Low-latency band-based pitch detection method, device and equipment based on neural network

Abstract

The present invention relates to a low-latency band-based pitch detection method, device and equipment based on a neural network. The method comprises: obtaining an audio signal. Constructing a pitch detection model based on a neural network, inputting the audio signal as sample training data into the filter of the pitch detection model for filtering, and obtaining multiple sub-band audio vectors. Each sub-band audio vector is subjected to feature extraction through a preset number of convolution layers in the neural network to obtain a first pitch sub-band. The second pitch signal is obtained by artificially correcting the audio signal. The first pitch sub-band and the second pitch signal are input into the pitch detection model for model training to obtain a pre-trained pitch detection model. The pre-trained pitch detection model proofreads the audio signal according to the loss function and outputs a low-latency pitch signal. The use of this method can significantly reduce the delay of pitch detection and improve the real-time performance and time domain resolution fineness of various pitch-related processing.

Description

Technical Field

[0001] The present invention relates to the field of audio detection technology, and in particular to a low-delay band-based pitch detection method, device and equipment based on a neural network. Background Art

[0002] Pitch detection, a core problem in audio processing and music information retrieval, involves accurately identifying the fundamental frequency of a note or sound from an audio signal. This technology has broad applications such as automatic music transcription, music teaching, sound analysis, and vocal training assistance.

[0003] Early pitch detection methods were mostly based on time-domain and frequency-domain analysis. Representative algorithms include autocorrelation, fast Fourier transform, and cepstrum. These algorithms analyze a whole block of audio data to estimate its pitch. This leads to latency issues, as they require a complete block of data for analysis. Furthermore, the time-domain resolution is limited, as they extract features from a single block of data.

[0004] In recent years, with the development of machine learning and artificial intelligence, pitch detection technology has significantly improved. Deep learning algorithms, particularly convolutional neural networks (CNNs) and recurrent neural networks (RNNs), have been used to improve detection accuracy. However, these algorithms are computationally complex and often require feature extraction based on spectrum or cepstrum calculations. Therefore, they suffer from the same real-time limitations as the aforementioned methods. Summary of the Invention

[0005] Based on this, it is necessary to provide a low-latency band pitch detection method, device and equipment based on a neural network that can reduce the pitch detection delay in order to address the above technical problems.

[0006] A low-latency band-based pitch detection method based on a neural network, the method comprising:

[0007] Get the audio signal.

[0008] A pitch detection model based on a neural network is constructed, and the audio signal is input as sample training data into the filter of the pitch detection model for filtering to obtain multi-channel sub-band audio vectors.

[0009] Each sub-band audio vector is subjected to feature extraction through a preset number of convolutional layers in the neural network to obtain a first pitch sub-band.

[0010] The second pitch signal is obtained by manually modifying the audio signal.

[0011] The first pitch subband and the second pitch signal are input into a pitch detection model for model training to obtain a pre-trained pitch detection model.

[0012] The pre-trained pitch detection model proofreads the audio signal according to the loss function and outputs a low-latency pitch signal.

[0013] In one embodiment, it also includes: constructing a pitch detection model based on a neural network, determining the filter order according to the sampling environment, inputting the audio signal as sample training data into the filter of the corresponding filter order in the pitch detection model for order filtering, and obtaining a multi-channel sub-band audio vector.

[0014] In one of the embodiments, it also includes: each sub-band audio vector is subjected to feature extraction through a preset number of convolution layers in the LSTM network architecture, and the extracted features are used to determine the probability of the zero-crossing rate of the sub-band audio vector using a softmax function, so as to obtain the sub-band audio vector corresponding to the maximum probability as the first pitch sub-band.

[0015] In one embodiment, the method further includes: calculating the energy characteristics of the sub-band audio vector using a preset number of convolutional layers in the LSTM network architecture:

[0016]

[0017] Wherein, E(n) is the energy feature of the sub-band audio vector, n is the current time point, x is the sub-band audio data, and L is the time window size for calculating energy.

[0018] In one embodiment, the method further includes: inputting the first pitch subband and the second pitch signal into a pitch detection model for model parameter training to obtain parameters of a neural network, optimizing the pitch detection model according to the parameters, and obtaining a pre-trained pitch detection model.

[0019] In one embodiment, the further step further includes: training a pitch detection model to calibrate the audio signal according to a cross entropy loss function, and outputting a low-latency pitch signal:

[0020]

[0021] Among them, Loss is the cross entropy loss function, N is the filter order, y i is the first pitch subband, It is a low-latency pitch signal.

[0022] A low-latency band-splitting pitch detection device based on a neural network, the device comprising:

[0023] The audio signal acquisition module is used to acquire the audio signal.

