Vehicle-mounted intelligent sound field management method and system and computer readable storage medium

Abstract

The invention discloses an active noise reduction method and system for a vehicle-mounted intelligent sound field and a computer readable storage medium. The method comprises the following steps: acquiring an external acoustic signal outside a vehicle; generating a reverse noise reduction wave based on the external acoustic signal, wherein the reverse noise reduction wave is used for counteracting the external acoustic signal; identifying a specific acoustic event in the external acoustic signal and a specific acoustic feature corresponding to the specific acoustic event; reconstructing reconstructed time domain waves with consistent features based on the specific acoustic events and the corresponding specific acoustic features; and playing the reconstructed time domain wave and the reverse noise reduction wave in the vehicle cabin through a loudspeaker. According to the management method of the vehicle-mounted intelligent sound field, the key external safe sound can be recognized and actively reconstructed while the background noise is continuously and actively reduced, and the key external safe sound is integrated into the cabin sound field, so that the extreme quiet and comfortable experience is provided on the premise that the safety is not sacrificed.

Description

Technical Field

[0001] This invention relates to the automotive field, and further to a method, system, and computer-readable storage medium for managing an in-vehicle intelligent sound field. Background Technology

[0002] Vehicle active and passive noise reduction technology is a combined noise reduction solution of "active cancellation + passive blocking," with the core being to comprehensively reduce in-vehicle noise and improve the quietness of the ride. "Active cancellation" uses technology to cancel in-vehicle noise and improve the quietness of the ride. The main principle is to use microphones to collect in-vehicle noise (such as engine noise, wind noise, tire noise) in real time, and the onboard processor quickly calculates and generates "anti-noise waves" with opposite phase and the same amplitude as the noise. The speakers release the anti-noise waves, which cancel out the original noise, thus achieving the noise reduction effect. "Passive blocking" passively blocks the transmission of noise, mainly covering mid-to-high frequency noise, and is achieved through physical structures and materials.

[0003] In recent years, with the widespread adoption of active and passive noise reduction technologies in vehicles, the quietness of vehicle cabins has significantly improved, especially in pure electric vehicles. However, a new problem has arisen: excessive quietness can isolate the driver from the external acoustic environment, creating an "auditory vacuum" or a "sense of isolation," which may pose significant safety hazards. Specifically:

[0004] 1) Loss of critical safety information. Important external sound signals, such as sirens of emergency vehicles, horns of other vehicles, shouts of pedestrians, and abnormal noises in the traffic environment (such as tire bursting), are indiscriminately canceled or weakened by the noise reduction system, making it impossible for drivers to effectively obtain this critical auditory safety information.

[0005] 2) Unnatural auditory experience. Currently, the mainstream direction of active cabin noise cancellation technology is to achieve maximum quietness. Being in a completely quiet environment for a long time may cause some passengers to experience ear pressure discomfort or motion sickness. The lack of natural environmental references does not conform to human long-term auditory habits.

[0006] For example, Chinese invention patent application number CN202310592558.2 discloses an in-vehicle multi-channel active noise cancellation system and method for improving sound quality. The system includes: an audio transceiver module, an audio data transmission module, a noise signal analysis and preprocessing module, a reference signal generation module, a FELMS-based multi-channel noise cancellation algorithm module, and a basic parameter setting module. The noise signal analysis and preprocessing module performs noise signal analysis and preprocessing on the noise source signal, then sends the reference signal required for control to the noise cancellation algorithm module. After combining the received error signal with the algorithm module's calculations, it emits an anti-noise signal to achieve noise cancellation. The technical solution disclosed in this invention patent can analyze the sound quality characteristics of in-vehicle noise and control it accordingly, improving in-vehicle sound quality while meeting the requirements for noise reduction in a large in-vehicle space, thereby enhancing the driving experience. However, this invention does not involve the active identification and enhancement of critical safety information, and it does not consider the safety issues caused by excessive noise reduction while maintaining noise reduction effectiveness and in-vehicle sound quality. Summary of the Invention

[0007] To address the aforementioned technical problems, the present invention aims to provide a method, system, and computer-readable medium for managing an in-vehicle intelligent sound field. The method for managing an in-vehicle intelligent sound field can identify and actively reconstruct key external safety sounds while continuously performing active noise reduction of background noise, and integrate them into the cabin sound field, thereby providing an ultimate quiet and comfortable experience without sacrificing safety.

