Laser pickup voice recognition degree improving method and device, computer equipment and storage medium
A deep learning-based method for laser sound pickup enhances speech recognition by removing noise and reverberation through sequential processing, improving clarity and adaptability across diverse environments.
Patent Information
- Application Number
- CN202411221284.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-02
- Publication Date
- 2024-12-17
AI Technical Summary
Existing laser sound pickup technologies face challenges in improving speech recognition due to noise level, equipment vibration, surface characteristics, and reverberation, with traditional methods lacking adaptability and efficiency, and deep learning-based methods requiring significant computational resources and high processing power.
A method and device for enhancing speech recognition using a deep learning-based approach that includes white noise removal, environmental noise removal, bandpass filtering, dereverberation, and high-frequency recovery models to improve speech clarity and adaptability.
The method effectively removes noise and reverberation, enhances high-frequency components, and improves speech recognition quality and adaptability across various environments.
Smart Images
Figure CN119152868A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of laser sound pickup, and in particular to a method, device, computer equipment and storage medium for improving the speech recognition of laser sound pickup. Background Art
[0002] Laser sound pickup technology uses the coherence principle of lasers to collect surface vibration signals caused by sound propagation and then recover audio signals. However, improving speech recognition is a key challenge in this field because the quality of laser interference signals is affected by many factors, such as the noise level of the laser and the acquisition equipment, equipment vibration, surface characteristics of the object, and reverberation.
[0003] At present, a variety of processing methods are used to improve the recognition of laser pickup equipment. Traditional signal processing methods mainly rely on signal processing and statistical modeling to process the characteristics of the collected sound signals. For example, spectral subtraction and Wiener filtering are widely used. Although these methods are reliable, they have relatively poor adaptability and are difficult to flexibly adapt to different environments and scenarios, so they have limited performance when dealing with complex acquisition conditions and signal loss problems. In contrast, deep learning-based processing methods use advanced neural network architectures to improve speech recognition by learning complex speech features. These methods have achieved remarkable results in scenarios such as network calls, voice collection, and headphone playback, effectively improving speech clarity and recognition. However, as the complexity of the model increases, these technologies require a large amount of training data and high computing power support, which limits their application in situations where resources are limited or real-time processing is required. In addition, current methods mainly focus on noise elimination and fail to fully address the special challenges of laser pickup, such as reverberation and signal loss.
[0004] Therefore, it is necessary to design a new method to improve the recognition of laser sound pickup and solve the problems of traditional noise reduction technology in scene adaptability and flexibility of use. Summary of the invention
[0005] The purpose of the present invention is to overcome the defects of the prior art and provide a method, device, computer equipment and storage medium for improving the speech recognition of laser pickup.
[0006] To achieve the above object, the present invention adopts the following technical solution: a method for improving the speech recognition of laser pickup, comprising:
[0007] Get the original audio;
[0008] Inputting the original audio into a white noise removal model to remove white noise to obtain a first target audio;
[0009] Inputting the first target audio into an environmental noise removal model to remove environmental noise to obtain a second target audio;
[0010] Performing bandpass filtering on the second target audio to obtain a third target audio;
[0011] Inputting the third target audio into a dereverberation model to remove the reverberation signal to obtain a fourth target audio;
[0012] Inputting the fourth target audio into a high frequency recovery model for high frequency recovery processing to obtain a fifth target audio;
[0013] The fifth target audio is output.
[0014] A further technical solution is: the white noise removal model is obtained by simulating white noise signals and / or actually collecting white noise signals as sample sets to train a deep learning network.
[0015] Its further technical solution is: the environmental noise removal model is obtained by training a deep learning network by simulating environmental noise signals and / or noise signals collected according to actual scenes and / or noise audio signals that are not collected in actual scenes but collected as sample sets.
[0016] A further technical solution is: performing bandpass filtering on the second target audio to obtain a third target audio, comprising:
[0017] The second target audio is subjected to bandpass filtering to limit the frequency domain range of the second target audio to a specific valid signal interval, so as to obtain a third target audio.