[0024] The sub-band audio vector acquisition module is used to build a pitch detection model based on a neural network, input the audio signal as sample training data into the filter of the pitch detection model for filtering, and obtain multiple sub-band audio vectors.

[0025] The first pitch subband acquisition module is used to extract features of each subband audio vector through a preset number of convolution layers in a neural network to obtain a first pitch subband.

[0026] The second pitch signal acquisition module is used to obtain the second pitch signal by manually correcting the audio signal.

[0027] The model training module is used to input the first pitch subband and the second pitch signal into the pitch detection model for model training to obtain a pre-trained pitch detection model.

[0028] The low-latency pitch signal acquisition module is used to pre-train the pitch detection model to calibrate the audio signal according to the loss function and output a low-latency pitch signal.

[0029] In one embodiment, it also includes: a first pitch subband acquisition module, which is also used to extract features of each sub-band audio vector through a preset number of convolution layers in the LSTM network architecture, and use the softmax function to determine the probability of the zero-crossing rate of the sub-band audio vector using the extracted features, so as to obtain the sub-band audio vector corresponding to the maximum probability as the first pitch sub-band.

[0030] A computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0031] Get the audio signal.

[0032] A pitch detection model based on a neural network is constructed, and the audio signal is input as sample training data into the filter of the pitch detection model for filtering to obtain multi-channel sub-band audio vectors.

[0033] Each sub-band audio vector is subjected to feature extraction through a preset number of convolutional layers in the neural network to obtain a first pitch sub-band.

[0034] The second pitch signal is obtained by manually modifying the audio signal.

[0035] The first pitch subband and the second pitch signal are input into a pitch detection model for model training to obtain a pre-trained pitch detection model.

[0036] The pre-trained pitch detection model proofreads the audio signal according to the loss function and outputs a low-latency pitch signal.

[0037] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the following steps:

[0038] Get the audio signal.

[0039] A pitch detection model based on a neural network is constructed, and the audio signal is input as sample training data into the filter of the pitch detection model for filtering to obtain multi-channel sub-band audio vectors.

[0040] Each sub-band audio vector is subjected to feature extraction through a preset number of convolutional layers in the neural network to obtain a first pitch sub-band.

[0041] The second pitch signal is obtained by manually modifying the audio signal.

[0042] The first pitch subband and the second pitch signal are input into a pitch detection model for model training to obtain a pre-trained pitch detection model.

[0043] The pre-trained pitch detection model proofreads the audio signal according to the loss function and outputs a low-latency pitch signal.

[0044] The neural network-based low-latency band-based pitch detection method, apparatus, and device first filter the audio signal through a filter to decompose it into multiple sub-band audio vectors. This sub-band processing method meticulously analyzes the signal's frequency components, helping to improve pitch detection accuracy. Each sub-band audio vector then undergoes feature extraction through a pre-set convolutional layer within the neural network, yielding a first pitch sub-band. Leveraging its powerful feature extraction capabilities, the convolutional neural network extracts pitch-related features from complex audio signals, improving detection accuracy. Next, the audio signal is manually corrected to obtain a second pitch signal, providing accurate pitch annotation and laying the foundation for supervised learning of the model. The first pitch sub-band and second pitch signal are then input into a pitch detection model for training, resulting in a pre-trained pitch detection model. This process combines the features extracted by the convolutional layer with the manually corrected pitch signal, improving the model's accuracy and robustness. Finally, the pre-trained model calibrates the audio signal according to a loss function and outputs a low-latency pitch signal. The cross-entropy loss function measures the difference between the model's predicted output and the actual pitch signal. Minimizing this loss function optimizes model performance. Throughout the entire process, the efficient feature extraction capabilities of convolutional neural networks, the parallel processing of multiple sub-band audio vectors, and the pre-training and proofreading mechanisms not only improve the accuracy of pitch detection but also significantly reduce processing latency. The optimized convolutional neural network architecture makes audio signal processing more efficient and reduces computational overhead; the sub-band decomposition method processes multiple frequency band signals in parallel, improving processing efficiency and reducing latency; combined with pre-trained models and cross-entropy loss function proofreading, while ensuring accuracy, the model's inference time is shortened, solving the problems of pitch detection accuracy and low latency, and ultimately achieving the technical effect of low-latency pitch detection, suitable for application scenarios such as real-time audio processing that require efficient pitch detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 1 is a flow chart of a low-latency band-splitting pitch detection method based on a neural network in one embodiment;

[0046] Figure 2 1 is a flow chart of a low-latency band-splitting pitch detection step based on a neural network in one embodiment;

[0047] Figure 3 Schematic diagram of a process for extracting energy generated by a filter using an LSTM network architecture in one embodiment;

[0048] Figure 4 1 is a block diagram of a low-latency band-splitting pitch detection device based on a neural network in one embodiment;

[0049] Figure 5 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0050] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0051] In one embodiment, Figure 1 As shown, a low-latency band-splitting pitch detection method based on a neural network is provided, comprising the following steps:

[0052] Step 102: Acquire an audio signal.

