An AI intelligent noise reduction method based on laser modulation voice

Abstract

This invention relates to the field of noise reduction technology, specifically to an AI-based intelligent noise reduction method for laser-modulated speech, comprising the following steps: Analog-to-digital conversion: A high-speed, high-precision analog-to-digital conversion module is used to convert analog data into digital signals, and noise reduction is performed only on human voices according to the usage scenario; Digital filtering: Bandpass filtering technology is used to filter out all out-of-band signals; AI intelligent processing: Spectral subtraction is used to eliminate background noise. By constructing a collaborative processing link of "high-precision analog-to-digital / digital-to-analog conversion—64th-order narrow transition band FIR filtering—dynamic spectral subtraction based on AI model library", for slow-changing and stable specific background noise such as mechanical vibration thermal noise of the laser itself and ambient optical path scattering interference, a 10-millisecond non-overlapping time slice and a joint decision mechanism of three parameters (mean, variance, and entropy) of Mel frequency cepstral coefficients are used to achieve accurate dynamic tracking of noise targets and adaptive stripping of speech signals.

Description

Technical Field

[0001] This invention relates to the field of noise reduction technology, and more specifically, to an AI-based intelligent noise reduction method for laser-modulated speech. Background Technology

[0002] To address noise cancellation in voice communication, various algorithms have been developed to analyze and modify audio signals, including:

[0003] Traditional digital signal processing methods:

[0004] Filtering technology: The audio frequency range is 20Hz~20KHz, and the human voice frequency range is 85Hz~1100Hz. For specific usage environments, digital filters (such as low-pass, high-pass, and band-pass filters) are designed to remove high and low frequency noise outside the specific frequency range.

[0005] Spectral subtraction: Spectral subtraction is a common spectrum-based noise reduction method that uses frequency domain analysis and signal-to-noise ratio estimation to reduce background noise. It first estimates the spectral characteristics of the background noise in the silence or non-speech segments of the speech signal, and then subtracts the estimated noise spectrum from the spectrum of the original signal to obtain a clear speech spectrum.

[0006] Noise thresholding: A volume threshold is set; signals below this threshold are considered noise and are removed. This method is suitable for scenarios where there is a significant difference between noise and speech volume, but it may cause loss of speech details and ending sounds.

[0007] Sampling-based noise reduction: This method is an effective noise reduction technique used in professional audio processing software, especially suitable for removing continuous and stable background noise. Its principle is to sample the noise waveform and then analyze and compare it with the waveform of the entire audio file, thereby automatically removing the noise. However, this method has the disadvantage of potentially damaging the original vocal quality, and the lower the signal-to-noise ratio, the greater the damage. Therefore, this method is generally not recommended for noise reduction of singing voices.

[0008] Differential input method: Receives at least two or more voice signals, and subtracts the reference noise signal from one signal with superimposed voice noise, thereby eliminating background noise.

[0009] Noise smoothing generally includes mean filtering and median filtering. Mean filtering replaces the value of the current pixel with the average value of the surrounding pixels. Mean filtering is effective at removing coarse noise, but it is not effective at removing fine-texture noise and high-frequency noise. Median filtering replaces the value of the current pixel with the median value of the surrounding pixels. Median filtering is effective at removing fine-texture noise, but it is not effective at removing coarse noise and high-frequency noise.

[0010] The core principle of laser-modulated speech is to convert sound vibration information into laser signal modulation, and then demodulate to restore the speech. Near-infrared lasers (commonly 1550nm wavelength) in the invisible light band are used to directionally illuminate the surface of objects with low acoustic impedance, such as paper cups, paper boxes, paper bags, tissue boxes, and biscuit boxes. The nanometer-level vibrations caused by the sound result in a Doppler frequency shift in the reflected light. The frequency shift is linearly related to the vibration amplitude, thus converting sound wave information into light waves and achieving high-precision sound acquisition. The receiver uses heterodyne detection technology to convert the light signal into an electrical signal. After noise reduction by a bandpass filter (300Hz-3.4kHz), a lock-in amplifier is used to extract the speech signal.

