A method and device for processing audio signals based on machine learning algorithm catboost

Abstract

An audio signal processing method and device based on a machine learning algorithm CatBoost, belonging to the fields of audio signal processing and computer technology. A spatial directional receiving array is used to collect real-time audio signals of a sound source, a signal preprocessing system is used to preprocess the collected real-time audio signals, and the preprocessed data is used as input data of the machine learning algorithm CatBoost. The present invention uses a spatial directional receiving array to arrange signal receivers. Due to the improved arrangement of the spatial directional receiving array, the signal fading and interference are reduced, thereby improving the signal reception quality. At the same time, the technical solution makes full use of the characteristics of the machine learning algorithm CatBoost that it is good at processing category features, and applies the machine learning algorithm CatBoost to the field of audio signal processing, thereby obtaining the beneficial effect of being able to handle missing values ​​and outliers of data well while maintaining category information.

Description

Technical Field

[0001] The present invention belongs to the field of audio signal processing and computer technology, and in particular to a method and device for processing audio signals by adopting a machine learning algorithm catboost. Background Art

[0002] In the prior art, when collecting audio signals from a certain sound source, it is easily affected by environmental noise or sounds from other sound sources. For example, when collecting audio signals from a certain sound source in a recording studio, it is easily disturbed by reverberation, environmental noise, etc. When collecting audio signals from a certain sound source in an environment such as a theater or concert hall, it is easily disturbed by other musical instruments, environmental noise, etc., resulting in reduced signal quality. Therefore, when collecting sound signals, there are technical problems such as the inability to effectively collect signals and reduced quality of collected signals due to the dispersion of sound sources during audio signal processing. Summary of the invention

[0003] The purpose of the present invention is to solve the above technical problems existing in the background technology, use a spatial directional receiving array to collect real-time audio signals, use a preprocessing system to process the collected real-time audio signals, and input the preprocessed audio data into a trained machine learning algorithm CatBoost model for processing. The specific solutions provided by the present invention are as follows:

[0004] An audio signal processing method based on a machine learning algorithm catboost comprises the following steps:

[0005] Step 1: Collect the real-time audio signal of the sound source;

[0006] A spatial directional receiving array is used to collect real-time audio signals from a sound source; wherein the spatial directional receiving array includes a signal receiver array, the signal receiver array is arranged according to a sound radiation pattern of the sound source, and the receiving direction of the signal receiver array is directed toward the sound source;

[0007] Step 2: Preprocess the collected audio signal;

[0008] The audio signal collected by the signal receiver array is input into the signal preprocessing system to enhance the signal strength. Specifically, the signal preprocessing system includes a preamplifier, a bandpass or low-pass filter; the collected audio signal is amplified by the preamplifier, and then a low-pass filter or a bandpass filter is used to remove high-frequency noise or unnecessary high-frequency signal components, and the processed signal is converted from the time domain to the frequency domain. Then the spatial spectrum method is used to separate and enhance the audio signal processed by the bandpass or low-pass filter. Specifically, at the beginning of signal acquisition or when there is a signal-free part during the signal acquisition process, the spectrum of the background noise is estimated, and the signal decomposition is completed by calculating the covariance matrix of the signal. Then the eigenvalues ​​of the extracted audio signal are used to decompose the audio signal into a noise subspace and a signal subspace. After the noise subspace is identified, the noise is suppressed by weighted subtraction; when subtracting the noise power spectrum, a weight factor is introduced to control the degree of noise suppression.

[0009] After the noise subspace is suppressed, the signal is optimized and adjusted using the dynamic range compression method. The optimization and adjustment of the signal using the dynamic range compression method specifically includes determining a level threshold at which compression begins, and the portion of the signal exceeding the level threshold is compressed. The compression ratio is set to determine the degree of compression when the input signal exceeds the threshold; the compression time is set to determine the time required for the compressor to start working and reach the full compression state; and the gain compensation is set to increase the overall volume of the compressed signal.

[0010] Step 3: Using a machine learning algorithm CatBoost to process the audio signal of the sound source;

[0011] Step 3.1: Collect audio samples of sound sources in different states, extract feature values ​​from the audio samples after preprocessing, and label the correct pitch for each audio sample as a training set and a test set for the machine learning method CatBoost;

[0012] Step 3.2: Build a model based on the machine learning algorithm CatBoost and train the model using the training set;

[0013] Step 3.3: Evaluate the performance of the model on the test set and adjust the model parameters based on the evaluation results;

[0014] Step 3.4: Get the trained CatBoost model for real-time audio signal processing.