[0008] To achieve the above objectives, the present invention aims to provide an active noise reduction method for in-vehicle intelligent sound fields, comprising:

[0009] Acquire external acoustic signals from outside the vehicle;

[0010] An inverse noise reduction wave is generated based on the external acoustic signal, and the inverse noise reduction wave is used to cancel out the external acoustic signal;

[0011] Identify specific acoustic events in the external acoustic signal and the specific acoustic features corresponding to those specific acoustic events;

[0012] Reconstruct a reconstructed time-domain wave with consistent characteristics based on the specific acoustic event and the corresponding specific acoustic features;

[0013] The reconstructed time-domain wave and the inverse noise-reduced wave are played in the vehicle cabin through a speaker.

[0014] In some implementations, the step of playing the reconstructed time-domain wave and the inverse noise-reduced wave in the vehicle cabin via a speaker is as follows:

[0015] The reverse noise-canceling wave and the reconstructed time-domain wave are played using the same loudspeaker. The driving signal of the same loudspeaker = reverse noise-canceling wave + α × reconstructed time-domain wave, where α is a gain coefficient associated with the specific acoustic event, which includes sirens, horns, ambulance sounds, train crossing sounds, and tire explosion sounds. The specific acoustic features include frequency, intensity, direction, and distance.

[0016] In some implementations, identifying specific acoustic events in the external acoustic signal and specific acoustic features corresponding to those events includes analyzing the external acoustic signal based on a convolutional neural network model, including:

[0017] The external acoustic signal is input into a trained convolutional neural network model to identify specific acoustic events in the external acoustic signal and the specific acoustic features corresponding to the specific acoustic events;

[0018] The training process of the convolutional neural network model is as follows:

[0019] An audio training set is formed by acquiring a collection of audio signals and noise audio with specific acoustic events;

[0020] Convert the audio signals in the audio training set into a spectrogram atlas;

[0021] The spectrogram set is input into a convolutional neural network model to train the model. The convolutional neural network model includes an input layer, a convolutional layer, a pooling layer, an activation function layer, a fully connected layer, and an output layer. The input layer receives the spectrogram as input. The convolutional layer uses multiple convolutional kernels to extract local features from the audio. The activation function introduces non-linearity. The fully connected layer maps the features extracted by the convolutional layer to the output dimension required for the classification task. The output layer converts the output into a probability distribution representing the likelihood of different sound events.

[0022] In some implementations, identifying specific acoustic events in the external acoustic signal and specific acoustic features corresponding to those events includes analyzing the external acoustic signal based on a recurrent neural network model, including:

[0023] The external acoustic signal is input into a trained recurrent neural network model to identify specific acoustic events in the external acoustic signal and the specific acoustic features corresponding to the specific acoustic events;

[0024] The training process of the recurrent neural network model is as follows:

[0025] An audio training set is formed by acquiring a collection of audio signals and noise audio with specific acoustic events;

[0026] Convert the audio signals in the audio training set into a spectrogram atlas;

[0027] The spectrogram set is input into a recurrent neural network model to train the model. The recurrent neural network model includes an input layer, several hidden layers, a fully connected layer, and an output layer. The hidden layers are used to extract temporal features. The input layer is used to receive the spectrogram as input. The fully connected layer establishes the association between the temporal features and the sound time category. The output layer is used to convert the output into a probability distribution to represent the probability of different sound events.