[0011] Laser-based voice acquisition is limited by various factors, such as:

[0012] Due to the limitation of the effective area of ​​the light spot, the reflectivity of the target object must be >10% (ordinary glass is about 4-8%).

[0013] The signal weakens by approximately 20dB for every 100 meters of distance (it needs to remain effective within 500 meters).

[0014] The temperature gradient causes a change in the refractive index of air, resulting in a beam drift of approximately 0.1 mrad / °C.

[0015] Rainfall attenuation coefficient: Loss increases by 3 dB / km when rainfall intensity is 5 mm / h.

[0016] Therefore, laser-modulated speech contains noise generated by the laser device and various complex background noises, necessitating a device to improve sound quality and remove various noises. Existing noise removal devices include:

[0017] Active noise cancellation: It uses electronic technology to generate reverse sound waves to cancel out noise, and is commonly found in headphones and some professional equipment.

[0018] Passive noise reduction devices: These devices use physical structures or sound-absorbing materials to block or absorb sound waves, such as sound insulation cotton and silencers.

[0019] Common ground isolator: It cuts off the grounding loop through physical isolation, and is specifically designed to deal with current noise and buzzing. It is the most effective solution.

[0020] Audio isolators: These use transformers or optoelectronic technology to isolate signals, filtering out a wider range of noise, such as radio frequency interference, while maintaining pure sound quality.

[0021] None of these methods can meet the noise reduction requirements of laser-modulated speech. Therefore, a dedicated noise reduction device was developed for this application scenario to extract high-definition speech signals. Summary of the Invention

[0022] To address the problems mentioned in the background art, this application provides an AI-based intelligent noise reduction method for laser-modulated speech, comprising the following steps:

[0023] Analog-to-digital conversion: A high-speed, high-precision analog-to-digital conversion module is used to convert analog data into digital signals. Noise reduction is performed only for human voice, depending on the usage scenario.

[0024] Digital filtering: Bandpass filtering technology is used to filter out all out-of-band signals;

[0025] AI intelligent processing: Uses spectral subtraction to eliminate background noise;

[0026] Digital-to-analog conversion: A high-speed, high-precision digital-to-analog conversion module is used to convert data into analog signals.

[0027] Audio output: Two output modes are available: headphones and speakers.

[0028] As a preferred embodiment of the present invention, bandpass filtering is an electronic filtering method that allows signals within a specific frequency range to pass through while suppressing other frequencies. It consists of a cascaded low-pass filter and a high-pass filter, where the high-pass filter filters out low-frequency signals and the low-pass filter filters out high-frequency signals, while retaining the selected frequency signals.

[0029] As a preferred embodiment of the present invention, the signal processed by digital filtering is a single-variable function sequence, and a one-dimensional 64th order Hamming window function method FIR digital filter is selected. For FIR digital filters, the higher the order, the narrower the transition band, and the computational load increases linearly. For each additional order, the amplitude-frequency response roll-off rate increases by 6 dB / octave. The width of the transition band is inversely proportional to the order. Higher order designs are used to achieve a narrower transition band.

[0030] As a preferred embodiment of the present invention, the specific parameters of the FIR digital filter are as follows:

[0031] Speech signal sampling rate: 10kHz;

[0032] Passband edge frequencies: fp1 = 50 Hz, fp2 = 4 kHz;

[0033] Stopband edge frequencies: fs1=0Hz, fs2=4.3kHz;

[0034] Passband ripple: less than 0.5 dB;

[0035] Minimum stopband attenuation: greater than 40dB;

[0036] Transition bandwidth: Δf = 300Hz.

