Game voice noise reduction and enhancement method based on AI voiceprint analysis

Abstract

The invention relates to the field of audio recognition and processing, in particular to a game voice noise reduction and enhancement method based on AI voiceprint analysis, which comprises the following steps: firstly, decoupling audio data into two groups of orthogonal feature streams, and carrying out parallel calculation on spectrum features for neural network input and trigger features for state judgment; and then, switching between a steady-state self-adaption mode and a transient locking mode according to the spectral flux change rate and the envelope saturation by utilizing a finite-state machine containing asymmetric hysteresis logic. In the steady-state mode, the system updates the dynamic background noise model in real time to execute denoising standardization; and in the transient mode, freezing statistic update and inhibiting high-frequency clipping distortion through a nonlinear gain function. And finally, coding the state flag into a priori gating signal, and injecting the priori gating signal into a neural network attention layer. According to the method, pollution of burst high-energy signals to a background model is effectively prevented, mistaken truncation of long-tail reverberation is avoided, and speech enhancement robustness and response speed in a complex acoustic environment are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of audio recognition and processing, specifically to a method for game voice noise reduction and enhancement based on AI voiceprint analysis. Background Technology

[0002] With the popularization of e-sports and online games, real-time voice interaction has become a core component of the gaming experience. Modern intelligent multimedia terminals are typically equipped with a pickup array consisting of multiple high signal-to-noise ratio (SNR) microelectromechanical system (MEMS) microphones, connected to an audio codec via a pulse density modulation (PDM) bus, and using a digital signal processor (DSP) and neural network acceleration unit (NNCU) to process the audio data. To achieve high-quality calls with limited transmission bandwidth, current mainstream technologies typically employ deep learning-based voice enhancement algorithms. However, in intense gaming scenarios, this conventional processing method based on traditional feature extraction and general neural networks faces significant challenges.

[0003] Firstly, game scenarios have extremely high dynamic range and suddenness. Players' loud shouts when emotionally agitated often cause physical overload on the microphone or produce physical abrupt changes in non-speech characteristics, such as popping or wind noise. This makes it easy for the system to misjudge the physical overload distortion of the microphone itself as valid speech features, or when using simple linear thresholding, the signal fluctuations around the threshold cause jitter in the state judgment, making it impossible to stably track the signal state.

[0004] Secondly, when the audio signal experiences amplitude saturation, severe nonlinear clipping distortion occurs, manifesting as spurious high-frequency harmonic components in the frequency domain. Traditional linear smoothing or general neural network models often struggle to distinguish these high-frequency artifacts caused by hardware distortion from genuine high-frequency speech details, lacking robust suppression techniques for such nonlinear distortions, thus affecting the final speech reconstruction quality and intelligibility. Summary of the Invention

[0005] To address the issue that traditional linear smoothing or general neural network models negatively impact the final speech reconstruction quality and intelligibility, this invention proposes a game speech denoising and enhancement method based on AI voiceprint analysis. The method includes: acquiring audio data and extracting its frequency domain features as a first feature stream; inputting the first feature stream into an AI voiceprint model; calculating the spectral flux, which includes the difference between the current frame's spectral amplitude and the previous frame's spectral amplitude, and the envelope saturation, which includes the percentage of samples within the current frame whose absolute values ​​reach a set maximum value; the spectral flux and envelope saturation together constitute a second feature stream; and based on the second feature stream, controlling a finite element that switches between a steady-state adaptive mode and a transient locked mode. The state machine switches from a steady-state adaptive mode to a transient locking mode in response to the spectral flux being greater than a high threshold set based on its historical average or the envelope saturation being greater than a set threshold; it switches back from the transient locking mode to a steady-state adaptive mode in response to the spectral flux being lower than a low threshold set based on the historical average when switching to the transient locking mode and remaining below it for a preset duration; when in steady-state adaptive mode, the first feature stream is normalized based on a background noise model; when in transient locking mode, a mode flag is input as a priori gating signal to the decoder contained in the AI ​​voiceprint model, and a nonlinear gain function is applied to the first feature stream to suppress high-frequency distortion.

[0006] This invention addresses the problem of traditional noise reduction algorithms failing to handle both stable background noise and sudden, high-dynamic sounds, such as shouts and explosions, in game scenarios. By introducing spectral flux and envelope saturation as a second feature stream and combining them with a finite state machine mechanism, the system can intelligently distinguish between steady-state adaptive and transient locking scenarios. In steady-state conditions, a background noise model is used for standardization to improve the signal-to-noise ratio. In transient modes where sudden high-energy signals are detected, the AI ​​model's decoding strategy is adjusted using a priori gating signals, and a nonlinear gain function is applied to effectively suppress high-frequency distortion at high volumes. Thus, while ensuring effective background noise reduction, lossless enhancement and protection of high-dynamic voice in games are achieved.