[0028] In some implementations, the step of generating a reconstructed acoustic signal based on the specific acoustic event and the corresponding specific acoustic features includes:

[0029] Determine the corresponding basic waveform template based on the specific acoustic event;

[0030] The specific acoustic features are injected into the base waveform template, and the waveform of the base waveform template is adjusted to generate the reconstructed acoustic waveform;

[0031] The reconstructed acoustic waveform is modulated based on a modulation function to generate a time-domain waveform corresponding to the specific acoustic event.

[0032] In some embodiments, the step of generating a reverse noise-canceling wave based on the external acoustic signal, the reverse noise-canceling wave being used to cancel out the external acoustic signal, includes:

[0033] The external acoustic signal is converted into frequency domain features to obtain the frequency distribution, amplitude, and phase of the external acoustic signal;

[0034] The phase of each frequency component is reversed by 180° to generate an inverse frequency domain signal, and the amplitude of the inverse frequency domain signal is adjusted based on the amplitude of the external acoustic signal.

[0035] The inverse frequency domain signal is converted into an inverse time domain signal to generate the inverse noise-reduced wave.

[0036] According to another aspect of this application, an active noise cancellation system for an in-vehicle smart speaker is further provided, comprising:

[0037] An acoustic signal acquisition module, comprising an external acquisition unit for acquiring external acoustic signals from outside the vehicle;

[0038] A noise reduction module is provided, wherein the noise reduction module is used to generate a reverse noise reduction wave based on the external acoustic signal, and the reverse noise reduction wave is used to cancel out the external acoustic signal;

[0039] An acoustic reconstruction module includes a feature extraction unit and a time-domain reconstruction unit. The feature extraction unit is used to identify specific acoustic events in the external acoustic signal and specific acoustic features corresponding to the specific acoustic events. The time-domain reconstruction unit is used to reconstruct a reconstructed time-domain wave with consistent characteristics based on the specific acoustic events and corresponding specific acoustic features.

[0040] An acoustic signal playback module, which is connected to the noise reduction module and the acoustic reconstruction module, is used to convert the inverse noise reduction wave and the reconstructed time-domain wave into sound signals.

[0041] In some embodiments, the acoustic signal playback module is used to play the reverse noise-canceling wave and the reconstructed time-domain wave using the same speaker, wherein the drive signal of the same speaker = reverse noise-canceling wave + α × reconstructed time-domain wave, where α is a gain coefficient associated with the specific acoustic event, including sirens, horns, ambulance sounds, train crossing sounds, and tire explosion sounds, and the specific acoustic features include frequency, intensity, direction, and distance.

[0042] In some implementations, the feature extraction unit is adapted to identify specific acoustic events in the external acoustic signal and specific acoustic features corresponding to the specific acoustic events using a convolutional neural network model or a recurrent neural network model.

[0043] According to another aspect of this application, a computer-readable storage medium is further provided, storing a computer program that can be loaded by a processor and execute the above-described active noise reduction method for in-vehicle intelligent sound field.

[0044] Beneficial effects:

[0045] 1. The vehicle-mounted intelligent sound field management method provided in this application can identify and actively reconstruct key external safety sounds while continuously performing active noise reduction of background noise, and integrate them into the cabin sound field, thereby providing an ultimate quiet and comfortable experience without sacrificing safety.

[0046] 2. The vehicle-mounted smart speaker management method provided in this application reconstructs a reconstructed time-domain wave with consistent characteristics after acquiring a specific acoustic event and corresponding specific acoustic features, and plays it in the vehicle cabin space, so that the people in the vehicle cabin can know the sound of a specific acoustic event outside the vehicle.