[0037] As a preferred embodiment of the present invention, the prerequisite for spectral subtraction in AI intelligent processing to eliminate background noise is:

[0038] Noise is a slow variable;

[0039] Quickly distinguish between noise and signal: Using AI training, the spectral characteristics of human voice are obtained in advance to generate an AI model library. Each 10ms data is used as a time slice. The spectral characteristics of each time slice are compared with the human voice characteristics in the model library. If they do not meet the requirements, they are identified as noise and automatically eliminated, and the noise target is dynamically replaced. If the human voice characteristics are met, the noise target is subtracted from the signal to remove the background noise.

[0040] The obtained continuous speech is divided into time slices, with each slice being 10ms in unit. The data in each time slice is then subjected to Fourier transform processing to convert it into a frequency signal.

[0041] As a preferred embodiment of the present invention, the specific operation of the Fourier transform is as follows:

[0042] Finite-length truncation: Truncate the sampled signal to 10ms, that is, 100 sampling points as a transformation unit, with no overlap between adjacent time slices;

[0043] Windowing: To reduce spectral leakage, a Hamming window is used to smooth signal edges and reduce sidelobe levels;

[0044] Zero padding: The sampling sequence is padded with zeros to 128 points to increase the density of the spectrum display, thereby increasing the number of FFT output points and making the curve smoother, without improving the true resolution;

[0045] Call the FFT algorithm to perform FFT calculation;

[0046] Amplitude normalization: to obtain the true amplitude, multiply the single-sided spectrum amplitude by 2 / 100, and multiply the DC component by only 1 / 100.

[0047] As a preferred embodiment of the present invention, the specific process of eliminating background noise by spectral subtraction is as follows:

[0048] MFCC extracts the spectral envelope of speech by simulating the characteristics of human hearing. Noise in MFCC coefficients is characterized by low energy, high entropy, and irregular distribution.

[0049] In the speech activity detection algorithm, speech frames are distinguished from non-speech frames by using a speech threshold in the AI ​​chip;

[0050] Initialize by setting the noise target signal to 0;

[0051] Obtain the spectral information of the time slice and calculate the statistical characteristics of the Mel frequency cepstral coefficients;

[0052] The mean, variance, and entropy of the Mel frequency cepstral coefficients are obtained and compared with the threshold values ​​of the mean, variance, and entropy of the Mel frequency cepstral coefficients stored in the AI ​​chip. If two or more conditions are met (mean and variance are greater than the threshold, and entropy is less than the threshold), it is determined to be a signal; otherwise, it is determined to be noise.

[0053] If it is noise, replace the original noise marker signal with the time slice signal and output 0 (no sound); if it is a speech signal, subtract the noise marker signal from the speech signal and output the speech signal.

[0054] As a preferred embodiment of the present invention, the specific process for obtaining the mean, variance, and entropy of the Mel frequency cepstral coefficients is as follows:

[0055] Obtain the Mel-frequency cepstral coefficients of human voice;

[0056] First-order high-pass filtering compensates for the attenuation of high-frequency resonant peaks in human voices and improves the signal-to-noise ratio;

[0057] The signal is cut into short frames of 20–30 ms, with each frame overlapping by 50%, frame length: 25 ms, frame shift: 10 ms, so that the non-stationary speech is approximated as a stationary signal;

[0058] Perform FFT on each frame to calculate the power spectrum, and pass it through 20–40 triangular Mel filters covering 0–8kHz to simulate the nonlinear frequency perception of the human ear, compress high-frequency redundancy, and highlight key frequency bands of speech.

[0059] Take the logarithmic energy and perform a discrete cosine transform, retain the first 12 coefficients C1–C12, compress the dynamic range, remove frequency domain correlation, and extract the spectral envelope;

[0060] The input to the Discrete Cosine Transform is the logarithmic energy of the output of the 24–40 Vimel filter bank, and the output consists of 12 coefficients C1–C12, with C0 retained separately.

[0061] Through deep learning, the mean, variance, and entropy thresholds of the Mel frequency cepstral coefficients that are closest to human voice are obtained.

[0062] As a preferred embodiment of the present invention, the analog-to-digital converter module is model ADS8864, and the digital-to-analog converter module is model DAC8563.