[0007] Furthermore, the method for calculating the historical mean includes: during the initial calibration period after system startup, calculating the arithmetic mean of the spectral flux of a set number of frames as the initial historical mean; when the finite state machine is in steady-state adaptive mode, updating the historical mean using an exponentially weighted moving average algorithm; and pausing updates when in transient locked mode.

[0008] Furthermore, the high-level threshold is the historical average. times, of which The value range is [2.5, 4.0]; the low-order threshold is the historical average value at the time of locking. times, of which The value range is [1.2, 1.5].

[0009] Furthermore, the standardization process based on the background noise model includes: independently maintaining a global mean vector and a variance vector for each frequency point of the first feature stream; updating the global mean vector and variance vector in real time using a recursive smoothing algorithm; and processing the features of the current frame based on the updated global mean vector and variance vector.

[0010] Furthermore, when in transient locking mode, the current frame data is prevented from participating in the update of the global mean vector by pausing the update command.

[0011] Furthermore, the nonlinear gain function is specifically as follows: ; in This refers to the corrected frequency band energy. For input frequency band energy; The frequency band energy threshold based on the background noise model is equal to the last updated global mean. The inhibitory factor has a value range of [1.0, 5.0]. The nonlinear exponent is set to a value of 2; This is a parameter tuning factor used to avoid situations where the denominator is extremely zero.

[0012] This invention utilizes suppression factors and nonlinear exponents, enabling the system to dynamically adjust the gain of high-energy signals based on the relationship between the current frequency band energy and the background noise threshold. This processing method can gently suppress high-frequency harmonic distortion exceeding the dynamic range while preserving the main energy of the speech, avoiding the harshness caused by simple hard limiting and significantly improving speech clarity in intense combat scenarios.

[0013] Furthermore, the second feature stream also includes the logarithm of the linear predictive coding residual energy of the current frame; in response to the logarithm of the residual energy being greater than the sum of its short-time mean and standard deviation multiplied by a set coefficient, the finite state machine is switched to transient locking mode.

[0014] This invention, by adding the logarithm of the linear predictive coding residual energy to the second feature stream, leverages the sensitivity of the LPC residual to vocal tract excitation sources, enabling the system to more quickly identify vocal cord vibrations or sudden speech signals. As a supplement to spectral flux judgment, it improves the accuracy of separating sudden speech from background noise and reduces the risk of initial syllables being misjudged as noise and removed.

[0015] Furthermore, the set coefficient ranges from 3.0 to 5.0.

[0016] Furthermore, inputting the pattern flag as a priori gating signal into the decoder contained in the AI ​​voiceprint model includes: broadcasting the scalar value of the pattern flag to expand it into a gating tensor with the same dimension as a specific intermediate feature tensor in the attention layer of the decoder; and performing element-wise multiplication of the gating tensor with the intermediate feature tensor to adjust the attention weights.

[0017] This invention expands the pattern flag broadcast into a gated tensor with the same dimension as the intermediate feature tensor and performs element-wise multiplication, forcing the AI ​​model's decoder to adjust its attention weights in transient mode. This allows the AI ​​model to dynamically change the focus of feature extraction based on the current acoustic state, making it more adaptable to the drastically changing acoustic environment in game voice than a single-weight model.

[0018] Furthermore, the method operates in a terminal system comprising the following components: a SoC based on a multi-core heterogeneous architecture, including at least one DSP and NPU; a pickup array composed of multiple MEMS microphones connected to the SoC; and an on-chip SRAM connected to the internal processor of the SoC for storing data containing the first feature stream.

[0019] This invention utilizes a SoC based on a multi-core heterogeneous architecture and on-chip SRAM direct access, fully leveraging the computing power and high-speed cache advantages of dedicated hardware accelerators. This ensures that when processing high-sampling-rate game voice and complex AI model inference, it can meet the low-latency requirements of real-time communication and avoids excessive consumption of general-purpose CPU resources that could lead to frame drops in the game.