[0047] 3. The vehicle-mounted smart speaker management method provided in this application fuses the inverse noise-canceling wave and the reconstructed time-domain wave and plays the inverse noise-canceling wave and the reconstructed time-domain wave using the same speaker. The output wave can be changed by adjusting the gain coefficient based on different specific acoustic events, and the cost can be reduced by playing it through the same speaker. Attached Figure Description

[0048] The preferred embodiments will now be described in a clear and easy-to-understand manner, in conjunction with the accompanying drawings, to further explain the above-mentioned characteristics, technical features, advantages, and implementation methods of the present invention.

[0049] Figure 1 This is a flowchart illustrating the active noise cancellation method for an in-vehicle smart speaker according to a preferred embodiment of the present invention.

[0050] Figure 2 This is a schematic diagram of the process for generating reverse noise reduction waves in the active noise reduction method of an in-vehicle smart speaker according to a preferred embodiment of the present invention.

[0051] Figure 3 This is a schematic diagram of the process for generating reconstructed time-domain waves in the active noise reduction method of an in-vehicle smart speaker according to a preferred embodiment of the present invention;

[0052] Figure 4 This is a block diagram of the active noise cancellation system of an in-vehicle smart speaker according to a preferred embodiment of the present invention. Detailed Implementation

[0053] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the specific implementation methods of the present invention will be described below with reference to the accompanying drawings. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings and other implementation methods can be obtained based on these drawings without any creative effort.

[0054] To keep the drawings concise, each figure only schematically shows the parts relevant to the invention, and these do not represent the actual structure of the product. Furthermore, to facilitate understanding, in some figures, only one of components with the same structure or function is schematically depicted, or only one is labeled. In this document, "one" not only means "only one," but can also mean "more than one."

[0055] It should also be further understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0056] In this document, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0057] Furthermore, in the description of this application, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0058] refer to Figures 1 to 3 This application provides an active noise cancellation method 100 for an in-vehicle smart speaker, the active noise cancellation method 100 for an in-vehicle smart speaker comprising:

[0059] S101: Acquire external acoustic signals from outside the vehicle;

[0060] S102: Generate an inverse noise reduction wave based on the external acoustic signal, the inverse noise reduction wave being used to cancel out the external acoustic signal;

[0061] S103: Identify specific acoustic events in the external acoustic signal and the specific acoustic features corresponding to the specific acoustic events;

[0062] S104: Reconstruct a reconstructed time-domain wave with consistent characteristics based on the specific acoustic event and the corresponding specific acoustic features;

[0063] S105: Play the reconstructed time-domain wave and the inverse noise-reduced wave in the vehicle cabin through a speaker.

[0064] In step S101 above, external acoustic signals are obtained by acquiring audio from outside the vehicle through an external acquisition unit located outside the vehicle. The external acquisition unit includes multiple microphones located outside the vehicle. For example, but not limited to, the microphones of the external acquisition unit may be microphones located next to the front-facing ADAS camera of the vehicle, microphones located at the base of the rearview mirror, etc. In some embodiments, the microphones of the external acquisition unit may also be additional microphones located on the vehicle body, such as on the top surface or side surface of the vehicle. It is understood that the number and location of the microphones included in the external acquisition unit should not constitute a limitation of this application.

[0065] Preferably, after acquiring the external acoustic signal from outside the vehicle, the external acoustic signal is subjected to noise reduction and beamforming processing to enhance the target sound source signal. Specifically, noise reduction can employ methods such as spectral subtraction, Wiener filtering, or adaptive noise cancellation; beamforming processing can employ delay-summation beamforming, minimum variance distortionless response, or adaptive beamforming.

[0066] In step S102 above, a reverse noise reduction wave is generated based on the external acoustic signal. This reverse noise reduction wave can cancel out the external acoustic signal, thereby achieving active noise reduction. Specifically, the external acoustic signal includes engine noise, tire noise, and wind noise inside the vehicle.