[0053] Specifically, the audio signal is obtained through a pitch recognition module in a vocal effector.

[0054] Step 104 : construct a pitch detection model based on a neural network, input the audio signal as sample training data into the filter of the pitch detection model for filtering, and obtain multi-channel sub-band audio vectors.

[0055] Specifically, 64 filters are constructed, using a 4th-order Butterworth bandpass filter. Alternatively, filters can be selected based on the properties of the audio signal, using other bandpass, lowpass, or highpass filters for band division processing. The human voice from 20 to 800 Hz is divided into 64 subbands in the log domain, and the width of each subband is log(2) / 12. This means that the range of the nth subband is:

[0056] Next, a neural network model is selected for feature extraction. This solution uses an LSTM (Long Short-Term Memory) recurrent neural network, but a dense, RNN, or custom network model is also possible. The 64 signals filtered by these 64 filters are then individually tested for energy and zero-crossing rate to obtain multiple sub-band audio vectors. Since signals are generally composed of fundamental frequencies and harmonics, these 64 filters exhibit significant regularity in the vowel segment. Furthermore, band filtering can be performed using a common elliptical bandpass filter bank, but other bandpass filters can also be used.

[0057] In step 106 , each sub-band audio vector is subjected to feature extraction through a preset number of convolution layers in a neural network to obtain a first pitch sub-band.

[0058] Step 108: Obtain a second pitch signal by manually correcting the audio signal.

[0059] Step 110 : Input the first pitch subband and the second pitch signal into a pitch detection model for model training to obtain a pre-trained pitch detection model.

[0060] In step 112 , the pre-trained pitch detection model calibrates the audio signal according to the loss function and outputs a low-latency pitch signal.

[0061] Specifically, a cross entropy (categorical_crossentropy) loss function is constructed to optimize the features of the input audio signal, and the output is a 64*1 vector.

[0062]

[0063] Among them, Loss is the cross entropy loss function, N is the filter order, y i is the first pitch subband, For low-latency pitch signals, the loss function type can also be adjusted according to the neural network model, such as Mean Absolute Error (MAE), Log Loss, Smooth L1 Loss, etc.

[0064] In the neural network-based low-latency band-based pitch detection method described above, the audio signal is first filtered and decomposed into multiple sub-band audio vectors. This sub-band processing method meticulously analyzes the signal's frequency components, helping to improve pitch detection accuracy. Each sub-band audio vector is then passed through a pre-set convolutional layer within the neural network for feature extraction, resulting in a first pitch sub-band. Leveraging its powerful feature extraction capabilities, convolutional neural networks extract pitch-related features from complex audio signals, improving detection accuracy. Next, the audio signal is manually corrected to obtain a second pitch signal, providing accurate pitch annotation and laying the foundation for supervised learning of the model. The first pitch sub-band and the second pitch signal are then fed into a pitch detection model for training, resulting in a pre-trained pitch detection model. This process combines the features extracted by the convolutional layer with the manually corrected pitch signal, improving the model's accuracy and robustness. Finally, the pre-trained model calibrates the audio signal using a cross-entropy loss function to output a low-latency pitch signal. The cross-entropy loss function measures the difference between the model's predicted output and the actual pitch signal. Minimizing this loss function optimizes model performance. Throughout the entire process, the efficient feature extraction capabilities of convolutional neural networks, the parallel processing of multiple sub-band audio vectors, and the pre-training and proofreading mechanisms not only improve the accuracy of pitch detection but also significantly reduce processing latency. The optimized convolutional neural network architecture makes audio signal processing more efficient and reduces computational overhead; the sub-band decomposition method processes multiple frequency band signals in parallel, improving processing efficiency and reducing latency; combined with pre-trained models and cross-entropy loss function proofreading, while ensuring accuracy, the model's inference time is shortened, solving the problems of pitch detection accuracy and low latency, and ultimately achieving the technical effect of low-latency pitch detection, suitable for application scenarios such as real-time audio processing that require efficient pitch detection.

[0065] In one embodiment, a pitch detection model based on a neural network is constructed, the filter order is determined according to the sampling environment, and the audio signal is input as sample training data into the filter of the corresponding filter order in the pitch detection model for step-by-step filtering to obtain a multi-channel sub-band audio vector.