[0020] The technical effects of this invention are as follows: This invention combines frequency domain features with feature flows such as spectral flux and envelope saturation, utilizing a finite state machine to intelligently switch between steady-state adaptive and transient locking modes. In steady-state mode, it uses a background noise model for precise normalization; in transient mode, it uses gating signals to regulate AI decoding attention and combines nonlinear functions to suppress high-frequency distortion. This solution effectively solves the problem of sound distortion and noise reduction failure caused by sudden loud volumes in game scenarios, achieving a balance between background noise reduction and dynamic voice fidelity. Attached Figure Description

[0021] Figure 1 This is a schematic flowchart illustrating the game voice noise reduction and enhancement method based on AI voiceprint analysis in an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the time-domain waveform comparison between the noisy original signal and the clean speech signal in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the relationship between the audio frame envelope saturation value and the judgment threshold in an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the curves of spectral flux change rate, dynamic reference value, and adaptive high and low thresholds over time in an embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the timeline comparing the state determination of the improved method and the traditional fixed threshold method in an embodiment of the present invention. Figure 6 This is a schematic diagram illustrating the energy comparison curves before and after feature value processing within the burst distortion interval in an embodiment of the present invention. Detailed Implementation

[0022] 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, not all, of the embodiments of the present invention. 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.

[0023] A preferred embodiment of the present invention operates in a highly integrated intelligent multimedia terminal system. The core processing unit of the system adopts an embedded system-on-a-chip (SoC) based on a multi-core heterogeneous architecture, including at least one digital signal processor (DSP) with an operating frequency of not less than 1 GHz and a neural network acceleration unit (NPU) with a computing power of not less than 2 TOPS.

[0024] At the physical level, the system is configured with... A pickup array consisting of high signal-to-noise ratio microelectromechanical systems (MEMS) microphones. The microphone array is connected to an audio codec via a pulse density modulation (PDM) bus, and the audio data stream is transferred to the processor's on-chip static random access memory (SRAM) via a direct memory access (DMA) channel.

[0025] To ensure the continuity and real-time nature of the data stream, the processor allocates and maintains a memory space with a capacity of [missing information]. The frame uses a first-in-first-out (FIFO) ring buffer. This buffer employs a two-pointer mechanism for reading and writing. When the write pointer catches up with the read pointer, the system prioritizes the read operation through an interrupt priority control strategy to ensure that the response latency is less than 10ms when handling sudden high-energy signals.

[0026] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0027] Example of a game voice noise reduction and enhancement method based on AI voiceprint analysis: like Figure 1 As shown, the game voice noise reduction and enhancement method based on AI voiceprint analysis of the present invention includes: S100: The preprocessed audio data is decoupled into an orthogonal feature stream that retains the spectral envelope information using a parallel computing channel, and the envelope saturation and spectral flux change rate are calculated simultaneously to quantify the physical state of the signal.

[0028] The system first preprocesses the acquired audio time-domain signal, including a pre-emphasis coefficient of 0.97 and frame-by-frame windowing, specifically with a frame length of 25ms and a frame shift of 10ms. For example... Figure 2 As shown, the diagram schematically illustrates the time-domain waveform comparison between the raw audio signal captured in a game scene and the ideal clean speech signal. The curve with drastic amplitude fluctuations and obvious spikes represents the raw signal containing sudden physical overloads and environmental noise, exhibiting extremely high dynamic range during certain time periods; while the curve with relatively stable amplitude represents the ideal reference speech signal. Subsequently, the processor constructs parallel computing channels to decouple the data stream into two sets of orthogonal features: In the first channel, the processor maps the time-domain signal to the frequency domain using a discrete Fourier transform, and calculates the Mel-frequency cepstral coefficients or logarithmic filter bank characteristics, i.e., the first feature stream. This feature stream preserves complete spectral envelope information and is directly used as the input vector for subsequent neural network models.

[0029] In the second channel, the processor synchronously calculates the trigger feature stream, also known as the second feature stream, used to quantify the signal energy state. In this embodiment, the trigger feature stream includes indicators in two dimensions: Envelope saturation: The processor iterates through all sample points in the current frame and counts the number of sample points whose absolute value reaches the bit depth limit, such as the number of 32767 sample points in 16-bit audio. Calculate its portion of the frame length ratio .

[0030] like Figure 3 As shown, the system performs statistical analysis on the sampling points within the current frame to calculate the envelope saturation. The prominent pulse waveforms in the figure represent the calculated envelope saturation values ​​for each frame, and the horizontal dashed lines represent the preset saturation judgment threshold. When the calculated saturation value exceeds this threshold, the system identifies that the frame is in a severe physical overload state, and this feature will serve as an important triggering condition for subsequent state machine decisions.