[0063] As a preferred embodiment of the present invention, the sampling rate of the analog-to-digital conversion module is set to 10KSamples / S during the analog-to-digital conversion process, and the effective bandwidth during the digital filtering process is set to 50Hz~4KHz.

[0064] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0065] This invention provides a high-fidelity AI-powered intelligent noise reduction method for laser-modulated voice transmission scenarios. Its advantages lie in constructing a collaborative processing chain of "high-precision analog-to-digital / digital-to-analog conversion—64th-order narrow-transition-band FIR filtering—dynamic spectral subtraction based on an AI model library." Targeting slow-varying and stable background noise such as laser mechanical vibration thermal noise and ambient optical path scattering interference, it utilizes a 10-millisecond-level non-overlapping time slice and a joint decision mechanism based on the mean, variance, and entropy of the Mel-frequency cepstral coefficients to achieve precise dynamic tracking of the noise target and accurate voice signal reduction. Adaptive stripping effectively eliminates the "musical noise" and consonant false cancellation problems that are easily generated by traditional spectral subtraction. At the same time, combined with first-order high-pass pre-emphasis to compensate for the attenuation of high-frequency resonant peaks in laser transmission, it retains sibilant and fricative details with an extremely narrow 300Hz transition band within a sampling rate of only 10 kHz and an effective bandwidth of 50 Hz to 4 kHz. It also matches the wide dynamic range of a high-precision analog-to-digital converter to reduce quantization noise, thereby improving the intelligibility, brightness and background quietness of laser speech. This allows the output audio to maintain a natural listening experience and high-fidelity reproduction capability while suppressing specific physical noise. Attached Figure Description

[0066] Figure 1 This is a flowchart illustrating the overall process of the present invention. Detailed Implementation

[0067] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0068] Please see Figure 1 As shown, a preferred embodiment of the present invention provides an AI-based intelligent noise reduction method for laser-modulated speech, comprising the following steps:

[0069] Analog-to-digital conversion: A high-speed, high-precision analog-to-digital conversion module is used to convert analog data into digital signals. Noise reduction is performed only for human voice, depending on the usage scenario.

[0070] Digital filtering: Bandpass filtering technology is used to filter out all out-of-band signals;

[0071] AI intelligent processing: Uses spectral subtraction to eliminate background noise;

[0072] Digital-to-analog conversion: A high-speed, high-precision digital-to-analog conversion module is used to convert data into analog signals.

[0073] Audio output: Two output modes are available: headphones and speakers.

[0074] Bandpass filtering is an electronic filtering method that allows signals within a specific frequency range to pass through while suppressing other frequencies. It consists of a cascaded low-pass filter and a high-pass filter. The high-pass filter filters out low-frequency signals, and the low-pass filter filters out high-frequency signals, retaining the selected frequency signal.

[0075] The signal processed by digital filtering is a single-variable function sequence. A one-dimensional 64th-order Hamming window function FIR digital filter is selected. For FIR digital filters, the higher the order, the narrower the transition band. The computational load increases linearly. For each additional order, the amplitude-frequency response roll-off rate increases by 6 dB / octave. The width of the transition band is inversely proportional to the order. Higher-order designs are used to achieve a narrower transition band.

[0076] The specific parameters of the FIR digital filter are as follows:

[0077] Speech signal sampling rate: 10kHz;

[0078] Passband edge frequencies: fp1 = 50 Hz, fp2 = 4 kHz;

[0079] Stopband edge frequencies: fs1=0Hz, fs2=4.3kHz;

[0080] Passband ripple: less than 0.5 dB;

[0081] Minimum stopband attenuation: greater than 40dB;

[0082] Transition bandwidth: Δf = 300Hz.

