axis represents frequency, and each element represents the energy intensity at the corresponding time and frequency.

[0100] S142: Convert the extracted features into a tensor to unify the data format and fully utilize the computational power of the deep learning framework of the CNN model.

[0101] S143: Stack the features of multiple audio files into a batch and input them into the CNN model;

[0102] Stacking inputs in batches enables parallel computing and hardware optimization, effectively improving the computational efficiency of CNN models, stabilizing gradient descent, and benefiting memory management and training stability. Specifically, as an example, feature matrices can be stacked into a four-dimensional tensor through feature extraction and data standardization, and then input into the CNN model.

[0103] S144: Continuously adjust the parameters of the CNN model.

[0104] Specifically, as an example, the continuous adjustment of the parameters of the CNN model can employ adjustment schemes such as hyperparameter adjustment, regularization, and model structure adjustment. These adjustment schemes can be flexibly selected according to specific adjustment needs, and can also be repeatedly used or arbitrarily combined.

[0105] Hyperparameter tuning is used to adjust the parameters of the CNN model by adjusting the learning rate, batch size, number of layers, and number of neurons per layer.

[0106] The number of layers and the number of neurons in each layer can be expressed as:

[0107]

[0108] Among them, H out H represents the height of the output feature map. in The height of the input feature map is represented by F, the size of the filter is represented by P, and the stride is represented by S.

[0109] Regularization is used to reduce overfitting by adding regularization terms or using Dropout layers to adjust the parameters of the CNN model;

[0110] Model structure adjustment is used to adjust the parameters of the CNN model by adding or removing convolutional layers and fully connected layers.

[0111] S160: The trained CNN model is used to process the real-time collected sound data and output preliminary recognition results.

[0112] S180: Perform timbre filtering on the preliminary recognition results according to a preset threshold to obtain the target sound.

[0113] Specifically, as an example, the above timbre filtering process includes:

[0114] S181: Extract the timbre features of the preliminary identification result based on the Mel-frequency cepstral coefficients (MFCCs) of the preliminary identification result; and capture the sound spectral characteristics of the preliminary identification result, and determine the spectral center of the preliminary identification result based on the spectral centroid of the sound spectral characteristics.

[0115] S182: Based on the timbre characteristics and the spectral center, perform timbre filtering processing on the preliminary identification result to obtain the target sound.

[0116] It should be noted that the AI-based voice recognition method provided by this invention can be widely used in various recording scenarios, such as meeting minutes, speech recognition, and audio monitoring.

[0117] As can be seen from the above embodiments, the artificial intelligence-based sound recognition method provided by the present invention utilizes artificial intelligence technology to continuously train the recognition model to improve the sound recognition effect of the model. Then, the trained recognition model is used to accurately recognize the sound. After the sound is recognized, the target sound is accurately found from multiple sound sources through timbre filtering, so as to solve the problem of background noise interference during the recording process.

[0118] Corresponding to the above-mentioned AI-based voice recognition, the present invention also provides an AI-based voice recognition system to perform voice recognition using the above-mentioned AI-based voice recognition method. Figure 2 This is a logic block diagram of an artificial intelligence-based voice recognition system according to an embodiment of the present invention.

[0119] like Figure 2 As shown, the artificial intelligence-based voice recognition system 200 provided in this embodiment includes:

[0120] The noise reduction preprocessing unit 210 is used to perform noise reduction preprocessing on the pre-recorded sound data in order to reduce the noise in the sound data;

[0121] The model training unit 220 is used to train a preset CNN model based on the noise-reduced preprocessed sound data, and to optimize the recognition ability of the CNN model by continuously adjusting the parameters of the CNN model.

[0122] The recognition processing unit 230 is used to process the real-time acquired sound data through a trained CNN model and output preliminary recognition results.

[0123] The timbre filtering unit 240 is used to perform timbre filtering processing on the preliminary recognition result according to a preset threshold to obtain the target sound.

[0124] Furthermore, the AI-based voice recognition system also includes a feature extraction unit and a data enhancement processing unit, wherein...

[0125] The feature extraction unit is used to convert sound waveform data into feature representations suitable for inputting into the CNN model, and to obtain the Mel-frequency coefficients, temporal features and spectrograms of the sound.

[0126] A data enhancement processing unit is used to perform data enhancement processing on the noise reduction preprocessed audio data; wherein, the data enhancement processing includes:

[0127] Random cropping is used to randomly crop audio segments of a fixed length from the noise-reduced preprocessed audio data.

[0128] Noise addition processing is used to add random noise to the noise-reduced preprocessed audio data;

[0129] Time warp processing is used to randomly change the playback speed of the pre-processed noise-reduced audio data;

[0130] Frequency distortion processing is used to randomly change the frequency of the pre-processed noise-reducing audio data.

[0131] More specific implementations of the aforementioned AI-based voice recognition system can be found in the foregoing description of embodiments of AI-based voice recognition methods, and will not be detailed here.