[0018] A further technical solution is: the reverberation signal includes the reflection and reverberation of the sound signal caused by the laser pickup target itself or the surrounding environment and the signal distortion caused by the vibration characteristics of the laser pickup target itself.
[0019] Its further technical solution is: the dereverberation model is obtained by training a deep learning network using simulation analysis data of the vibration characteristics of the laser pickup target object and the actually collected audio data as a sample set.
[0020] Its further technical solution is: the high-frequency recovery model is obtained by training a generative deep learning network by bandpass processing the standard audio of a specific language as a sample set.
[0021] The present invention also provides a device for improving the speech recognition of laser pickup, comprising:
[0022] An original audio acquisition unit, used for acquiring original audio;
[0023] A white noise removal unit, configured to input the original audio into a white noise removal model to remove white noise, so as to obtain a first target audio;
[0024] an environmental noise removal unit, configured to input the first target audio into an environmental noise removal model to remove environmental noise, so as to obtain a second target audio;
[0025] a bandpass filtering unit, configured to perform bandpass filtering on the second target audio to obtain a third target audio;
[0026] a reverberation signal removal unit, configured to input the third target audio into a dereverberation model to remove the reverberation signal, so as to obtain a fourth target audio;
[0027] A high frequency recovery processing unit, used for inputting the fourth target audio into a high frequency recovery model for high frequency recovery processing to obtain a fifth target audio;
[0028] An output unit, configured to output the fifth target audio.
[0029] The present invention further provides a computer device, which includes a memory and a processor. The memory stores a computer program, and the processor implements the above method when executing the computer program.
[0030] The present invention also provides a storage medium, wherein the storage medium stores a computer program, and the computer program implements the above method when executed by a processor.
[0031] Compared with the prior art, the present invention has the following beneficial effects: the present invention removes white noise from the original audio, then removes environmental noise, and then processes the audio through bandpass filtering to obtain signals within a specific frequency range. Subsequently, a dereverberation model is applied to remove the reverberation signal, and finally the audio is processed through a high-frequency recovery model to enhance the high-frequency part, and the optimized audio is output, thereby improving the recognition of laser pickup and solving the problems of traditional noise reduction technology in scene adaptability and flexibility of use.
[0032] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying any creative work.
[0034] Figure 1 A schematic diagram of an application scenario of the method for improving speech recognition of laser pickup provided by an embodiment of the present invention;
[0035] Figure 2A schematic diagram of a process flow of a method for improving speech recognition of laser pickup provided by an embodiment of the present invention;
[0036] Figure 3 A schematic block diagram of a device for improving speech recognition of laser pickup provided by an embodiment of the present invention;
[0037] Figure 4 A schematic block diagram of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0038] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0039] It should be understood that when used in this specification and the appended claims, the terms "include" and "comprises" indicate the presence of described features, integers, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.
[0040] It should also be understood that the terms used in this specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.
[0041] It should be further understood that the term "and / or" used in the present description and the appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.
[0042] See also Figure 1 and Figure 2 , Figure 1 A schematic diagram of an application scenario of the method for improving speech recognition of laser pickup provided by an embodiment of the present invention. Figure 2A schematic flow chart of a method for improving the speech recognition of laser pickup provided in an embodiment of the present invention. The method for improving the speech recognition of laser pickup is applied in a server. The server interacts with the laser pickup to pick up the original audio, and then sequentially inputs the white noise removal model and the environmental noise removal model, performs bandpass filtering, and then inputs the dereverberation model and the high-frequency recovery model to form the final target audio, thereby improving the recognition of laser pickup and solving the problems of traditional noise reduction technology in scene adaptability and flexibility of use.
[0043] Figure 2 FIG. 1 is a flow chart of a method for improving speech recognition of laser pickup provided by an embodiment of the present invention. Figure 2 As shown, the method includes the following steps S110 to S170.
[0044] S110: Obtain original audio.
