A mobile device recording full-link automatic audio enhancement method and system

Abstract

The application discloses a kind of mobile equipment recording full-link automatic audio enhancement method and system, mainly related to audio signal processing technical field.It includes: decoding to input audio, obtain multi-channel floating point data;Optimal single channel is automatically selected by multidimensional quality score;Detect and repair clipping distortion;Execute high-pass filter to remove direct current and infrasound interference;Power frequency trap wave elimination power grid humming;Based on the minimum value statistics and decision-oriented wiener filtering of deviation correction are realized adaptive noise reduction, and can be controlled by user intensity parameter;Peak normalization to noise reduction signal;Encoding output enhanced audio.The beneficial effects of the application are that it realizes full-automatic recording enhancement without music noise under the condition of no mute section in fixed order full-link pipeline.

Description

Technical Field

[0001] This invention relates to the field of audio signal processing technology, specifically to a method and system for automatic audio enhancement across the entire recording process of mobile devices. Background Technology

[0002] Post-processing of recordings on mobile devices generally suffers from the problem of "independent research of each module and lack of system integration". In terms of channel selection, existing solutions mainly include: fixed selection of a certain channel, selection based on a single RMS energy or peak level, or selection based on simple signal-to-noise ratio that relies on the silent segment as a reference. None of these solutions comprehensively consider multi-dimensional quality factors such as signal-to-noise ratio, dynamic range, high-frequency interference, and clipping, nor do they perform typological analysis of the correlation between the two channels. The selection results are unstable in complex recording environments.

[0003] In terms of noise suppression, mainstream methods include spectral subtraction, Wiener filtering, neural network-based noise reduction, and minimum statistical methods. Spectral subtraction is prone to producing music noise artifacts; Wiener filtering requires accurate noise priors; neural network models are large and have high power consumption for mobile inference; although minimum statistical methods do not require silent segments and are suitable for steady-state noise, they have an inherent defect of systematically underestimating noise power.

[0004] Furthermore, in multi-step processing pipelines, modules such as high-pass filtering, notch filtering, noise reduction, and level normalization are usually arranged in an empirical order of execution, lacking systematic analysis based on signal processing principles, which leads to severe negative interference between steps.

[0005] Therefore, there is an urgent need for a method and system for automatic audio enhancement across the entire recording process of mobile devices to solve the above problems. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for automatic audio enhancement of the entire recording process of a mobile device. It achieves fully automatic recording enhancement without music noise under conditions of no silence segments by using a fixed sequence of the entire recording process.

[0007] To achieve the above objectives, the present invention employs the following technical solution: On one hand, the present invention provides a method for automatic audio enhancement throughout the entire recording process of a mobile device, comprising the following steps: Step S1: Decode the input audio file to obtain multi-channel floating-point PCM data and sampling rate and channel number information; Step S2: Evaluate the channel quality of the multi-channel floating-point PCM data, select an optimal single channel, and output the audio signal of that channel; Step S3: Perform time-domain clipping repair processing on the selected single-channel audio signal. If clipping distortion exists, smooth the clipped waveform through soft clipping transformation. Step S4: Perform high-pass filtering on the signal processed in step S3 to remove DC bias and infrasound interference; Step S5: Perform power frequency notch filtering on the high-pass filtered signal to eliminate power grid fundamental frequency and its harmonic interference; Step S6: Perform adaptive frequency domain noise reduction on the signal after power frequency notch filtering. Estimate the noise power spectrum based on the minimum value statistical method and introduce bias correction. Combine decision-oriented prior signal-to-noise ratio estimation and Wiener gain calculation to perform noise reduction. Adjust the degree of noise reduction through user intensity parameters. Step S7: Perform peak level normalization processing on the noise-reduced signal to adjust the signal peak to the preset target level; Step S8: Encode the normalized mono floating-point PCM signal into the target audio format and output the enhanced audio file.