[0031] Spectral flux change rate: The processor calculates the spectral amplitude vector of the current frame. Compared with the spectrum amplitude vector of the previous frame The Euclidean distance between them is given by the following formula: ; in Represents frequency index, This represents the total number of frequency points.

[0032] In another embodiment, to distinguish between "human shouting" and "non-speech abrupt changes" in complex scenarios such as wind noise or microphone physical overload, linear predictive coding (LPC) analysis can be introduced. Specifically, the processor is configured to calculate the LPC prediction residual energy of the current frame signal. have: ; in The original signal; Based on the past Linear prediction value for each sampling point. Processor calculates. The logarithm of the value, and its short-time mean are maintained. with standard deviation The decision logic is as follows: If (in If the setting is preferably 3.0 to 5.0, then a physical abrupt change in the signal that is not a speech characteristic is determined. It utilizes the physical difference that human voice has short-term correlation while physical impact noise has randomness, thereby achieving more accurate anomaly detection at the feature level.

[0033] S200 operates a finite state machine with asymmetric hysteresis characteristics, switching between steady-state adaptive and transient locking modes based on dynamically updated historical average flux reference values ​​and dual trigger thresholds, and preventing state jitter.

[0034] To address the jitter issue in linear threshold determination under non-stationary signals, the processor runs a finite state machine (FSM) with hysteresis characteristics, based on... and Switch between steady-state adaptive mode and transient locking mode.

[0035] S201. During the calibration period, calculate the arithmetic mean and use the exponentially weighted moving average algorithm in steady-state adaptive mode to iteratively update the historical average flux benchmark value to eliminate the contamination of the benchmark value by sudden noise.

[0036] Firstly, the system initially after power-on or reset... Frames (e.g.) A frame is defined as a calibration period. During this period, the state machine is forced into a steady state, and the processor only performs calculations. The arithmetic mean and assigned to The initial state is set to prevent false triggering caused by zero initialization. Then, when the calibration period ends and the state machine is in steady-state adaptive mode, an exponentially weighted moving average algorithm is used for iterative updates, as shown below: ; in The forgetting factor is preferably set to 0.95 to 0.99, and updates are performed only when the state machine is in steady-state adaptive mode to eliminate the contamination of the reference value by sudden noise. This represents the rate of change of spectral flux calculated by S100.

[0037] S202. By comparing the rate of change of spectral flux with the high-level trigger threshold, the transient locking mode is entered, and when it falls back below the low-level recovery threshold, the release timer mechanism is activated to avoid erroneous truncation of the long-tail reverberation.

[0038] steady state Transient determination: The processor will calculate in real time. High-level trigger threshold Compare. Set. ,in The preferred range is [2.5, 4.0]. If... The system immediately sets the status flag and switches to transient lockout mode.

[0039] transient Steady-state determination: The system introduces a release timer mechanism. Only when... Falling back to the low recovery threshold When the following time is reached, the timer starts counting down, as set in this embodiment. ,in , The baseline value is frozen at the time of locking.

[0040] If, within the preferred 300ms countdown window, the feature value of any frame exceeds [a certain threshold] again... The processor immediately performs a timer reset operation, clears the count value to zero, and waits for the signal to stabilize.

[0041] like Figure 4 As shown, the processor tracks changes in frequency domain characteristics in real time. The sharply fluctuating solid line in the figure represents the calculated rate of change of spectral flux; the smooth solid line at the bottom represents the historical average flux baseline value updated iteratively using an exponentially weighted moving average algorithm; the two dashed lines at the top represent the high-level trigger threshold and low-level recovery threshold dynamically generated based on this baseline value, respectively. It can be seen that as the baseline value fluctuates slowly, the high and low thresholds also exhibit an adaptive adjustment trend, rather than remaining fixed.

[0042] The above operation ensures that the state machine only releases the lock after the signal energy has completely decayed and remained stable, thus avoiding the erroneous truncation of the long-tailed reverberation at the physical level.

[0043] like Figure 5As shown, the state determination results of the traditional fixed threshold method and the asymmetric hysteresis logic of the present invention are compared. The square wave exhibiting frequent up-and-down transitions represents the state determination of the traditional method when signal energy fluctuates, showing significant state jitter; while another square wave, which maintains a longer duration and has stable state switching, represents the state flag of the present invention. Particularly during the signal energy decline phase, the state flag of the present invention does not immediately reset but remains locked, demonstrating the protective effect of the release timer mechanism on long-tailed signals.