[0045] In this embodiment, the original audio refers to the vibration signal collected by aiming at a specific target object through a laser pickup device. These vibration signals can be generated by the target object itself or by the vibration of the target object caused by the surrounding environment sound.
[0046] These vibration signals captured by the laser pickup device are processed by voice conversion technology to form original audio signals.
[0047] S120: Input the original audio into a white noise removal model to remove white noise, so as to obtain a first target audio.
[0048] In this embodiment, the first target audio refers to the remaining audio after removing white noise from the original audio.
[0049] Specifically, white noise is noise with uniform frequency distribution and equal intensity, usually introduced by laser pickup equipment or other electronic devices. Removing white noise helps improve the clarity and audibility of audio, ensuring that subsequent processing steps can more effectively process the target signal.
[0050] The white noise removal model is obtained by simulating white noise signals and / or actually collecting white noise signals as sample sets to train a deep learning network, wherein the deep learning network includes but is not limited to a recurrent neural network (RNN), a long short-term memory network (LSTM), a gated recurrent unit (GRU), and a Transformer.
[0051] Specifically, training a white noise removal model requires a large amount of training data, which can be simulated white noise signals or actually collected white noise mixed signals. Simulated white noise can be obtained by random generation, while the actually collected white noise signal is the real white noise in the environment collected by a specific device; the input to the model is usually the pre-processed original audio signal, including a mixture of the target signal and the noise signal. The output is the audio signal with the main white noise component removed after being processed by the model; during the training process, it is very important to choose an appropriate loss function to measure the difference between the model output and the target output (clear signal after noise removal). Commonly used loss functions include mean square error (MSE), mean absolute error (MAE), etc., which are used to guide the optimization of model parameters and the training process. In the training stage, the model parameters need to be iterated through a large amount of data so that the model can learn how to effectively distinguish the target signal from the white noise and remove the latter from the former. In this process, the model will adjust its own weights and parameters to minimize the white noise component in the input signal; after training, the model needs to be evaluated and optimized. Evaluation can be performed through performance on the validation set or test set, including indicators such as improvement in signal-to-noise ratio (SNR) and clarity. Optimization may involve adjusting the model structure, hyperparameter adjustment, and other means to further improve the model's generalization ability and actual application effect.
[0052] S130: Input the first target audio into an environmental noise removal model to remove environmental noise, so as to obtain a second target audio.
[0053] In this embodiment, the second target audio refers to the audio remaining after removing the environmental noise from the first target audio. Environmental noise includes various sounds from the surrounding environment, such as wind noise, mechanical noise, etc. Removing environmental noise can make the audio cleaner and highlight the sound characteristics of the target object itself.
[0054] The environmental noise removal model is obtained by training a deep learning network by simulating environmental noise signals and / or noise signals collected according to actual scenes and / or noise audio signals that are not collected in actual scenes but collected as sample sets.
[0055] Specifically, first, choose a deep learning architecture suitable for processing audio signals. Common choices include recurrent neural networks (RNN), long short-term memory networks (LSTM), gated recurrent units (GRU), Transformers, etc. These architectures can effectively capture the timing characteristics and long-term dependencies of audio signals.
[0056] Training data is the key to training an environmental noise removal model. Data can be obtained from multiple sources:
[0057] Simulate environmental noise signals: Simulate noise in various environments, such as street noise, wind noise, etc., by randomly generating them.
[0058] Noise signals collected in actual scenarios: actual noise signals collected in specific environments using professional equipment.
[0059] Collected noisy audio signals: Noise samples collected from public databases or scene-specific recordings to expand the diversity of training datasets.
[0060] The collected raw audio and noise signals are preprocessed. Preprocessing may include audio waveform normalization, time-frequency conversion (such as short-time Fourier transform), slicing and segmentation, etc., to facilitate model input and processing.
[0061] Design the input and output formats of the model. The typical input is a mixed audio signal (including the target signal and noise), and the output is a processed clear target signal (noise component removed).