[0008] Preferably, step S2 specifically includes: For each channel, divide the frame into 20-millisecond frames, calculate the root mean square energy of each frame, and extract the frame-level root mean square energy. The following statistics for the domain: 10th percentile root mean square value 50th percentile root mean square value 90th percentile root mean square value And the peak amplitude of the entire segment; Five quality indicators were calculated based on statistical measures: Signal-to-noise ratio ; Dynamic range ; Zero crossing rate , which is the proportion of the number of sign changes between adjacent sampling points to the total number of sampling points; Clipping rate is the proportion of frames whose root mean square value exceeds the threshold of 0.99. The mute rate is the percentage of frames whose root mean square value is below a threshold of 0.001. Synthetic channel quality fraction, expressed as: ; Among them, noise floor penalty :Will Convert to The value is denoted as ,like ,but Otherwise, it is 0; zero-crossing penalty :like ,but Otherwise, it is 0; clipping penalty ,in The peak amplitude corresponding to the whole segment value; While calculating the quality score of each channel, the dual-channel signal is downsampled to no more than 100,000 points, and then the normalized cross-correlation coefficient is calculated. And classify the channel relationships according to the cross-relationship number: if If determined to be a mono copy in phase, choose any channel; if And if the similarity is not greater than 0.999, it is considered highly similar, and the path with the higher quality score is selected; if If the signal is determined to be negatively correlated, a warning will be output based on the quality score; otherwise, it will be determined to be an independent dual-microphone signal, and the channel with the highest quality score will be selected as the optimal single channel.

[0009] Preferably, step S3 specifically includes: Detect whether there are sampling points in the selected channel signal with an absolute value exceeding 0.99; if so, apply a hyperbolic tangent soft clipping transform to the entire signal segment. ,in If it does not exist, skip the clipping repair step.

[0010] Preferably, in step S4, a second-order Butterworth high-pass filter is used with a cutoff frequency of 100Hz and a quality factor of [missing information]. The filter coefficients are calculated using the RBJ audio EQ formula and implemented through a recursive difference equation.

[0011] Preferably, in step S5, the frequency is based on the power grid base frequency. and its first four harmonics , , Four second-order IIR notch filters are cascaded and applied, with each filter having a quality factor of [missing value]. The notch bandwidth is less than 3Hz, and The default is 50Hz or 60Hz.

[0012] Preferably, step S6 includes: Separate the signal by frame length Frame division, when the sampling rate is no greater than 16kHz When greater than 16kHz Frame shift is After applying a Hann window to each frame, a real FFT is performed to obtain the complex spectrum. and instantaneous power spectrum ; Smoothing the power spectrum, smoothing coefficient A smooth power spectrum was obtained. ;Continuous tracking sliding minimum value It is periodically reset every 1.5 seconds, and a deviation correction factor is used. Calculate noise power estimation ; Calculate the posterior signal-to-noise ratio The prior signal-to-noise ratio was estimated using a decision-oriented method. : ; in, For decision-making orientation coefficient, ; This is the Wiener gain for the previous frame; Calculate the Wiener gain for the current frame. Among them, the spectral base constraint Determined by a linear mapping of the user strength parameter: and will The value is limited to the range [0.005, 0.06]. The value range is [0,1]; Applying gain to the complex spectrum After performing a real IFFT, the time-domain signal is synthesized by the overlapping addition method, and the amplitude is recovered by the Hann window normalization factor to obtain the denoised signal.

[0013] Preferably, step S7 includes: Peak amplitude of the entire signal after scanning noise reduction Calculate the gain ,in The preset target level has a default value of [value]. ;Will Constrained within the range of [0.0, 100.0]; apply this linear gain to the entire signal and cut the result to [-1.0, 1.0].