[0015] The signal receiver array includes a linear array, a circular array, a spherical array and a cellular array, which are selected according to the target application, the required direction control accuracy, the frequency range, the expected signal processing complexity and the cost-effectiveness; wherein the linear array adopts a uniform linear array or a non-uniform linear array, and the spacing between adjacent units of the uniform linear array is set to half the wavelength; for the non-uniform linear array, the spacing is adjusted according to a specific optimization goal. The composition of the cellular array is to distribute the linear array in 6 uniform linear direction areas, and the signal receiver is arranged in each area using the linear array, and the areas are connected by a circular array or a spherical array to cover the area not covered by the linear array.

[0016] When the sound source is a double bass, the real-time audio signal from the double bass is preprocessed and input into the trained catboost model, which outputs the pitch deviation value and string tension deviation value compared with the standard pitch. Then a mapping function is created to convert the pitch deviation and string tension deviation predicted by the trained CatBoost model into specific tuning actions.

[0017] The present invention also provides an audio signal processing device based on a machine learning algorithm catboost, the device comprising an intelligent pitch correction component and a tuning device; wherein the intelligent pitch correction component controls the tuning device, and the tuning device tunes the double bass;

[0018] The intelligent pitch correction component specifically includes a display device, a central processing unit and a control system; the display device sets the target pitch value and string tension value of the strings through an interface, and displays the changes in the tension, amplitude and frequency of the strings in real time, and provides manual and automatic tuning options; the central processing unit can run the program stored on the processor, and when the program is executed, it can implement the above-mentioned trained CatBoost model and mapping function; the control system controls the tuning device to execute the tuning action determined by the mapping function according to the data fed back by the central processing unit;

[0019] The tuning device is used to tune the double bass specifically as follows: the tuning device includes a screw assembly, a drive assembly, and an overload protection device; the drive assembly receives instructions from the control system to drive the screw assembly, the screw assembly is vertically installed inside the head of the double bass, the screw assembly includes a shaft body, a shaft cap, and a shaft seat, wherein the shaft body is connected to the worm on the tuning knob on the head of the double bass through a worm gear, the shaft body is provided with a thread for winding the strings, the drive assembly drives the tuning knob, and the tuning knob in turn drives the shaft body to rotate, thereby adjusting the strings; the shaft cap covers the outside of the shaft body, and the shaft seat is provided with a hole matching the shaft body for installing the shaft body; the overload protection device is used to automatically cut off power in the event of an overload.

[0020] The beneficial effects of the present invention are as follows:

[0021] The present invention adopts a spatial directional receiving array to arrange signal receivers. Due to the improved arrangement of the spatial directional receiving array, the fading and interference of the signal are reduced, thereby improving the effect of the signal reception quality. Multiple signal receivers in the receiving array are used to form a signal pointing to a specific direction, enhance the receiving ability of the signal in the specific direction, and suppress interference in other directions. By adjusting the phase and amplitude of the received signal of each receiver in the array, the direction of the signal source is estimated and tracked. The received signal is preprocessed by a noise suppression algorithm to extract useful information, such as the direction, frequency, amplitude, etc. of the signal.

[0022] At the same time, the present invention makes full use of the characteristics of the machine learning algorithm CatBoost that it is good at processing category-type features, and applies the machine learning algorithm CatBoost to the field of audio signal processing. Since the machine learning algorithm CatBoost is applied to the processing of audio signals, when processing audio data signals, the beneficial effect of being able to well process missing values ​​and abnormal values ​​of data while maintaining category information is obtained, which has good adaptability to incomplete or abnormal data that may occur in audio signal processing. At the same time, the machine learning algorithm CatBoost can automatically identify the features that contribute most to model prediction, which helps to screen out key parameters in audio signal processing and improve processing efficiency and accuracy. The machine learning algorithm CatBoost adopts an adaptive learning rate adjustment strategy, which can converge quickly and obtain better model performance, and helps the model to automatically adjust the learning speed during the training process to achieve better optimization effects. The machine learning algorithm CatBoost reduces the need for hyperparameter tuning, reduces the probability of overfitting, and makes the model more versatile and robust. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a workflow diagram for processing audio signals using a spatial directional receiving array and a machine learning algorithm, catboost.