[0066] In one embodiment, Figure 3 As shown, each sub-band audio vector is extracted through the number of convolution layers preset in the LSTM network architecture. The extracted features are then used to determine the probability of the sub-band audio vector's zero-crossing rate using the softmax function. The sub-band audio vector corresponding to the maximum probability is taken as the first pitch sub-band. The length signal of the sub-band audio vector is calculated using the number of convolution layers preset in the LSTM network architecture:

[0067]

[0068] Wherein, E(n) is the energy feature of the sub-band audio vector, n is the current time point, x is the sub-band audio data, and L is the time window size for calculating energy.

[0069] It is worth noting that the features of the sub-band audio vector can be extracted through the LSTM network architecture. It is also possible to use the computer autocorrelation coefficient method, calculate the cepstrum, etc., use energy as a feature, and send it into the neural network to obtain the probability of each sub-band.

[0070] In one embodiment, the first pitch subband and the second pitch signal are input into a pitch detection model for model parameter training to obtain parameters of a neural network, and the pitch detection model is optimized according to the parameters to obtain a pre-trained pitch detection model.

[0071] In one embodiment, the pitch detection model is trained to calibrate the audio signal according to the cross entropy loss function and output a low-latency pitch signal:

[0072]

[0073] Among them, Loss is the cross entropy loss function, N is the filter order, y i is the first pitch subband, It is a low-latency pitch signal.

[0074] It should be understood that although Figure 1-2 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. In addition, Figure 1-2 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0075] In one embodiment, Figure 4 As shown, a low-latency band-based pitch detection device based on a neural network is provided, comprising: an audio signal acquisition module 402, a sub-band audio vector acquisition module 404, a first pitch sub-band acquisition module 406, a second pitch signal acquisition module 408, a model training module 410, and a low-latency pitch signal acquisition module 412, wherein:

[0076] The audio signal acquisition module 402 is configured to acquire an audio signal.

[0077] The sub-band audio vector acquisition module 404 is used to build a pitch detection model based on a neural network, input the audio signal as sample training data into the filter of the pitch detection model for filtering, and obtain multiple sub-band audio vectors.

[0078] The first pitch sub-band acquisition module 406 is configured to perform feature extraction on each sub-band audio vector through a preset number of convolution layers in a neural network to obtain a first pitch sub-band.

[0079] The second pitch signal acquisition module 408 is configured to obtain a second pitch signal by manually modifying the audio signal.

[0080] The model training module 410 is configured to input the first pitch subband and the second pitch signal into a pitch detection model for model training to obtain a pre-trained pitch detection model.

[0081] The low-latency pitch signal acquisition module 412 is used to pre-train the pitch detection model to calibrate the audio signal according to the loss function and output a low-latency pitch signal.

[0082] In one embodiment, the first pitch subband acquisition module 406 is also used to extract features from each sub-band audio vector through a preset number of convolution layers in the LSTM network architecture, and use the softmax function to determine the probability of the zero-crossing rate of the sub-band audio vector using the extracted features, so as to obtain the sub-band audio vector corresponding to the maximum probability as the first pitch sub-band.

[0083] For the specific limitations of the low-latency band pitch detection device based on a neural network, please refer to the limitations of the low-latency band pitch detection method based on a neural network above, which will not be repeated here. The various modules in the above-mentioned low-latency band pitch detection device based on a neural network can be implemented in whole or in part by software, hardware and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules.

[0084] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 5As shown. The computer device includes a processor, memory, network interface, display screen and input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a low-latency band pitch detection method based on a neural network is implemented. The display screen of the computer device can be a liquid crystal display or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad provided on the computer device housing, or an external keyboard, touchpad or mouse.

[0085] Those skilled in the art will understand that Figure 4-5 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention and does not constitute a limitation on the computer device to which the solution of the present invention is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0086] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0087] Get the audio signal.

[0088] A pitch detection model based on a neural network is constructed, and the audio signal is input as sample training data into the filter of the pitch detection model for filtering to obtain multi-channel sub-band audio vectors.

[0089] Each sub-band audio vector is subjected to feature extraction through a preset number of convolutional layers in the neural network to obtain a first pitch sub-band.

[0090] A second pitch signal is obtained by manually modifying the audio signal.

[0091] The first pitch subband and the second pitch signal are input into a pitch detection model for model training to obtain a pre-trained pitch detection model.

[0092] The pre-trained pitch detection model proofreads the audio signal according to the loss function and outputs a low-latency pitch signal.