[0062] Choose an appropriate loss function to measure the difference between the model output and the target output. Common loss functions include mean square error (MSE), mean absolute error (MAE), etc., which are selected according to the needs of the specific task.
[0063] The model is trained using the prepared training data set. During the training process, the model optimizes parameters through the back propagation algorithm to minimize the selected loss function. The goal of the training is to enable the model to effectively learn how to accurately remove environmental noise from the input signal.
[0064] After training, the model needs to be evaluated. The evaluation can be performed on a validation set or a test set, and usually includes calculating indicators such as signal-to-noise ratio (SNR) and auditory evaluation (such as subjective hearing test). Based on the evaluation results, it may be necessary to adjust the model structure, hyperparameters, or optimize the training strategy to further improve the performance and generalization ability of the model.
[0065] After training, the environmental noise removal model can be integrated into practical applications, such as audio editing software, voice communication systems, etc. During deployment, the computational efficiency, real-time performance, and stability of the model need to be considered to ensure that it can be effectively run and applied in real environments.
[0066] S140: Perform bandpass filtering on the second target audio to obtain a third target audio.
[0067] In this embodiment, the third target audio refers to audio formed after performing bandpass filtering on the second target audio.
[0068] Bandpass processing is a common digital signal processing method, which aims to limit the target audio signal to a specific frequency domain range. The selection of this frequency domain range is based on the effective signal range collected in the current environment to ensure that the audio part that can be clearly distinguished by the human ear is retained. Specifically, bandpass processing uses filters or related technologies to allow only signals within a specific frequency range to pass, while signals of other frequencies are eliminated. This can effectively enhance the audibility and clarity of the target signal, and is suitable for various application scenarios such as speech recognition and audio enhancement.
[0069] Specifically, the second target audio is subjected to bandpass filtering, and the frequency domain range of the second target audio is limited to a specific valid signal interval, so as to obtain the third target audio.
[0070] This step can further highlight the frequency characteristics of the target signal and improve the clarity and intelligibility of the audio.
[0071] S150: Input the third target audio into a dereverberation model to remove the reverberation signal to obtain a fourth target audio.
[0072] In this embodiment, the fourth target audio refers to the audio remaining after the reverberation signal in the third target audio is removed.
[0073] The reverberation signal includes the reflection and reverberation of the sound signal caused by the laser pickup target object itself or the surrounding environment, and the signal distortion caused by the vibration characteristics of the laser pickup target object itself.
[0074] Specifically, the dereverberation model is obtained by training a deep learning network using simulation analysis data of the vibration characteristics of the laser pickup target object and actually collected audio data as a sample set.
[0075] Theoretical data on the impact of the target object's vibration characteristics on the audio signal are generated through simulation software or physical simulation tools. These data are used in the initial training and verification stages of the model.
[0076] Actual audio data: Audio samples collected from real scenes, including the actual mixture of target signals and reverberation signals. These data are the main source for training deep learning models, ensuring the accuracy and robustness of the models in real situations.
[0077] Choose an appropriate deep learning architecture, such as a recurrent neural network (RNN), transformer, or other architecture suitable for time series processing. These architectures are able to handle the temporal characteristics of audio signals and learn complex feature representations during training. Adjust the model structure and hyperparameters to maximize the performance and generalization ability of the model. This includes the choice of loss function, adjustment of learning rate, and application of regularization techniques to prevent overfitting.
[0078] Preprocess the collected audio data, including feature extraction, normalization, and time / frequency domain conversion. Use the prepared training data set to train the deep learning model. During the training process, the model updates parameters through the back-propagation algorithm and optimizer (such as Adam) to minimize the error between the predicted target signal and the actual target signal. Monitor the performance of the model during the training process and use the validation set to evaluate and debug the model. This helps to discover and solve potential training problems such as overfitting or underfitting.