[0132] The artificial intelligence-based voice recognition method and system according to the present invention have been described above by way of example with reference to the accompanying drawings. However, those skilled in the art should understand that various modifications can be made to the artificial intelligence-based voice recognition method and system proposed in the present invention without departing from the scope of the invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A voice recognition method based on artificial intelligence, applied to electronic devices, characterized in that, include: The pre-recorded audio data is subjected to noise reduction preprocessing to reduce noise in the audio data; The preset CNN model is trained based on the noise-reduced preprocessed audio data. The recognition ability of the CNN model is optimized by continuously adjusting the parameters of the CNN model. The trained CNN model is used to process the real-time collected sound data and output preliminary recognition results. The initial recognition results are subjected to timbre filtering based on a preset threshold to obtain the target sound.

2. The voice recognition method based on artificial intelligence according to claim 1, characterized in that, The data formats of the pre-recorded audio data include WAV, AIPP, MP3, and AQF.

3. The voice recognition method based on artificial intelligence according to claim 2, characterized in that, The noise reduction preprocessing includes: Normalization is used to scale the amplitude of the original audio signal to a preset range. Trimming or padding is used to trim the original audio signal, which has been scaled to a preset range, to a preset length; or, to add extra audio signal to the trimmed audio signal by padding to prevent audio loss caused by trimming. Downsampling is used to reduce the sampling rate of an audio signal that has been cropped or padded to a preset length in order to reduce the MIPS of the signal processing.

4. The artificial intelligence-based voice recognition method according to claim 3, characterized in that, Before training the CNN model, a feature extraction step is also included to convert the noise-reduced sound data into a feature representation suitable for inputting into the CNN model, and to obtain the Mel-spectral coefficients, temporal features and spectrograms in the sound data.

5. The artificial intelligence-based voice recognition method according to claim 4, characterized in that, Following the feature extraction step, the extracted audio data undergoes data augmentation processing, and then a pre-defined CNN model is trained based on the augmented audio data; wherein the data augmentation processing includes: Random cropping is used to randomly crop audio segments of a fixed length from the noise-reduced preprocessed audio data. Noise addition processing is used to add random noise to the pre-processed audio data after noise reduction. Time warp processing is used to randomly change the playback speed of the pre-processed noise-reduced audio data; Frequency distortion processing is used to randomly change the frequency of the pre-processed noise-reducing audio data.

6. The voice recognition method based on artificial intelligence according to claim 5, characterized in that, Following data augmentation, the following is also included: Feature extraction is performed on the enhanced audio data, and the extracted features are converted into a tensor. During the training of the CNN model, the features of multiple audio files are stacked into batches and input into the CNN model.

7. The voice recognition method based on artificial intelligence according to claim 1, characterized in that, The parameters of the CNN model are continuously adjusted using the following adjustment scheme: Hyperparameter tuning is used to adjust the parameters of the CNN model by adjusting the learning rate, batch size, number of layers, and number of neurons per layer; Regularization is used to reduce overfitting by adding regularization terms or using Dropout layers to adjust the parameters of the CNN model; Model structure adjustment is used to adjust the parameters of the CNN model by adding or removing convolutional layers and fully connected layers.

8. The artificial intelligence-based voice recognition method according to claim 1, characterized in that, The timbre filtering process includes: Based on the Mel-spectral cepstral coefficients of the preliminary identification result, the timbre features of the preliminary identification result are extracted; and the sound spectral characteristics of the preliminary identification result are captured, and the spectral center of the preliminary identification result is determined based on the spectral centroid of the sound spectral characteristics. Based on the timbre characteristics and the spectral center, the preliminary identification results are subjected to timbre filtering processing to obtain the target sound.

9. An artificial intelligence-based voice recognition system, comprising performing voice recognition using the artificial intelligence-based voice recognition method as described in any one of claims 1-8; including: A noise reduction preprocessing unit is used to perform noise reduction preprocessing on pre-recorded audio data to reduce noise in the audio data; The model training unit is used to train a preset CNN model based on the noise-reduced preprocessed audio data, and to optimize the recognition ability of the CNN model by continuously adjusting the parameters of the CNN model. The recognition and processing unit is used to process the real-time acquired sound data through a trained CNN model and output preliminary recognition results. The timbre filtering unit is used to perform timbre filtering processing on the preliminary recognition results according to a preset threshold to obtain the target sound.

10. The voice recognition system based on artificial intelligence according to claim 9, characterized in that, It also includes a feature extraction unit and a data augmentation processing unit; among which, The feature extraction unit is used to convert the noise-reduced sound data into a feature representation suitable for inputting into a CNN model, and to obtain the Mel spectrum coefficients, temporal features and spectrograms in the sound data; The enhancement processing unit is used to perform data enhancement processing on the audio data extracted by the feature extraction unit; wherein, the data enhancement processing includes: Random cropping is used to randomly crop audio segments of a fixed length from the noise-reduced preprocessed audio data. Noise addition processing is used to add random noise to the pre-processed audio data after noise reduction. Time warp processing is used to randomly change the playback speed of the pre-processed noise-reduced audio data; Frequency distortion processing is used to randomly change the frequency of the pre-processed noise-reducing audio data.

Images