[0044] S300: Perform mutually exclusive processing logic on the spectral features based on the discrete state flags output by the finite state machine to update the background noise model in steady state and suppress high-frequency distortion in transient state through a nonlinear gain function.

[0045] Based on the discrete state output by S200, the processor performs mutually exclusive processing logic on the Mel-frequency cepstral coefficients or logarithmic filter bank characteristics of the first channel.

[0046] S301. In steady-state mode, enable write access to statistics, use recursive smoothing algorithm to track the energy benchmark at each frequency point in real time, and build a dynamic background noise model to perform denoising and standardization processing to eliminate the influence of environmental background noise.

[0047] The execution logic of the steady-state adaptive mode is as follows: When the state machine is in a steady state, the system determines that the current environment is a reliable background noise range. The processor enables write access to the statistics register and performs dynamic background noise modeling based on fine-grained frequency domain.

[0048] Specifically, the processor independently maintains a set of statistical state variables, namely the global mean vector, for each frequency point of the data stream in the first channel. With variance vector Using a recursive smoothing algorithm, the system tracks and updates the energy reference and fluctuation range of each frequency point in real time under the current acoustic environment. This process is equivalent to filtering out transient disturbances through low-pass filtering, thereby constructing a dynamic background noise model that can adaptively follow the slowly time-varying characteristics of the acoustic channel.

[0049] Subsequently, based on the dynamic background noise model, the processor performs denoising and normalization processing on the current input frame, mapping frequency bands of different energy levels to a standard feature space to eliminate the impact of microphone frequency response differences and environmental background noise on subsequent neural network inference.

[0050] S302. In transient mode, high-energy frames are prohibited from participating in global mean iteration, and the spectrum is modified by a nonlinear gain function driven by the input energy independent variable to achieve hard suppression of nonlinear distortion artifacts.

[0051] The execution logic of transient locking mode is as follows: The processor first pulls the write enable signal low through gating logic, physically preventing the data of the current high-energy frame from participating in the global averaging. The iterative calculation prevents abnormally large amplitude data from raising the global mean, thus avoiding subsequent normal speech frames from being incorrectly normalized to the low amplitude range.

[0052] Then, to address the high-frequency harmonic distortion caused by clipping, the processor applies a nonlinear gain function to correct the spectrum. Let the input spectral energy be... Corrected energy The calculation formula is as follows: ; in For the frequency band energy threshold, directly call the last updated value in Mode A state. value; This is a suppression factor with a value range of [1.0, 5.0], used to adjust the suppression strength of abnormal signals; It is a non-linear exponent, with a preferred value of 2, used to control the attenuation slope of large signals; This is a parameter tuning factor used to avoid situations where the denominator is extremely zero.

[0053] The above formula constructs a system based on the input energy. Independent variable-driven voltage-controlled attenuator. When When the denominator is 1, the gain is 1, and the signal is direct; when Significantly exceeds the steady-state threshold As the denominator increases exponentially, the gain coefficient rapidly approaches zero. Through the above calculations, the system achieves hard suppression of the amplitude of high-frequency artifacts caused by nonlinear distortion, rather than simply performing linear smoothing.

[0054] like Figure 6 As shown, this demonstrates the nonlinear restoration effect on eigenvalues ​​in transient-locked mode. The curve with sharp peaks represents the original logarithmic energy characteristics before processing, especially in the burst distortion range highlighted in the background, where high-frequency artifacts caused by clipping exist. The other curve, with its peaks significantly flattened and smoother, represents the output characteristics after correction by the nonlinear gain function. The comparison clearly shows that this method effectively suppresses anomalously large high-frequency components, achieving hard suppression of nonlinear distortion.

[0055] S400: The state machine pattern flag is encoded as a priori gating signal and extended into a gating tensor through a broadcast operation and injected into the attention layer of the neural network to dynamically adjust the decoder's response weights to features of different frequency bands.

[0056] The tiered feature vectors are then fed into the AI ​​voiceprint recognition model on the NPU. To improve the model's robustness, the system adds a mode flag to the state machine ( The encoding is a priori gating signal.

[0057] Gated injection logic: In order to interact with the dimension of The neural network intermediate layer feature tensors are aligned, and the processor performs a broadcast operation to transfer the scalars. Expanded to a gated tensor of the same dimension Subsequently, The input is fed into the attention layer of the decoder.