[0024] Figure 2 This is a workflow diagram for processing and tuning the audio signal of a double bass using a spatial directional receiving array and a machine learning algorithm, catboost. DETAILED DESCRIPTION

[0025] The preferred embodiments of the present invention are shown in the accompanying drawings.

[0026] Example 1

[0027] like Figure 1 As shown, it is a specific flow chart of an audio signal processing method based on a machine learning model catboost provided by the present invention, and the specific steps are as follows:

[0028] Step 1: Collect the real-time audio signal of the sound source;

[0029] Specifically, this embodiment uses a spatial directional receiving array to collect real-time audio signals from a sound source; wherein the spatial directional receiving array includes a signal receiver array, which is arranged according to the sound radiation pattern of the sound source, and the receiving direction of the signal receiver array is directed toward the sound source. For a uniform linear array, the ideal spacing between adjacent units is approximately half the wavelength λ (λ / 2) to avoid phase synchronization problems; for a non-uniform linear array, the spacing can be adjusted according to a specific optimization goal.

[0030] When designing a signal receiver array, the specific spacing or angle values, the arrangement of the sub-arrays, and the parameter settings depend on many factors, including the target application, the required directional control accuracy, the frequency range, the expected signal processing complexity, and cost-effectiveness. Specifically, the signal receivers can be arranged in a mixed manner of linear arrays, circular arrays, spherical arrays, and cellular arrays.

[0031] This embodiment uses a honeycomb array to arrange the signal receivers. The honeycomb array is composed of distributing the linear array in 6 uniform linear direction areas. The linear array is used to arrange the signal receivers in each area. The areas are connected by a circular array or a spherical array to cover the areas not covered by the linear array. The linear array uses a uniform linear array or a non-uniform linear array. The spacing between adjacent units of the uniform linear array is set to half the wavelength to avoid phase synchronization problems; for the non-uniform linear array, the spacing can be adjusted according to specific optimization goals. To arrange the signal receiver array according to the circular array is to arrange the signal receiver array in a circular manner around a center point according to a circular plane in a two-dimensional space in order to collect signals from different directions. To arrange the signal receiver according to the spherical array is to evenly distribute the signal receivers on a spherical surface in a three-dimensional space according to a spherical surface.

[0032] Step 2: Preprocess the collected audio signal;

[0033] The audio signal collected by the signal receiver array is input into the signal preprocessing system to enhance the signal strength. Specifically, the signal preprocessing system includes a preamplifier and a bandpass filter;

[0034] The audio signal collected in step 1 may contain echoes reflected by the target, noise and other interference. Therefore, the role of the preamplifier is to increase the amplitude of the signal to ensure sufficient signal strength in the subsequent processing stage. This step is particularly important because the signal may be very weak at long distances or under low signal-to-noise ratio conditions. The bandwidth of the preamplifier ranges from 20Hz to 800Hz, the noise figure is between 0dB and 5dB, and the output impedance is above 10KΩ to reduce signal attenuation.

[0035] After being amplified by the preamplifier, the signal still contains noise. At this time, a low-pass filter or a band-pass filter is used to remove high-frequency noise or unnecessary high-frequency signal components and retain the main signal frequency range. The processed signal is converted from the time domain to the spatial domain.

[0036] Then use the band-stop filter to suppress interference signals within a specific frequency range, and adjust the filter cutoff frequency or gain according to the strength of the noise and signal to enhance the signal strength. For example, the ground reflection signal may need to be suppressed, and the band-stop filter can accurately filter out this part of the frequency. Through the fine adjustment of the band-stop filter, the performance of signal processing can be further optimized, and the accuracy and reliability of target detection can be improved.

[0037] Furthermore, preprocessing also includes using the spatial spectrum method to separate and enhance the audio signal processed by the bandpass filter. The combination of the spatial spectrum method with the preamplifier and the bandstop filter and the bandpass filter is mainly to enhance the performance of signal processing. The following are the specific steps:

[0038] The processed signal is converted from the time domain to the frequency domain. This step usually involves Fourier transform or other forms of spatial domain conversion algorithms to analyze the spatial distribution characteristics of the signal. In the spatial domain, the spectral method is applied to analyze the spectral characteristics of the signal, extract the eigenvalues ​​of the audio signal, and suppress environmental noise.