[0014] On the other hand, the present invention also provides a mobile device recording end-to-end automatic audio enhancement system for implementing the above-described mobile device recording end-to-end automatic audio enhancement method, comprising: The input decoding module is used to decode the input audio file into multi-channel floating-point PCM data and extract the sampling rate and number of channels; The channel selection module is used to automatically evaluate the quality of multi-channel floating-point PCM data and select the optimal single channel. The time-domain clipping repair module is used to perform condition-triggered soft clipping repair on the selected single-channel audio signal; High-pass filter module is used to filter out DC bias and infrasound interference; Power frequency notch filter module, used to eliminate the power grid fundamental frequency and its harmonics; An adaptive frequency domain noise reduction module is used for noise reduction based on bias-corrected minimum statistical noise estimation and decision-oriented Wiener filtering, and provides a user intensity adjustment interface; The level normalization module is used to adjust the peak level of the noise-reduced signal to a preset target. The encoding output module is used to encode and output the processed mono PCM data as an enhanced audio file; The modules are connected sequentially according to the signal processing order.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs an end-to-end full-link enhancement pipeline in a defined signal processing sequence, organically integrating channel selection, time-domain preprocessing, frequency-domain noise reduction, and level normalization into a collaborative system. The output of each step provides the optimal input conditions for the next step, fundamentally eliminating the negative interference between steps caused by the independent implementation and arbitrary order of each module in traditional schemes, making the overall enhancement effect significantly better than the simple splicing of each module. 2. This invention adopts a multi-dimensional quality scoring mechanism based on frame-level root mean square distribution percentiles, using the 10th percentile as a noise floor proxy. It can comprehensively evaluate multi-dimensional indicators such as signal-to-noise ratio, dynamic range, zero-crossing rate, and clipping rate without relying on silent segments. It also classifies the relationship between the two channels by combining normalized cross-correlation coefficients, which can accurately select the optimal channel in the whole recording scenario with sound, overcoming the one-sidedness of traditional single index schemes. 3. This invention compensates for the systematic underestimation of noise power by the minimum statistical method by introducing a deviation correction factor, and adopts a strategy that combines decision-oriented prior signal-to-noise ratio estimation with Wiener gain spectral base constraint, which effectively suppresses inter-frame gain jumps and fundamentally eliminates the music noise artifacts commonly found in traditional spectral subtraction methods. At the same time, it linearly maps the internal spectral base constraint parameters to a single intensity adjustment parameter for the user, realizing intuitive noise reduction control. 4. The pure digital signal processing architecture of this invention does not require a neural network model, consumes little computing resources, is fully adaptable to offline processing scenarios on mobile devices, and balances processing quality and engineering feasibility. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the system structure of the present invention. Detailed Implementation

[0017] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined in this application.

[0018] In this invention, terms such as "upper," "lower," "left," "right," "front," "back," "vertical," "horizontal," "side," and "bottom" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used only to facilitate the description of the structural relationships of the various components or elements of this invention and do not specifically refer to any component or element in this invention. They should not be construed as limiting the invention.

[0019] Example 1: like Figure 1 As shown, this embodiment provides a method for automatic audio enhancement across the entire recording process of a mobile device, including the following steps: Step S1: Decode the input audio file to obtain multi-channel floating-point PCM data and sampling rate and channel number information; Step S2: Evaluate the channel quality of the multi-channel floating-point PCM data, select the optimal single channel, and output the audio signal of that channel; Step S3: Perform time-domain clipping repair processing on the selected single-channel audio signal. If clipping distortion exists, smooth the clipped waveform through soft clipping transformation. Step S4: Perform high-pass filtering on the signal processed in step S3 to remove DC bias and infrasound interference; Step S5: Perform power frequency notch filtering on the high-pass filtered signal to eliminate power grid fundamental frequency and its harmonic interference; Step S6: Perform adaptive frequency domain noise reduction on the signal after power frequency notch filtering. Estimate the noise power spectrum based on the minimum value statistical method and introduce bias correction. Combine decision-oriented prior signal-to-noise ratio estimation and Wiener gain calculation to perform noise reduction. Adjust the degree of noise reduction through user intensity parameters. Step S7: Perform peak level normalization processing on the noise-reduced signal to adjust the signal peak to the preset target level; Step S8: Encode the normalized mono floating-point PCM signal into the target audio format and output the enhanced audio file.