[0079] Use independent test datasets to evaluate the performance of the trained model, including the quality assessment of the audio after dereverberation and the reconstruction accuracy of the target signal. Convert the trained model into a format that can be used in practical applications, such as TensorFlow model or ONNX model, so as to be deployed and integrated on different platforms.
[0080] S160: Input the fourth target audio into a high frequency recovery model for high frequency recovery processing to obtain a fifth target audio.
[0081] In this embodiment, the fifth target audio refers to the audio formed after the fourth target audio is processed by high-frequency restoration. High-frequency restoration is intended to enhance the missing or attenuated high-frequency part of the audio through a generative deep learning model or other technologies. This helps make the audio more natural and realistic, and improves its auditory effect and perceived quality.
[0082] Specifically, the high-frequency recovery model is obtained by training a generative deep learning network by performing bandpass processing on standard audio of a specific language as a sample set.
[0083] In this embodiment, a large amount of standard audio data is collected, such as speech clips of Mandarin Chinese. These data need to be bandpass filtered, that is, the parts with too high and too low frequencies are removed and concentrated in the medium frequency range, so that the model can focus on learning the recovery of high-frequency details. The collected audio data is preprocessed, including feature extraction, time-frequency domain conversion, etc., so that the model can effectively process and learn.
[0084] Choose a generative model architecture suitable for audio processing tasks, such as GAN (Generative Adversarial Network) or Transformer. These architectures are able to learn complex audio feature representations during training and generate high-quality high-frequency signals. Adjust the model's structure, loss function, and hyperparameters to maximize the model's performance and generalization ability. This includes optimizing the generator and discriminator (for GAN), as well as applying techniques such as attention mechanisms and layer normalization (for Transformer).
[0085] The generative deep learning model is trained using the prepared bandpass processed audio dataset. During the training process, the model learns to recover high-frequency details by minimizing the difference between the restored audio and the original high-frequency audio. Appropriate optimization strategies, such as the Adam optimizer, are used to adjust the model parameters to improve training efficiency and convergence speed.
[0086] Use an independent validation set to evaluate the performance of the trained model in recovering high-frequency information. Evaluation metrics can include signal-to-noise ratio (SNR), spectrogram comparison, etc. to verify whether the high-frequency details generated by the model are accurate and clear. Adjust and improve the model based on the evaluation results to further optimize its ability to recover high-frequency information.
[0087] Convert the trained generative model into a format that can be used in practical applications, such as TensorFlow model or ONNX model. Deploy the model to the audio processing system for real-time or batch processing of low-quality audio, thereby improving the auditory perception and recognition accuracy of the audio.
[0088] The high-frequency restoration model is able to effectively learn and restore lost high-frequency details from bandpass processed audio, thereby improving audio quality and clarity.
[0089] S170: Output the fifth target audio.
[0090] After the above processing steps, the fifth target audio is an optimized and enhanced audio signal. The output audio is usually clearer and easier to understand than the original audio, and is suitable for application scenarios that require high-quality audio, such as speech recognition, remote communication, audio recording, etc.
[0091] For the training process of the models mentioned above, the order and selection of the models can be flexibly adjusted according to the specific scenarios and requirements during the model training process. For example, the white noise removal model can be combined with the environmental noise removal model to simplify the processing steps. In some cases, if the impact of reverberation is small, the processing of the reverberation removal model can be omitted and the high-frequency recovery model can be directly applied. This flexibility allows the model to be optimized and adjusted according to actual application requirements to improve processing efficiency and effect.
[0092] The method of this embodiment is based on the traditional processing method and deep learning technology, and extracts effective audio signals through multi-step processing, including the removal of white noise, environmental noise and reverberation. Subsequently, the deep learning generative AI technology is used to restore the high frequency of these effective signals to produce clearer and easier to identify audio signals.