[0039] Specifically, at the beginning of signal acquisition or when there is a signal-free part during the signal acquisition process, the spectrum of the background noise is estimated, and the signal decomposition is completed by calculating the covariance matrix of the signal; then the eigenvalues ​​of the extracted audio signal are used to decompose the audio signal into a noise subspace and a signal subspace; after the noise subspace is identified, the noise is suppressed by weighted subtraction; when subtracting the noise power spectrum (including background noise and environmental noise), a weight factor is introduced to control the degree of noise suppression to prevent excessive suppression from causing signal distortion; after the noise subspace is suppressed, the dynamic range compression method is used to optimize and adjust the signal.

[0040] The extracted characteristic values ​​of the audio signal are used to analyze the spectrum to identify the main frequency range or mode of the noise;

[0041] The dynamic range compression method is used to reduce the difference between the maximum spectrum and the minimum spectrum, including determining a level threshold at which compression starts, and the signal portion exceeding the level threshold is compressed;

[0042] The level threshold is usually set according to the dynamic range of the audio signal and the desired compression effect. A level between the maximum peak and the minimum peak is selected as the threshold. The maximum peak is 100dB SPL, the minimum peak is 50dB SPL, and 70dB SPL is selected as the threshold.

[0043] And set the compression ratio to determine the degree of compression when the input signal exceeds the threshold; the compression ratio represents the relationship between the output signal after the input signal exceeds the threshold and the input signal that does not exceed the threshold. When the input signal exceeds the threshold, the output signal will be compressed to half of the original value. The compression ratio is set to 2:1. For example, if the input signal is 80dB SPL, the output signal is compressed to 40dB SPL.

[0044] The compression time is set to determine the time required for the compressor to start working and reach the full compression state; the compression time is the time required for the compressor to fully compress from detecting that the signal exceeds the level threshold. The compression time is set to 50ms.

[0045] Set the release time, which includes the time it takes the compressor to recover from full compression to an uncompressed state, to 200ms

[0046] Since compression will reduce the overall volume, setting gain compensation is used to increase the overall volume of the compressed signal to offset the effect of volume reduction during compression and make it match the loudness of the original signal; the average sound pressure level of the original signal is 60dB SPL, which is reduced to 55dB SPL after compression. By increasing the output gain by +5dB, the average sound pressure level of the compressed signal returns to 60dB SPL to match the original signal.

[0047] The spectrum method can help identify the location, size and other information of the target by calculating the power spectrum density or correlation function in different directions. Using the results of the spatial spectrum method, the target signal is enhanced to further improve the contrast between the target and the background noise.

[0048] Step 3: Using a machine learning method CatBoost to process the audio signal of the sound source;

[0049] Step 3.1: Collect audio samples of the sound source in different states, extract feature values ​​from the audio samples, label each audio sample with the correct pitch as a training label for the machine learning method CatBoost; and divide the data into a training set and a test set according to a certain ratio.

[0050] Step 3.2: Build a regression model based on the machine learning method CatBoost and train the regression model;

[0051] Step 3.3: Use the training set to train the regression model and evaluate the performance of the model on the test set, including calculating indicators such as accuracy, confusion matrix, AUC-ROC curve, etc.

[0052] Step 3.4: Based on the evaluation results, adjust the model parameters to improve performance; this can be achieved through grid search, random search, or Bayesian optimization. Save the trained machine learning algorithm CatBoost model for later use.

[0053] Specifically, when using the CatBoost algorithm for training, it usually includes: 1. Configuration of the training environment: Select a suitable GPU or CPU, and sufficient memory to support the training process. Make sure that you are running a Linux or Windows system that supports Python. Install the latest version of Python and install the CatBoost library through pip or conda. Make sure that common data processing and machine learning libraries such as NumPy, Pandas and scikit-learn are installed. 2. Use of loss function: CatBoost provides a variety of loss functions to choose from, depending on the type of task: This application uses `train_test_split` or `RMSE` (mean square error) as the loss function. 3. The method used for hyperparameter optimization, CatBoost provides an automatic hyperparameter tuning function, manually defines a hyperparameter grid, and then uses cross-validation to find the best combination. Evaluate the performance of the CatBoost model of the machine learning algorithm, including evaluating the calculation accuracy of the CatBoost model of the machine learning algorithm. The method of evaluating model performance includes measuring the proportion of samples predicted correctly and intuitively displaying the relationship between the model prediction results and the actual results. According to the evaluation results, the parameters of the CatBoost model of the machine learning algorithm are adjusted by grid search, random search or Bayesian optimization method to improve the performance, and the training and optimization of the catboost model of the machine learning algorithm are completed. During and after the training, the performance of the model is evaluated using appropriate evaluation indicators (such as mean square error, root mean square error, etc.). Ensure that the model has good performance.