[0020] Specifically, step S1 is the input decoding process, which includes: The user recorded a conference audio segment using the built-in microphone array of the mobile device and stored it as a dual-channel WAV file with a sampling rate of 48kHz and a bit depth of 16bit. During the recording, the bottom main microphone was directly facing the sound source and the signal was clear. The top auxiliary microphone had obvious current humming and random electrical noise due to hand-held obstruction and circuit differences. The recording lasted for about 30 minutes and no silent segments were created during the recording. The system first reads the file header of the WAV file, parses the audio encoding format, sampling rate and number of channels, and converts the original 16-bit integer PCM data into dual-channel Float32 floating-point PCM data based on the parsing results, normalizing the values ​​to the range of [-1.0, 1.0]. This step unifies inputs of different formats into a standard data format that can be directly manipulated by subsequent processing modules.

[0021] Step S2 is the channel selection process, which includes: After obtaining dual-channel floating-point data, the system enters the automatic channel selection stage. Each channel is analyzed independently, with frames divided into 20-millisecond frames (non-overlapping). At a 48kHz sampling rate, each frame contains 960 sampling points. The root mean square energy (RMSE) is calculated for each frame, resulting in frame-level RMSE sequences for both channels. The RMSE values ​​are then converted to... After the domain, the 10th percentile root mean square value of each channel is extracted. 50th percentile root mean square value and the 90th percentile root mean square value At the same time, the maximum absolute value of the entire signal segment is recorded as the peak amplitude; Taking the bottom main microphone channel as an example, in its frame-level root mean square value distribution, the 10th percentile is approximately The 50th percentile is approximately The 90th percentile is approximately The peak amplitude is approximately The signal-to-noise ratio of this channel was calculated to be... The dynamic range is The zero-crossing rate is 0.08, the clipping rate is 0, and the silence rate is 0.05. The penalty values ​​for each item are calculated as follows: noise floor Less than The noise floor penalty is 0; the zero-crossing rate is 0.08, which is less than 0.15, so the zero-crossing rate penalty is 0; the clipping penalty is 0. Substituting these values ​​into the mass fraction formula: ; The top auxiliary microphone channel, due to obstruction and circuit noise, has a 10th percentile of approximately [missing value]. The 50th percentile is approximately The 90th percentile is approximately The peak amplitude is approximately The calculated signal-to-noise ratio is 18dB, the dynamic range is 30dB, the zero-crossing rate is 0.18, the clipping rate is 0, the mute rate is 0.03, and the penalty values ​​for each parameter are: noise floor. Less than The noise floor penalty is 0; the zero-crossing rate of 0.18 is greater than 0.15, so the zero-crossing rate penalty is... Clipping penalty is 0; quality score ; While calculating the scores for each channel, the system downsamples the dual-channel signals. The original signal length, converted to approximately 86,400,000 points based on the sampling rate, is downsampled proportionally to 100,000 points. The normalized cross-correlation coefficients for the two channels are then calculated, and the two downsampled signals are denoted as follows: and Calculate the inner product of the two. and their respective L2 norms and To obtain the cross-correlation number In actual recording, due to the different positions of the two microphones and the presence of additional noise in the auxiliary channel, The calculated result is approximately 0.72, which does not meet the requirements. The in-phase copy condition is not satisfied. The signal is highly similar to the given signal and not less than -0.15, therefore it is determined to be an independent dual-microphone signal. According to the processing strategy for independent dual-microphone signals, the channel with the highest quality score is selected, namely the bottom main microphone channel. As the optimal single channel, it simultaneously generates a structured report containing score details for each channel and conclusions on channel relationships.