[0083] The prerequisite for spectral subtraction in AI intelligent processing to eliminate background noise is:

[0084] Noise is a slow variable; this is determined by the acquisition scenario of the laser-modulated speech (mainly for indoor scenarios), and its noise mainly includes two aspects:

[0085] 1. Laser noise: Lasers generate some noise during operation, primarily due to heat and mechanical vibration from the internal optical and electronic components. This is considered system noise and remains relatively constant throughout operation. While this noise is relatively minor, it can interfere with the transmission and reception of sound signals in high-precision applications.

[0086] 2. Environmental Noise: Environmental noise is also a significant factor in laser sound transmission. For example, dust, water vapor, and other impurities in the air may scatter and absorb the laser beam, affecting the transmission quality of the optical signal. Furthermore, mechanical vibrations and electromagnetic interference from the external environment can also impact the laser sound transmission system. However, because the environment in which the voice is acquired is relatively stable, this noise changes slowly over time and does not fluctuate significantly.

[0087] Quickly distinguish between noise and signal: Using AI training, the spectral characteristics of human voice are obtained in advance to generate an AI model library. Each 10ms data is used as a time slice. The spectral characteristics of each time slice are compared with the human voice characteristics in the model library. If they do not meet the requirements, they are identified as noise and automatically eliminated, and the noise target is dynamically replaced. If the human voice characteristics are met, the noise target is subtracted from the signal to remove the background noise.

[0088] The obtained continuous speech is divided into time slices, with 10ms as a slice unit. The data of each time slice is processed by Fourier transform and converted into frequency signals.

[0089] AI training process: Build a training platform using DeepSeek; collect speech signals (corpus) online; filter the corpus, keeping only pure human voices; use the retained corpus to train on the training platform to obtain the frequency features of the speech; store the frequency features in the AI ​​model library.

[0090] The specific operations of Fourier transform are as follows:

[0091] Finite-length truncation: Truncate the sampled signal to 10ms, that is, 100 sampling points as a transformation unit, with no overlap between adjacent time slices;

[0092] Windowing: To reduce spectral leakage, a Hamming window is used to smooth signal edges and reduce sidelobe levels;

[0093] Zero padding: The sampling sequence is padded with zeros to 128 points to increase the density of the spectrum display, thereby increasing the number of FFT output points and making the curve smoother, without improving the true resolution;

[0094] Call the FFT algorithm to perform FFT calculation;

[0095] Amplitude normalization: to obtain the true amplitude, multiply the single-sided spectrum amplitude by 2 / 100, and multiply the DC component by only 1 / 100.

[0096] The specific process of eliminating background noise using spectral subtraction is as follows:

[0097] MFCC extracts the spectral envelope of speech by simulating the characteristics of human hearing. Noise in MFCC coefficients is characterized by low energy, high entropy, and irregular distribution.

[0098] In the speech activity detection algorithm, speech frames are distinguished from non-speech frames by using a speech threshold in the AI ​​chip;

[0099] Initialize by setting the noise target signal to 0;

[0100] Obtain the spectral information of the time slice and calculate the statistical characteristics of the Mel frequency cepstral coefficients;

[0101] The mean, variance, and entropy of the Mel frequency cepstral coefficients are obtained and compared with the threshold values ​​of the mean, variance, and entropy of the Mel frequency cepstral coefficients stored in the AI ​​chip. If two or more conditions are met (mean and variance are greater than the threshold, and entropy is less than the threshold), it is determined to be a signal; otherwise, it is determined to be noise.

[0102] If it is noise, replace the original noise marker signal with the time slice signal and output 0 (no sound); if it is a speech signal, subtract the noise marker signal from the speech signal and output the speech signal.

[0103] The specific process for obtaining the mean, variance, and entropy of the Mel frequency cepstral coefficients is as follows:

[0104] Obtain the Mel-frequency cepstral coefficients of human voice;

[0105] First-order high-pass filtering compensates for the attenuation of high-frequency resonant peaks in human voices and improves the signal-to-noise ratio;

[0106] The signal is cut into short frames of 20–30 ms, with each frame overlapping by 50%, frame length: 25 ms, frame shift: 10 ms, so that the non-stationary speech is approximated as a stationary signal;

[0107] Perform FFT on each frame to calculate the power spectrum, and pass it through 20–40 triangular Mel filters covering 0–8kHz to simulate the nonlinear frequency perception of the human ear, compress high-frequency redundancy, and highlight key frequency bands of speech.