[0058] When the transient lock-in mode is indicated, the decoder dynamically adjusts the attention weights based on the prior information to reduce the response value to high-frequency dimensional features; When the steady-state adaptive mode is indicated, the decoder tends to finely remove background steady-state noise.

Claims

1. A method for game voice denoising and enhancement based on AI voiceprint analysis, characterized in that, The method includes: acquiring audio data and extracting its frequency domain features as a first feature stream, and inputting the first feature stream into an AI voiceprint model; Calculate the spectral flux, which includes the difference between the spectral amplitude of the current frame and the spectral amplitude of the previous frame, and the envelope saturation, which includes the percentage of the number of sampling points in the current frame whose absolute values ​​reach a set maximum value. The spectral flux and the envelope saturation together constitute the second feature flow. Based on the second feature flow, control a finite state machine that switches between steady-state adaptive mode and transient locked mode. In response to the spectral flux being greater than a high threshold set based on its historical average or the envelope saturation being greater than a set threshold, the system switches from steady-state adaptive mode to transient locking mode; in response to the spectral flux being lower than a low threshold set based on the historical average when switching to the transient locking mode and continuing for a preset duration, the system switches back from transient locking mode to steady-state adaptive mode. When in steady-state adaptive mode, the first feature stream is normalized based on the background noise model; when in transient locked mode, the mode flag is used as a priori gating signal input to the decoder contained in the AI ​​voiceprint model, and a nonlinear gain function is applied to the first feature stream to suppress high-frequency distortion.

2. The game voice noise reduction and enhancement method based on AI voiceprint analysis according to claim 1, characterized in that, The method for calculating the historical mean includes: During the initial calibration period after system startup, the arithmetic mean of the spectral flux of a set number of frames is calculated as the initial historical mean. When the finite state machine is in steady-state adaptive mode, the historical mean is updated using an exponentially weighted moving average algorithm; when it is in transient locked mode, the update is paused.

3. The game voice noise reduction and enhancement method based on AI voiceprint analysis according to claim 1, characterized in that, The high threshold is the historical average. times, of which The value range is [2.5, 4.0]; The low-level threshold is the historical average value at the time of locking. times, of which The value range is [1.2, 1.5].

4. The game voice noise reduction and enhancement method based on AI voiceprint analysis according to claim 1, characterized in that, The standardization process based on the background noise model includes: The global mean vector and variance vector are maintained independently for each frequency point of the first feature stream; The global mean vector and variance vector are updated in real time using a recursive smoothing algorithm. The features of the current frame are processed based on the updated global mean vector and variance vector.

5. The game voice noise reduction and enhancement method based on AI voiceprint analysis according to claim 4, characterized in that, When in transient locking mode, the current frame data is prevented from participating in the update of the global mean vector by pausing the update command.

6. The game voice noise reduction and enhancement method based on AI voiceprint analysis according to claim 1, characterized in that, The nonlinear gain function is specifically: ； in This refers to the corrected frequency band energy. For input frequency band energy; The frequency band energy threshold based on the background noise model is equal to the last updated global mean. The inhibitory factor has a value range of [1.0, 5.0]. The nonlinear exponent is set to a value of 2; This is a parameter tuning factor used to avoid situations where the denominator is extremely zero.

7. The game voice noise reduction and enhancement method based on AI voiceprint analysis according to claim 1, characterized in that, The second feature stream also includes the logarithm of the linear prediction coding residual energy of the current frame; In response to the logarithm of the residual energy being greater than the sum of its short-time mean and standard deviation multiplied by a set coefficient, the finite state machine is switched to transient locked mode.

8. The game voice noise reduction and enhancement method based on AI voiceprint analysis according to claim 7, characterized in that, The set coefficient ranges from 3.0 to 5.

0.

9. The game voice noise reduction and enhancement method based on AI voiceprint analysis according to claim 1, characterized in that, The pattern flag is input as a priori gating signal into the decoder included in the AI ​​voiceprint model, including: The scalar value of the pattern flag is broadcast to expand it into a gated tensor with the same dimension as the specific intermediate feature tensor in the attention layer of the decoder; The gating tensor is element-wise multiplied with the intermediate feature tensor to adjust the attention weights.

10. The game voice noise reduction and enhancement method based on AI voiceprint analysis according to claim 1, characterized in that, The method operates on a terminal system that includes the following components: A SoC based on a multi-core heterogeneous architecture includes at least one DSP and NPU; A pickup array consisting of multiple MEMS microphones connected to the SoC; An on-chip SRAM connected to the internal processor of the SoC is used to store data containing the first feature stream.

Images