[0022] Step S3 is the time-domain clipping repair process, which includes: After selecting a channel, the system scans the signal of that channel point by point to check for any sampling points with an absolute value exceeding 0.99. The detection results show that the peak amplitude is approximately... The corresponding linear value is approximately 0.89, and no sampling point exceeds the 0.99 threshold. Therefore, it is determined that there is no clipping distortion. According to the conditional triggering mechanism, this step is skipped, and the process proceeds directly to the next stage. If clipping due to ADC overload exists in the actual recording, a flat top close to ±1.0 will appear in the signal. In this case, the system will process the entire signal according to... Perform soft clipping transformation point by point to smooth the steep edges caused by clipping.

[0023] Step S4 is the high-pass filtering process, which includes: The system sends the mono Float32 signal to a high-pass filter module. The filter type is a second-order Butterworth high-pass filter with a cutoff frequency set to 100Hz and a quality factor of [missing information]. Pick (Approximately 0.7071), the filter coefficients are calculated using the RBJ audio EQ formula: first, the angular frequency is calculated based on the cutoff frequency and sampling rate. Then calculate the intermediate quantity. Based on this, feedforward coefficients are generated. , , and feedback coefficient , The filtering process uses a recursive difference equation to calculate the current output point by point. ,in Sample the current input. and To input history, and For outputting historical data, this filter effectively suppresses DC bias and infrasound interference below 100Hz in the signal.

[0024] Step S5 is the power frequency notch filtering process, which includes: The high-pass filtered signal enters the power frequency notch filter module. In this embodiment, the power grid standard is 50Hz, therefore the base frequency is... Set to 50Hz, the system constructs four second-order IIR notch filters sequentially, targeting the fundamental frequency of 50Hz and its four preceding harmonics of 100Hz, 150Hz, and 200Hz, respectively. The quality factor of each notch filter is... All values ​​are set to 35. Precise notch filtering is performed on the target frequency using an extremely narrow bandwidth within 3Hz. Taking a 50Hz notch filter as an example: Calculate the angular frequency. , Based on this, filter coefficients are generated. The four filters are cascaded in the order of 50Hz, 100Hz, 150Hz, and 200Hz. The output of the previous notch filter is used as the input of the next notch filter. Each filter independently maintains its own state variables. After four-stage cascading, the power grid frequency and its harmonic interference are precisely removed, while the voice fundamental frequency region above 100Hz is almost unaffected.

[0025] Step S6 is the adaptive frequency domain noise reduction process, including: After completing the above time-domain preprocessing, the signal has the characteristics of being clean, free of DC bias, and free of power frequency harmonic interference, providing the optimal input conditions for frequency domain noise estimation. The system first determines the framing parameters based on the sampling rate. The current sampling rate is 48kHz, which is greater than 16kHz, therefore the frame length is... Take 1024 points, frame shifting That is, 512 points. The signal is divided into overlapping frames starting from the beginning, with a frame length of 1024 points and a frame shift of 512 points. A Hann window function is applied to each frame. , The value ranges from 0 to 1023. After windowing, a 1024-point real-valued FFT is performed on each frame to obtain the complex spectrum of that frame. and instantaneous power spectrum , Frequency index; The noise power spectrum estimation employs the minimum statistical method and introduces a bias correction. The system performs first-order recursive smoothing on the power spectrum value at each frequency point using a smoothing coefficient α = 0.98. ,in This represents the smoothed power value for the corresponding frequency point in the previous frame, while continuously tracking the sliding minimum of the smoothed power spectrum for each frequency point over a past period. The system is configured to periodically reset the sliding minimum window approximately every 1.5 seconds to maintain its ability to track noise changes. After obtaining the sliding minimum, the noise power is estimated according to... The calculation includes a deviation correction factor of 1.67 to compensate for the systematic underestimation of noise power by the minimum statistical method. Based on noise power estimation, the system calculates the posterior signal-to-noise ratio frame by frame and frequency by frequency. Prior signal-to-noise ratio Estimation using a decision-oriented approach: ; in This is the Wiener gain for the frequency corresponding to the previous frame. The a posteriori signal-to-noise ratio (SNR) of the frequency point corresponding to the previous frame is used for the first frame or the first frame after noise reset. As an initial estimate of the prior signal-to-noise ratio; Wiener gain calculation Among them, the spectral base constraint Intensity parameters set by the user Sure, This is a single noise reduction intensity parameter that the user can adjust in the application interface, with a value range of [0,1]. The system will... Linear mapping to spectral base constraint values: And limited to the closed interval [0.005, 0.06], when the user... When adjusted to 0, floor=0.06, achieving mild noise reduction to preserve the naturalness of the audio as much as possible; when adjusted to 1... This enables powerful noise reduction to fully suppress background noise. This mapping relationship allows users to intuitively control the level of noise reduction with just a single slider, without needing to understand the internal parameters. The gain application stage multiplies the Wiener gain by a complex spectrum at each frequency point to obtain the denoised spectrum. ,right Perform a 1024-point real IFFT to convert the frequency domain signal back to the time domain. Since the framing uses 50% overlap, there is a 512-point overlap area between adjacent frames. The system adds the 0th to 511th sampling points of the current frame's IFFT result to the corresponding 512 sampling points in the previous frame's processing result to complete the overlap-addition synthesis. The synthesized signal is divided by the normalization factor of the Hann window to restore it to the original amplitude level. The system can perform this noise reduction processing asynchronously in an independent thread or isolated computing environment. After processing, the noise reduction signal is returned to the main thread through a callback mechanism to avoid blocking the user interface.