[0108] Take the logarithmic energy and perform a discrete cosine transform, retain the first 12 coefficients C1–C12, compress the dynamic range, remove frequency domain correlation, and extract the spectral envelope;

[0109] The input to the Discrete Cosine Transform is the logarithmic energy of the output of the 24–40 Vimel filter bank, and the output consists of 12 coefficients C1–C12, with C0 retained separately.

[0110] Through deep learning, the mean, variance, and entropy thresholds of the Mel frequency cepstral coefficients that are closest to human voice are obtained.

[0111] By using deep learning models (such as deep neural networks) to learn the pattern differences between speech and noise, more accurate noise recognition and suppression can be achieved, especially in handling complex and non-steady-state environmental noise (such as street noise and crowd noise).

[0112] The analog-to-digital converter module is model ADS8864, and the digital-to-analog converter module is model DAC8563.

[0113] During the analog-to-digital conversion process, the sampling rate of the analog-to-digital conversion module is set to 10KSamples / S, and the effective bandwidth during the digital filtering process is set to 50Hz~4KHz.

[0114] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. An AI-powered intelligent noise reduction method for laser-modulated speech, characterized in that, Includes the following steps: Analog-to-digital conversion: A high-speed, high-precision analog-to-digital conversion module is used to convert analog data into digital signals. Noise reduction is performed only for human voice, depending on the usage scenario. Digital filtering: Bandpass filtering technology is used to filter out all out-of-band signals; AI intelligent processing: Uses spectral subtraction to eliminate background noise; Digital-to-analog conversion: A high-speed, high-precision digital-to-analog conversion module is used to convert data into analog signals; Audio output: Two output modes are available: headphones and speakers.

2. The AI ​​intelligent noise reduction method for laser-modulated speech according to claim 1, characterized in that, Bandpass filtering is an electronic filtering method that allows signals within a specific frequency range to pass through while suppressing other frequencies. It consists of a cascaded low-pass filter and a high-pass filter. The high-pass filter filters out low-frequency signals, and the low-pass filter filters out high-frequency signals, retaining the selected frequency signal.

3. The AI-powered intelligent noise reduction method for laser-modulated speech according to claim 1, characterized in that, The signal processed by digital filtering is a single-variable function sequence. A one-dimensional 64th-order Hamming window function FIR digital filter is selected. For FIR digital filters, the higher the order, the narrower the transition band. The computational load increases linearly. For each additional order, the amplitude-frequency response roll-off rate increases by 6 dB / octave. The width of the transition band is inversely proportional to the order. Higher-order designs are used to achieve a narrower transition band.

4. The AI ​​intelligent noise reduction method for laser-modulated speech according to claim 3, characterized in that, The specific parameters of the FIR digital filter are as follows: Speech signal sampling rate: 10kHz; Passband edge frequencies: fp1 = 50 Hz, fp2 = 4 kHz; Stopband edge frequencies: fs1=0Hz, fs2=4.3kHz; Passband ripple: less than 0.5 dB; Minimum stopband attenuation: greater than 40dB; Transition bandwidth: Δf = 300Hz.

5. The AI-powered intelligent noise reduction method for laser-modulated speech according to claim 1, characterized in that, The prerequisite for spectral subtraction in AI intelligent processing to eliminate background noise is: Noise is a slow variable; Quickly distinguish between noise and signal: Using AI training, the spectral characteristics of human voice are obtained in advance to generate an AI model library. Each 10ms data is used as a time slice. The spectral characteristics of each time slice are compared with the human voice characteristics in the model library. If they do not meet the requirements, they are identified as noise and automatically eliminated, and the noise target is dynamically replaced. If the human voice characteristics are met, the noise target is subtracted from the signal to remove the background noise. The obtained continuous speech is divided into time slices, with each slice being 10ms in unit. The data in each time slice is then subjected to Fourier transform processing to convert it into a frequency signal.