[0054] Input the real-time audio source data into the trained machine learning algorithm catboost model, and output the pitch deviation value after comparison with the standard pitch. Through the following steps, a model is constructed to input the real-time audio signal and predict the difference between the pitch standard value and the real-time pitch value, including the following steps: First, convert the real-time audio signal into a form that can be processed by the computer, and adjust the sampling rate to ensure that the sampling rate of the audio signal is suitable for the input requirements of the model. Extract useful features from the audio signal, such as Mel-frequency cepstral coefficients (MFCCs), energy, zero-crossing rate, etc. These features can capture key information of the audio signal, such as pitch, tone, rhythm, etc. Prepare a large number of audio samples and their corresponding pitch standard values ​​as training data sets for training models. Use the trained CatBoost model to predict the difference between the pitch standard value and the real-time pitch value.

[0055] Example 2

[0056] In order to facilitate understanding of the present invention, the present invention is described by taking a double bass as an example of a sound source. However, the present invention can be applied to other sound sources other than a double bass and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to provide a more thorough and comprehensive understanding of the disclosure of the present invention.

[0057] Combined with Figure 2 , a spatial directional receiving array is used to collect and preprocess audio signals from sound sources such as double bass. The specific implementation method is to use multiple signal receivers to form an array, and these signal receivers have a wide frequency response and high sensitivity to capture the low-frequency musical sound of the double bass. The signal receivers are arranged according to the shape and sound radiation pattern of the double bass to ensure the comprehensive collection of the sound of the double bass. The receiving pattern of the signal receiver array is directed to the double bass to enhance the sound from the double bass and reduce the background noise. The collected sound is analyzed using the spatial spectrum method to extract musical elements such as timbre, pitch, and rhythm. The spatial spectrum method is applied to separate and enhance the sound of the double bass, while suppressing the musical sound of other instruments or environmental noise. The spectrum space method is an audio signal processing technology commonly used to reduce background noise. The basic idea of ​​this method is to identify the noise component in the spectrum of the signal, and then subtract the spectrum of the noise from the total spectrum to retain the useful components of the signal.

[0058] Here are the specific steps of the spatial spectrum method: First, the spectrum of the background noise needs to be estimated. This is usually done at the beginning of the recording, or by finding a signal-free (silent) part in the signal. Signal decomposition is done by calculating the covariance matrix of the signal, which decomposes the signal into a noise subspace and a signal subspace. Then eigenvalue decomposition is used to find the subspaces of the noise and signal. The noise subspace is identified and the noise is suppressed by weighted subtraction. When subtracting the noise power spectrum, a weight factor is introduced to control the degree of noise suppression and prevent over-suppression from causing signal distortion. After the noise subspace is suppressed, the reconstructed signal is completed by combining the remaining signal subspace with the residual part of the noise subspace. The processed signal is optimized and adjusted using dynamic range compression. Dynamic range compression is an audio processing technique used to reduce the dynamic range of a signal, that is, to reduce the difference between the maximum volume and the minimum volume.

[0059] For the double bass, this means that the volume remains at a relatively consistent and clear level during both intense and soft passages. The specific steps to implementing dynamic range compression are: Determine the level threshold at which compression begins, and the portion of the signal above this threshold will be compressed. Set the ratio. The compression ratio determines the degree of compression when the input signal exceeds the threshold. For example, a 4:1 ratio means that the output signal only increases by 1dB every time the input signal exceeds the threshold by 4dB. Set the compression time. The compression time determines the time it takes for the compressor to start working and reach the full compression state. For the double bass, a fast compression time can catch the beginning of a note and prevent a sudden increase in volume. Set the release time. The release time defines the time it takes for the compressor to recover from the full compression state to the uncompressed state. Set the gain compensation. Since compression will reduce the overall volume, the gain compensation is used to increase the overall volume of the compressed signal to match the loudness of the original signal. The compression parameters are dynamically adjusted according to the intensity of the double bass.