[0026] Step S7 is the peak AGC normalization process, which includes: After noise reduction, the system scans the absolute value of the entire signal segment to obtain the peak amplitude. Calculate the target gain ,in Take the default value ,Will The signal is limited to the range [0.0, 100.0] to prevent excessive amplification of extremely quiet signals. Each sample point of the entire signal is multiplied by [the specified value]. The value is calculated, and results exceeding 1.0 are truncated to 1.0, and results below -1.0 are truncated to -1.0, thus completing the normalization of the signal peak level. This step is performed after noise reduction to ensure that the level normalization is calculated based on the peak value of the clean signal and does not amplify the residual noise floor.

[0027] Step S8 is the encoding output process, which includes: The normalized mono Float32 PCM signal is sent to the encoding output module. The system calls the corresponding encoder to convert the Float32 data into the target format according to the user's preset output format (MP3 in this embodiment). The encoding parameters adopt the default high-quality configuration, and the output bitrate is consistent with or slightly lower than that of the input file. Finally, the enhanced MP3 audio file is generated and saved to the device storage space.

[0028] like Figure 2 As shown, this embodiment also provides a mobile device recording end-to-end automatic audio enhancement system, including: The input decoding module is used to decode the input audio file into multi-channel floating-point PCM data and extract the sampling rate and number of channels; The channel selection module is used to automatically evaluate the quality of multi-channel floating-point PCM data and select the optimal single channel. The time-domain clipping repair module is used to perform condition-triggered soft clipping repair on the selected single-channel audio signal; High-pass filter module is used to filter out DC bias and infrasound interference; Power frequency notch filter module, used to eliminate the power grid fundamental frequency and its harmonics; An adaptive frequency domain noise reduction module is used for noise reduction based on bias-corrected minimum statistical noise estimation and decision-oriented Wiener filtering, and provides a user intensity adjustment interface; The level normalization module is used to adjust the peak level of the noise-reduced signal to a preset target. The encoding output module is used to encode and output the processed mono PCM data as an enhanced audio file; The modules are connected sequentially according to the signal processing order.

[0029] Example 2: This embodiment describes an implementation method in which an end-to-end neural network noise reduction is used to replace the cascaded DSP noise reduction section in an alternative solution.