[0093] This implementation method not only takes into account the particularity of laser pickup signal acquisition and provides a corresponding signal processing solution, but also fully combines the technical advantages in the field of deep learning. By adopting a simple step-by-step small model, the ideal processing effect can be effectively achieved, thereby improving the audio quality and clarity, making the audio more usable and accurate in the application. In response to the problems existing in the laser pickup process, it is proposed to adopt a step-by-step processing method according to the physical process of pickup to eliminate the unfavorable factors in laser pickup, such as high-frequency loss, reverberation, and noise of the laser itself. Compared with the data of other training models, more targeted data can be used for training, thereby achieving an ideal speech enhancement effect, thereby improving the recognition of laser pickup.
[0094] In summary, the method of this embodiment processes audio signals in multiple stages, gradually removes different types of noise and improves sound quality, providing higher scene adaptability and flexibility of use. Traditional noise reduction technology may be limited to specific types of noise or difficult to adapt to complex environmental changes, while this method utilizes a variety of models and processing steps to more accurately process different noise sources. This flexibility enables it to perform better in various practical application scenarios and can effectively improve the quality and efficiency of speech recognition and audio processing.
[0095] The above-mentioned method for improving the speech recognition of laser pickup removes white noise from the original audio, then removes the ambient noise, and then processes the audio through bandpass filtering to obtain signals within a specific frequency range. Subsequently, a dereverberation model is applied to remove the reverberation signal, and finally the audio is processed through a high-frequency recovery model to enhance the high-frequency part and output optimized audio, thereby improving the recognition of laser pickup and solving the problems of traditional noise reduction technology in scene adaptability and flexibility of use.
[0096] Figure 3 is a schematic block diagram of a laser pickup speech recognition improvement device 300 provided in an embodiment of the present invention. Figure 3 As shown, corresponding to the above laser pickup voice recognition improvement method, the present invention also provides a laser pickup voice recognition improvement device 300. The laser pickup voice recognition improvement device 300 includes a unit for executing the above laser pickup voice recognition improvement method, and the device can be configured in a server. Figure 3 The laser pickup speech recognition improving device 300 includes an original audio acquisition unit 301, a white noise removal unit 302, an environmental noise removal unit 303, a bandpass filtering unit 304, a reverberation signal removal unit 305, a high frequency restoration processing unit 306 and an output unit 307.
[0097] The original audio acquisition unit 301 is used to acquire the original audio; the white noise removal unit 302 is used to input the original audio into the white noise removal model to remove the white noise to obtain the first target audio; the environmental noise removal unit 303 is used to input the first target audio into the environmental noise removal model to remove the environmental noise to obtain the second target audio; the bandpass filtering unit 304 is used to perform bandpass filtering on the second target audio to obtain the third target audio; the reverberation signal removal unit 305 is used to input the third target audio into the dereverberation model to remove the reverberation signal to obtain the fourth target audio; the high-frequency recovery processing unit 306 is used to input the fourth target audio into the high-frequency recovery model for high-frequency recovery processing to obtain the fifth target audio; the output unit 307 is used to output the fifth target audio.
[0098] In one embodiment, the bandpass filtering unit 304 is configured to perform bandpass filtering on the second target audio to limit the frequency domain range of the second target audio to a specific valid signal interval to obtain the third target audio.
[0099] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the above-mentioned laser pickup speech recognition improvement device 300 and each unit can refer to the corresponding description in the aforementioned method embodiment, and for the convenience and brevity of description, it will not be repeated here.
[0100] The above-mentioned laser pickup speech recognition improving device 300 can be implemented in the form of a computer program, and the computer program can be used in the following ways: Figure 4 Runs on the computer device shown.
[0101] See also Figure 4 , Figure 4 1 is a schematic block diagram of a computer device provided in an embodiment of the present application. The computer device 500 may be an independent server or a server cluster composed of multiple servers.
[0102] See also Figure 4 The computer device 500 includes a processor 502 , a memory and a network interface 505 connected via a system bus 501 , wherein the memory may include a non-volatile storage medium 503 and an internal memory 504 .
[0103] The non-volatile storage medium 503 can store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions, and when the program instructions are executed, the processor 502 can execute a method for improving the speech recognition of laser pickup.