[0060] The specific steps of combining the data processed by the above-mentioned spatial directional receiving array with the CatBoost model for intelligent tuning of the double bass include: collecting audio samples of each string of the double bass in different tuning states, extracting features such as fundamental frequency, pitch, timbre, string tension, etc. from the audio samples, and marking the correct pitch and corresponding string tension for each sample as the training label of the CatBoost model. The CatBoost model is trained, the data set is divided into a training set and a test set, and a regression model is built using the CatBoost library. The input is the extracted audio features and string tension features, and the output is the pitch deviation and string tension deviation, that is, the deviation from the standard pitch and standard tension. The tuning algorithm logic is designed, which adjusts the pitch and string tension of the string according to the prediction results of the CatBoost model. The CatBoost model is integrated in the tuning algorithm so that it can receive audio features and output the pitch deviation and string tension deviation. A mapping function is created to convert the pitch deviation and string tension deviation predicted by the trained CatBoost model into specific tuning actions, such as string stretching or relaxation. The tuning device is then controlled to execute the tuning action determined by the mapping function.

[0061] The spatial directional receiving array described above is used to collect and process the real-time audio of the double bass, and the CatBoost model is used to predict the pitch deviation and string tension deviation in real time, and trigger the tuning algorithm. At the same time, a user interface is designed to allow the user to start the tuning process and view the prediction results. Manual and automatic tuning options are provided, as well as the parameters of the CatBoost model and the tuning algorithm are adjusted. The pitch deviation and string tension deviation predicted by the CatBoost model are converted into specific tuning actions.

[0062] The intelligent pitch correction component of the double bass controls the tuning device, and the intelligent pitch correction component includes a display device, a central processor, and a control system. The display device sets the target pitch value and string tension value of the string through an interface, and displays the changes in the tension, amplitude, and frequency of the string in real time. The display device allows the user to start the tuning process, view the prediction results, and provide manual and automatic tuning options, as well as adjust the CatBoost model and tuning algorithm parameters through the interface. The central processing unit calculates the string pitch and string tension values ​​that need to be adjusted through the CatBoost model based on the preset pitch standard and the needs of the performer. The control system adjusts the speed and direction of the micro motor of the tuning device based on the feedback data to achieve automatic adjustment and calibration of the string tension. When it is detected that the string tension deviates from the standard value, the central processing unit sends an instruction to the control system, and the control system controls the tuning device to adjust the strings. If the string tension is too low, the system will increase the number of string windings to increase the tension. If the tension is too high, the number of windings will be reduced to reduce the tension. This dynamic adjustment process will continue until the string tension value reaches or approaches the preset standard value.

[0063] The tuning device includes a screw assembly, a drive assembly, and an overload protection device. The drive assembly receives instructions from the central processor of the control system to drive the screw assembly. The drive assembly includes a micro motor, which is responsible for providing power. The screw assembly is vertically installed inside the headstock. The screw assembly includes a shaft body, a shaft cap, and a shaft seat. The shaft body is connected to the worm on the tuning knob through a worm gear. The shaft body is provided with a thread for winding the strings. The micro motor drives the tuning knob, and the tuning knob in turn drives the shaft body to rotate to adjust the strings. Through the action of the thread on the shaft body, the winding length of the string on the shaft body can be changed, thereby adjusting the tension of the string. The shaft cap covers the outside of the shaft body, which plays a role in protecting the shaft body and providing a grip point. A hole matching the shaft body is provided in the shaft seat for installing the shaft body. The overload protection device automatically cuts off the power in the event of an overload to protect the screw assembly from damage.

[0064] The technical features of the above-described embodiments can be combined arbitrarily. In order to make the description concise, all possible combinations of the technical features in the above-described embodiments are not 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. The above-described embodiments only express several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they cannot be understood as limiting the scope of the patent of the present invention. It should be pointed out that for ordinary technicians in this field, without departing from the concept of the present invention, several variations and improvements can be made, which all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be based on the attached claims.