[0030] In the automatic audio enhancement method for mobile device recording described in Embodiment 1 above, the adaptive frequency domain noise reduction step S6 can be replaced by an end-to-end noise reduction scheme based on neural networks. At this time, the processing flow of steps S1 to S5 remains unchanged, still executing input decoding, channel selection, time-domain clipping repair, high-pass filtering, and power frequency notch filtering in sequence to provide a pre-processed clean signal for the noise reduction stage. During the noise reduction stage, the system loads a pre-trained lightweight speech noise reduction neural network model, which can adopt publicly available architectures such as RNNoise or DTLN. The pre-processed time-domain signal is divided into frames according to the frame length and format required by the model, and each frame is fed into the neural network for inference. The model outputs a denoised time-domain signal frame, which is then synthesized into a complete denoised signal using an overlap-addition method. The noise reduction intensity is determined by the model's internal parameters, and users can indirectly control the intensity by selecting different pre-trained models (such as a light noise reduction model or a strong noise reduction model). The peak AGC normalization in step S7 and the encoding output in step S8 after noise reduction are the same as in Embodiment 1.

[0031] In this alternative, the corresponding adaptive frequency domain noise reduction module is replaced with a neural network noise reduction module, while the composition and connection relationships of other system modules remain unchanged.

[0032] It should be noted that neural network alternatives generally outperform pure DSP solutions in noise reduction quality, especially in non-stationary noise scenarios. However, the model file size is large (several MB to tens of MB), significantly increasing the inference computation load and placing higher demands on the performance and power consumption of mobile processors. In contrast, the pure DSP solution described in Example 1 requires no model file, has minimal memory footprint, and can achieve low-power real-time or offline processing on various mobile devices. In practical engineering deployments, the appropriate solution can be flexibly selected based on the hardware capabilities of the target device and the application scenario.

[0033] The above is a detailed description of the preferred embodiments of the present invention, but the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A method for automatic audio enhancement across the entire recording process of a mobile device, characterized in that, Includes the following steps: Step S1: Decode the input audio file to obtain multi-channel floating-point PCM data and sampling rate and channel number information; Step S2: Evaluate the channel quality of the multi-channel floating-point PCM data, select an optimal single channel, and output the audio signal of that channel; Step S3: Perform time-domain clipping repair processing on the selected single-channel audio signal. If clipping distortion exists, smooth the clipped waveform through soft clipping transformation. Step S4: Perform high-pass filtering on the signal processed in step S3 to remove DC bias and infrasound interference; Step S5: Perform power frequency notch filtering on the high-pass filtered signal to eliminate power grid fundamental frequency and its harmonic interference; Step S6: Perform adaptive frequency domain noise reduction on the signal after power frequency notch filtering. Estimate the noise power spectrum based on the minimum value statistical method and introduce bias correction. Combine decision-oriented prior signal-to-noise ratio estimation and Wiener gain calculation to perform noise reduction. Adjust the degree of noise reduction through user intensity parameters. Step S7: Perform peak level normalization processing on the noise-reduced signal to adjust the signal peak to the preset target level; Step S8: Encode the normalized mono floating-point PCM signal into the target audio format and output the enhanced audio file.

2. The automatic audio enhancement method for the entire recording process of a mobile device according to claim 1, characterized in that, Step S2 is as follows: For each channel, divide the frame into 20-millisecond frames, calculate the root mean square energy of each frame, and extract the frame-level root mean square energy. The following statistics for the domain: 10th percentile root mean square value 50th percentile root mean square value 90th percentile root mean square value And the peak amplitude of the entire segment; Five quality indicators were calculated based on statistical measures: Signal-to-noise ratio ; Dynamic range ; Zero crossing rate , which is the proportion of the number of sign changes between adjacent sampling points to the total number of sampling points; Clipping rate is the proportion of frames whose root mean square value exceeds the threshold of 0.

99. The mute rate is the percentage of frames whose root mean square value is below a threshold of 0.

001. Synthetic channel quality fraction, expressed as: ； Among them, noise floor penalty :Will Convert to The value is denoted as ,like ,but Otherwise, it is 0; zero-crossing penalty :like ,but Otherwise, it is 0; clipping penalty ,in The peak amplitude corresponding to the whole segment value; While calculating the quality score of each channel, the dual-channel signal is downsampled to no more than 100,000 points, and then the normalized cross-correlation coefficient is calculated. And classify the channel relationships according to the cross-relationship number: if If determined to be a mono copy in phase, choose any channel; if And if the similarity is not greater than 0.999, it is considered highly similar, and the path with the higher quality score is selected; if If the signal is determined to be negatively correlated, a warning will be output based on the quality score; otherwise, it will be determined to be an independent dual-microphone signal, and the channel with the highest quality score will be selected as the optimal single channel.