[0104] The processor 502 is used to provide computing and control capabilities to support the operation of the entire computer device 500 .
[0105] The internal memory 504 provides an environment for the operation of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute a method for improving the speech recognition of laser pickup.
[0106] The network interface 505 is used to communicate with other devices over the network. Figure 4 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present application, and does not constitute a limitation on the computer device 500 to which the solution of the present application is applied. The specific computer device 500 may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.
[0107] The processor 502 is used to run the computer program 5032 stored in the memory to implement the following steps:
[0108] Acquire original audio; input the original audio into a white noise removal model to remove white noise to obtain a first target audio; input the first target audio into an environmental noise removal model to remove environmental noise to obtain a second target audio; perform bandpass filtering on the second target audio to obtain a third target audio; input the third target audio into a dereverberation model to remove reverberation signals to obtain a fourth target audio; input the fourth target audio into a high frequency recovery model to perform high frequency recovery processing to obtain a fifth target audio; and output the fifth target audio.
[0109] The white noise removal model is obtained by simulating white noise signals and / or actually collecting white noise signals as sample sets to train a deep learning network.
[0110] The environmental noise removal model is obtained by training a deep learning network by simulating environmental noise signals and / or noise signals collected according to actual scenes and / or noise audio signals that are not collected in actual scenes but collected as sample sets.
[0111] The reverberation signal includes the reflection and reverberation of the sound signal caused by the laser pickup target object itself or the surrounding environment, and the signal distortion caused by the vibration characteristics of the laser pickup target object itself.
[0112] The dereverberation model is obtained by training a deep learning network using simulation analysis data of the vibration characteristics of the laser pickup target and actually collected audio data as a sample set.
[0113] The high-frequency recovery model is obtained by training a generative deep learning network using the standard audio of a specific language as a sample set after bandpass processing.
[0114] In one embodiment, when the processor 502 implements the step of performing bandpass filtering on the second target audio to obtain the third target audio, the processor 502 specifically implements the following steps:
[0115] The second target audio is subjected to bandpass filtering to limit the frequency domain range of the second target audio to a specific valid signal interval, so as to obtain a third target audio.
[0116] It should be understood that in the embodiment of the present application, the processor 502 may be a central processing unit (CPU), and the processor 502 may also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.
[0117] It can be understood by those skilled in the art that all or part of the processes in the method for implementing the above embodiment can be completed by instructing the relevant hardware through a computer program. The computer program includes program instructions, and the computer program can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiment of the above method.
[0118] Therefore, the present invention also provides a storage medium. The storage medium may be a computer-readable storage medium. The storage medium stores a computer program, wherein when the computer program is executed by a processor, the processor executes the following steps:
[0119] Acquire original audio; input the original audio into a white noise removal model to remove white noise to obtain a first target audio; input the first target audio into an environmental noise removal model to remove environmental noise to obtain a second target audio; perform bandpass filtering on the second target audio to obtain a third target audio; input the third target audio into a dereverberation model to remove reverberation signals to obtain a fourth target audio; input the fourth target audio into a high frequency recovery model to perform high frequency recovery processing to obtain a fifth target audio; and output the fifth target audio.
[0120] The white noise removal model is obtained by simulating white noise signals and / or actually collecting white noise signals as sample sets to train a deep learning network.
[0121] The environmental noise removal model is obtained by training a deep learning network by simulating environmental noise signals and / or noise signals collected according to actual scenes and / or noise audio signals that are not collected in actual scenes but collected as sample sets.
[0122] The reverberation signal includes the reflection and reverberation of the sound signal caused by the laser pickup target object itself or the surrounding environment, and the signal distortion caused by the vibration characteristics of the laser pickup target object itself.
[0123] The dereverberation model is obtained by training a deep learning network using simulation analysis data of the vibration characteristics of the laser pickup target and actually collected audio data as a sample set.
[0124] The high-frequency recovery model is obtained by training a generative deep learning network using the standard audio of a specific language as a sample set after bandpass processing.