Claims

1. An audio signal processing method based on a machine learning algorithm catboost, characterized in that: The following steps are involved: Step 1: Collect the real-time audio signal of the sound source; A spatial directional receiving array is used to collect real-time audio signals from the sound source; The spatial directional receiving array includes a signal receiver array, and the receiving direction of the signal receiver array is directed toward the sound source; Step 2: Preprocess the collected audio signal; The collected audio signal is amplified by a preamplifier, and then a low-pass filter or a band-pass filter is used to remove high-frequency noise or unnecessary signal components, and the processed signal is converted from the time domain to the frequency domain; then the spatial spectrum method is used to separate and enhance the audio signal processed by the band-pass filter or the low-pass filter; specifically, at the beginning of signal collection or when there is a signal-free part during the signal collection process, the spectrum of the background noise is estimated, and the signal decomposition is completed by calculating the covariance matrix of the signal; then the eigenvalues ​​of the extracted audio signal are used to decompose the audio signal into a noise subspace and a signal subspace; after identifying the noise subspace, the noise is suppressed by weighted subtraction; when subtracting the noise power spectrum, a weight factor is introduced to control the degree of noise suppression; After the noise subspace is suppressed, the signal is optimized and adjusted using a dynamic range compression method, specifically including determining a level threshold at which compression begins, a portion of the signal exceeding the level threshold being compressed, setting a compression ratio, and determining a degree of compression when the input signal exceeds the threshold; Set the compression time to determine the time required from the start of compression to the full compression state; Setting the gain compensation is used to increase the overall volume of the compressed signal; Step 3: Using a machine learning algorithm CatBoost to process the audio signal of the sound source; Step 3.1: Collect audio samples of sound sources in different states, extract feature values ​​from the audio samples after preprocessing, and mark the correct pitch for each audio sample as a training set and a test set for the machine learning algorithm CatBoost; Step 3.2: Build a model based on the machine learning algorithm CatBoost and train the model using the training set; Step 3.3: Evaluate the performance of the model on the test set and adjust the model parameters based on the evaluation results; Step 3.4: Get the trained CatBoost model for real-time audio signal processing.

2. The audio signal processing method based on the machine learning algorithm catboost according to claim 1, characterized in that: The signal receiver array includes a linear array, a circular array, a spherical array and a cellular array, which are selected according to the target application, the required direction control accuracy, the frequency range, the expected signal processing complexity and the cost-effectiveness; wherein the linear array adopts a uniform linear array or a non-uniform linear array, and the spacing between adjacent units of the uniform linear array is set to half the wavelength; for the non-uniform linear array, the spacing is adjusted according to the optimization target.

3. The audio signal processing method based on the machine learning algorithm catboost according to claim 2 is characterized in that: The signal receiver array is a honeycomb array, which is constructed by distributing linear arrays evenly in six straight line direction areas. Signal receivers are arranged in each area using the linear array, and the areas are connected by circular arrays or spherical arrays to cover areas not covered by the linear arrays.

4. The audio signal processing method based on the machine learning algorithm catboost according to claim 3 is characterized in that: The sound source is a double bass; the real-time audio signal emitted by the double bass is preprocessed and then input into a trained catboost model, which outputs a pitch deviation value and a string tension deviation value compared with a standard pitch.

5. The audio signal processing method based on the machine learning algorithm catboost according to claim 4 is characterized in that: The method also includes: creating a mapping function to convert the pitch deviation and string tension deviation predicted by the trained CatBoost model into a specific tuning action.

6. A device for processing audio signals using the method of claim 5 based on the machine learning algorithm catboost, characterized in that: The device comprises an intelligent pitch correction component and a tuning device; wherein the intelligent pitch correction component controls the tuning device, and the tuning device tunes the double bass; The intelligent pitch correction component specifically includes a display device, a central processing unit and a control system; the display device sets the target pitch value and the string tension value of the string through an interface, and displays the changes in the string tension, amplitude and frequency in real time, and provides manual and automatic tuning options; the central processing unit can run the program stored on the processor, and when the program is executed, it can implement the trained CatBoost model and mapping function as described in claim 5; the control system controls the tuning device to execute the tuning action determined by the mapping function according to the data fed back by the central processing unit; The tuning device is used to tune the double bass specifically as follows: the tuning device includes a screw assembly, a drive assembly, and an overload protection device; the drive assembly receives instructions from the control system to drive the screw assembly, the screw assembly is vertically installed inside the head of the double bass, the screw assembly includes a shaft body, a shaft cap, and a shaft seat, wherein the shaft body is connected to the worm on the tuning knob on the head of the double bass through a worm gear, the shaft body is provided with a thread for winding the strings, the drive assembly drives the tuning knob, and the tuning knob in turn drives the shaft body to rotate, thereby adjusting the strings; the shaft cap covers the outside of the shaft body, and the shaft seat is provided with a hole matching the shaft body for installing the shaft body; the overload protection device is used to automatically cut off power in the event of an overload.

Images