6. The AI ​​intelligent noise reduction method for laser-modulated speech according to claim 5, characterized in that, The specific operations of Fourier transform are as follows: Finite-length truncation: Truncate the sampled signal to 10ms, that is, 100 sampling points as a transformation unit, with no overlap between adjacent time slices; Windowing: To reduce spectral leakage, a Hamming window is used to smooth signal edges and reduce sidelobe levels; Zero padding: The sampling sequence is padded with zeros to 128 points to increase the density of the spectrum display, thereby increasing the number of FFT output points and making the curve smoother, without improving the true resolution; Call the FFT algorithm to perform FFT calculation; Amplitude normalization: to obtain the true amplitude, multiply the single-sided spectrum amplitude by 2 / 100, and multiply the DC component by only 1 / 100.

7. The AI ​​intelligent noise reduction method for laser-modulated speech according to claim 1, characterized in that, The specific process of eliminating background noise using spectral subtraction is as follows: MFCC extracts the spectral envelope of speech by simulating the characteristics of human hearing. Noise in MFCC coefficients is characterized by low energy, high entropy, and irregular distribution. In the speech activity detection algorithm, speech frames are distinguished from non-speech frames by using a speech threshold in the AI ​​chip; Initialize by setting the noise target signal to 0; Obtain the spectral information of the time slice and calculate the statistical characteristics of the Mel frequency cepstral coefficients; The mean, variance, and entropy of the Mel frequency cepstral coefficients are obtained and compared with the threshold values ​​of the mean, variance, and entropy of the Mel frequency cepstral coefficients stored in the AI ​​chip. If two or more conditions are met, it is determined to be a signal; otherwise, it is determined to be noise. If it is noise, replace the original noise mark signal with the time slice signal and output 0; if it is a speech signal, subtract the noise mark signal from the signal and output the result.

8. The AI ​​intelligent noise reduction method for laser-modulated speech according to claim 7, characterized in that, The specific process for obtaining the mean, variance, and entropy of the Mel frequency cepstral coefficients is as follows: Obtain the Mel-frequency cepstral coefficients of human voice; First-order high-pass filtering compensates for the attenuation of high-frequency resonant peaks in human voices and improves the signal-to-noise ratio; The signal is cut into short frames of 20–30 ms, with each frame overlapping by 50%, frame length: 25 ms, frame shift: 10 ms, so that the non-stationary speech is approximated as a stationary signal; Perform FFT on each frame to calculate the power spectrum, and pass it through 20–40 triangular Mel filters covering 0–8kHz to simulate the nonlinear frequency perception of the human ear, compress high-frequency redundancy, and highlight key frequency bands of speech. Take the logarithmic energy and perform a discrete cosine transform, retain the first 12 coefficients C1–C12, compress the dynamic range, remove frequency domain correlation, and extract the spectral envelope; The input to the Discrete Cosine Transform is the logarithmic energy of the output of the 24–40 Vimel filter bank, and the output consists of 12 coefficients C1–C12, with C0 retained separately. Through deep learning, the mean, variance, and entropy thresholds of the Mel frequency cepstral coefficients that are closest to human voice are obtained.

9. The AI-based intelligent noise reduction method for laser-modulated speech according to claim 1, characterized in that, The analog-to-digital converter module is model ADS8864, and the digital-to-analog converter module is model DAC8563.

10. The AI-powered intelligent noise reduction method for laser-modulated speech according to claim 1, characterized in that, During the analog-to-digital conversion process, the sampling rate of the analog-to-digital conversion module is set to 10KSamples / S, and the effective bandwidth during the digital filtering process is set to 50Hz~4KHz.

Images