[0125] In one embodiment, when the processor executes the computer program to implement the step of performing bandpass filtering on the second target audio to obtain the third target audio, the processor specifically implements the following steps:
[0126] The second target audio is subjected to bandpass filtering to limit the frequency domain range of the second target audio to a specific valid signal interval, so as to obtain a third target audio.
[0127] The storage medium may be a USB flash drive, a mobile hard disk, a read-only memory (ROM), a magnetic disk, or an optical disk, etc., which are computer-readable storage media that can store program codes.
[0128] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.
[0129] In the several embodiments provided by the present invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of each unit is only a logical function division, and there may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed.
[0130] The steps in the method of the embodiment of the present invention can be adjusted in order, combined and deleted according to actual needs. The units in the device of the embodiment of the present invention can be combined, divided and deleted according to actual needs. In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.
[0131] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions for a computer device (which can be a personal computer, terminal, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention.
[0132] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.
Claims
1. A method for improving speech recognition of laser pickup, characterized in that: include: Get the original audio; Inputting the original audio into a white noise removal model to remove white noise to obtain a first target audio; Inputting the first target audio into an environmental noise removal model to remove environmental noise to obtain a second target audio; Performing bandpass filtering on the second target audio to obtain a third target audio; Inputting the third target audio into a dereverberation model to remove the reverberation signal to obtain a fourth target audio; Inputting the fourth target audio into a high frequency recovery model for high frequency recovery processing to obtain a fifth target audio; The fifth target audio is output.
2. The method for improving the speech recognition of laser pickup according to claim 1, characterized in that: The white noise removal model is obtained by simulating white noise signals and / or actually collecting white noise signals as sample sets to train a deep learning network.
3. The method for improving speech recognition of laser pickup according to claim 1, characterized in that: The environmental noise removal model is obtained by training a deep learning network by simulating environmental noise signals and / or noise signals collected according to actual scenes and / or noise audio signals that are not collected in actual scenes but collected as sample sets.
4. The method for improving speech recognition of laser pickup according to claim 1, characterized in that: The step of performing bandpass filtering on the second target audio to obtain a third target audio includes: The second target audio is subjected to bandpass filtering to limit the frequency domain range of the second target audio to a specific valid signal interval, so as to obtain a third target audio.
5. The method for improving speech recognition of laser pickup according to claim 1, characterized in that: The reverberation signal includes the reflection and reverberation of the sound signal caused by the laser pickup target object itself or the surrounding environment, and the signal distortion caused by the vibration characteristics of the laser pickup target object itself.
6. The method for improving the speech recognition of laser pickup according to claim 5, characterized in that: The dereverberation model is obtained by training a deep learning network using simulation analysis data of the vibration characteristics of the laser pickup target object and actually collected audio data as a sample set.
7. The method for improving speech recognition of laser pickup according to claim 1, characterized in that: The high-frequency recovery model is obtained by training a generative deep learning network using the standard audio of a specific language as a sample set after bandpass processing.
8. A device for improving the speech recognition of laser pickup, characterized in that: include: An original audio acquisition unit, used for acquiring original audio; A white noise removal unit, configured to input the original audio into a white noise removal model to remove white noise, so as to obtain a first target audio; an environmental noise removal unit, configured to input the first target audio into an environmental noise removal model to remove environmental noise, so as to obtain a second target audio; a bandpass filtering unit, configured to perform bandpass filtering on the second target audio to obtain a third target audio; a reverberation signal removal unit, configured to input the third target audio into a dereverberation model to remove the reverberation signal, so as to obtain a fourth target audio; A high frequency recovery processing unit, used for inputting the fourth target audio into a high frequency recovery model for high frequency recovery processing to obtain a fifth target audio; An output unit, configured to output the fifth target audio.
9. A computer device, characterized in that: The computer device comprises a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the computer program, the method according to any one of claims 1 to 7 is implemented.
10. A storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 7 is implemented.