[0093] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0094] Get the audio signal.

[0095] A pitch detection model based on a neural network is constructed, and the audio signal is input as sample training data into the filter of the pitch detection model for filtering to obtain multi-channel sub-band audio vectors.

[0096] Each sub-band audio vector is subjected to feature extraction through a preset number of convolutional layers in the neural network to obtain a first pitch sub-band.

[0097] A second pitch signal is obtained by manually modifying the audio signal.

[0098] The first pitch subband and the second pitch signal are input into a pitch detection model for model training to obtain a pre-trained pitch detection model.

[0099] The pre-trained pitch detection model proofreads the audio signal according to the loss function and outputs a low-latency pitch signal.

[0100] Those skilled in the art will appreciate that all or part of the processes in the above-described embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above-described embodiments. Among them, any reference to memory, storage, database or other media used in the embodiments provided by the present invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct RAM bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM).

[0101] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0102] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the patent for this invention shall be determined by the appended claims.

Claims

1. A low-latency band-splitting pitch detection method based on a neural network, characterized in that: The method comprises: Get audio signal; Constructing a pitch detection model based on a neural network, inputting the audio signal as sample training data into a filter of the pitch detection model for filtering, and obtaining a multi-channel sub-band audio vector; Each of the sub-band audio vectors is subjected to feature extraction through a preset number of convolutional layers in the neural network to obtain a first pitch sub-band; Obtaining a second pitch signal by manually modifying the audio signal; Inputting the first pitch subband and the second pitch signal into the pitch detection model for model training to obtain a pre-trained pitch detection model; The pre-trained pitch detection model proofreads the audio signal according to the loss function and outputs a low-latency pitch signal.

2. The method according to claim 1, characterized in that Constructing a pitch detection model based on a neural network, inputting the audio signal as sample training data into the filter of the pitch detection model for filtering, and obtaining a multi-channel sub-band audio vector, including: A neural network-based pitch detection model is constructed, the filter order is determined according to the sampling environment, the audio signal is input as sample training data into the filter corresponding to the filter order in the pitch detection model for hierarchical filtering, and a multi-channel sub-band audio vector is obtained.

3. The method according to claim 1, characterized in that Each of the sub-band audio vectors is subjected to feature extraction through a preset number of convolutional layers in the neural network to obtain a first pitch sub-band, including: Each sub-band audio vector is subjected to feature extraction through the number of convolutional layers preset in the LSTM network architecture, and the extracted features are used to determine the probability of the zero-crossing rate of the sub-band audio vector using the softmax function, so as to obtain the sub-band audio vector corresponding to the maximum probability as the first pitch sub-band.

4. The method according to claim 3, characterized in that Inputting the first pitch subband and the second pitch signal into the pitch detection model for model training to obtain a pre-trained pitch detection model, including: The first pitch subband and the second pitch signal are input into the pitch detection model for model parameter training to obtain parameters of the neural network, and the pitch detection model is optimized according to the parameters to obtain a pre-trained pitch detection model.

5. The method according to any one of claims 1 to 4, characterized in that The pre-trained pitch detection model calibrates the audio signal according to the loss function and outputs a low-latency pitch signal, including: Train the pitch detection model, calibrate the audio signal according to the cross-entropy loss function, and output a low-latency pitch signal: Among them, Loss is the cross entropy loss function, N is the filter order, y i is the first pitch subband, It is a low-latency pitch signal.

6. A low-latency band-splitting pitch detection device based on a neural network, characterized in that: The device comprises: An audio signal acquisition module, used to acquire an audio signal; A sub-band audio vector acquisition module is used to build a pitch detection model based on a neural network, input the audio signal as sample training data into the filter of the pitch detection model for filtering, and obtain multiple sub-band audio vectors; A first pitch subband acquisition module is configured to extract features of each subband audio vector through a preset number of convolution layers in the neural network to obtain a first pitch subband; A second pitch signal acquisition module, configured to obtain a second pitch signal by manually correcting the audio signal; a model training module, configured to input the first pitch subband and the second pitch signal into the pitch detection model for model training to obtain a pre-trained pitch detection model; A low-latency pitch signal acquisition module is used for the pre-trained pitch detection model to proofread the audio signal according to the loss function and output a low-latency pitch signal.

7. The device according to claim 6, characterized in that The first pitch subband acquisition module is also used to extract features of each subband audio vector through a preset number of convolution layers in the LSTM network architecture, and use the softmax function to determine the probability of the zero-crossing rate of the subband audio vector using the extracted features, so as to obtain the subband audio vector corresponding to the maximum probability as the first pitch subband.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 5 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 5 are implemented.

Images