3. The automatic audio enhancement method for the entire recording process of a mobile device according to claim 1, characterized in that, Step S3 is as follows: Detect whether there are sampling points in the selected channel signal with an absolute value exceeding 0.99; if so, apply a hyperbolic tangent soft clipping transform to the entire signal segment. ,in If it does not exist, skip the clipping repair step.

4. The automatic audio enhancement method for the entire recording process of a mobile device according to claim 1, characterized in that, In step S4, a second-order Butterworth high-pass filter is used with a cutoff frequency of 100Hz and a quality factor of [missing information]. The filter coefficients are calculated using the RBJ audio EQ formula and implemented through a recursive difference equation.

5. The automatic audio enhancement method for the entire recording process of a mobile device according to claim 1, characterized in that, In step S5, according to the power grid base frequency and its first four harmonics , , Four second-order IIR notch filters are cascaded and applied, with each filter having a quality factor of [missing value]. The notch bandwidth is less than 3Hz, and The default is 50Hz or 60Hz.

6. The automatic audio enhancement method for the entire recording process of a mobile device according to claim 1, characterized in that, Step S6 includes: Separate the signal by frame length Frame division, when the sampling rate is no greater than 16kHz When greater than 16kHz Frame shift is After applying a Hann window to each frame, a real FFT is performed to obtain the complex spectrum. and instantaneous power spectrum ; Smoothing the power spectrum, smoothing coefficient A smooth power spectrum was obtained. ;Continuous tracking sliding minimum value It is periodically reset every 1.5 seconds, and a deviation correction factor is used. Calculate noise power estimation ; Calculate the posterior signal-to-noise ratio The prior signal-to-noise ratio was estimated using a decision-oriented method. : ； in, For decision-making orientation coefficient, ; This is the Wiener gain for the previous frame; Calculate the Wiener gain for the current frame. Among them, the spectral base constraint Determined by a linear mapping of the user strength parameter: and will The value is limited to the range [0.005, 0.06]. The value range is [0,1]; Applying gain to the complex spectrum After performing a real IFFT, the time-domain signal is synthesized by the overlapping addition method, and the amplitude is recovered by the Hann window normalization factor to obtain the denoised signal.

7. The automatic audio enhancement method for the entire recording process of a mobile device according to claim 1, characterized in that, Step S7 includes: Peak amplitude of the entire signal after scanning noise reduction Calculate the gain ,in The preset target level has a default value of [value]. ;Will Constrained within the range of [0.0, 100.0]; apply this linear gain to the entire signal and cut the result to [-1.0, 1.0].

8. A mobile device recording end-to-end automatic audio enhancement system, used to implement the mobile device recording end-to-end automatic audio enhancement method as described in any one of claims 1-7, characterized in that, include: The input decoding module is used to decode the input audio file into multi-channel floating-point PCM data and extract the sampling rate and number of channels; The channel selection module is used to automatically evaluate the quality of multi-channel floating-point PCM data and select the optimal single channel. The time-domain clipping repair module is used to perform condition-triggered soft clipping repair on the selected single-channel audio signal; High-pass filter module is used to filter out DC bias and infrasound interference; Power frequency notch filter module, used to eliminate the power grid fundamental frequency and its harmonics; An adaptive frequency domain noise reduction module is used for noise reduction based on bias-corrected minimum statistical noise estimation and decision-oriented Wiener filtering, and provides a user intensity adjustment interface; The level normalization module is used to adjust the peak level of the noise-reduced signal to a preset target. The encoding output module is used to encode and output the processed mono PCM data as an enhanced audio file; The modules are connected sequentially according to the signal processing order.

Images