[0067] refer to Figure 2 Specifically, step S102 above further includes:

[0068] S1021: Convert the external acoustic signal into frequency domain features to obtain the frequency distribution, amplitude, and phase of the external acoustic signal;

[0069] S1022: The phase of each frequency component is reversed by 180° to generate an inverse frequency domain signal, and the amplitude of the inverse frequency domain signal is adjusted based on the amplitude of the internal acoustic signal.

[0070] S1023: Convert the inverse frequency domain signal into an inverse time domain signal to generate the inverse noise-reduced wave.

[0071] In step S1021 above, the microphone of the external acquisition unit converts the sound wave signal outside the vehicle cabin into an analog electrical signal, and then converts the analog electrical signal into a digital signal through an analog-to-digital converter. During this process, irrelevant signals such as microphone background noise and circuit interference can be filtered out simultaneously. After acquiring the internal acoustic signal in digital form, a fast Fourier transform is used to convert the internal acoustic signal from a time-domain signal to a frequency-domain signal to decompose the key parameters of each frequency component, including frequency value, amplitude, and phase. In some embodiments, after acquiring the frequency-domain signal, the frequency-domain signal is classified and filtered to distinguish between periodic noise (such as engine roaring sound 50-100Hz) and non-periodic noise (such as instantaneous noise such as road bumps), focusing on low-frequency and mid-frequency components (such as 20-500Hz) that can be canceled by active noise cancellation.

[0072] In step 1022 above, the phase of each frequency component is reversed, preferably by 180°, so that the reversed sound wave will undergo "destructive interference" when it meets the original sound wave. After reversing the phase of each frequency component, the amplitude of the reversed signal is adjusted based on the amplitude of the original signal to ensure that the amplitude of the inverted signal is completely consistent with that of the original signal. That is, the amplitude of the reversed sound wave signal is the same as that of the original sound wave signal, the only difference being that the phases are opposite by 180°.

[0073] In step 1023 above, the inverse frequency domain signal generated in step S1022 is converted into an inverse time domain signal by inverse Fourier transform (IFT) or inverse fast Fourier transform (IFFT), which is to generate the final inverse noise-reduced wave.

[0074] Specifically, the external acoustic signal is the original noise, which is denoted as... Through analog-to-digital converter Reference signal converted to digital signal form .

[0075] Furthermore, an internal acquisition unit (error microphone) is also installed inside the vehicle's cabin to acquire internal acoustic signals (residual noise) within the cabin. The internal acquisition unit includes multiple microphones located within the vehicle's cabin; for example, but not limited to, microphones located in the headrests, ceiling, etc., are also present. A speaker for playing the reconstructed time-domain wave and the inverse noise-reduced wave is also located within the vehicle's cabin. The inverse noise-reduced wave played by the speaker travels from its output to the internal acquisition unit via a path of speaker—air—internal acquisition unit. , for Domain symbol represents frequency domain characteristics. The transmission path is established through factory calibration or real-time adaptive methods. The mathematical model (filter function) is denoted as... .use For reference signal Filtering is performed to obtain the filtered reference signal. , , where * represents convolution operation, corresponding to the linear superposition of time-domain signals. Simulated reference signal The signal characteristics after passing through the speaker to the internal acquisition unit ensure that the anti-phase sound wave calculated by the subsequent algorithm can adapt to system distortion.

[0076] Will With the current filter coefficients of the adaptive FIR filter Performing linear convolution yields a preliminary inverse frequency domain signal, denoted as... ,

[0077] ;

[0078] in For filter tap index, Indicates delay The filtered reference signal after the step.

[0079] ,in For length.

[0080] Digital signals are converted to digital signals via a digital-to-analog converter (DAC). Converted into an analog signal and played through a speaker, this forms the inverse noise-reduced wave. .

[0081] In step S103 above, the external acoustic signal is analyzed in real time based on a pre-trained deep learning model to identify specific acoustic events and their corresponding specific acoustic features. The specific acoustic events include sirens, horns, ambulance sounds, railway crossing sounds, and tire explosions, etc., and the specific acoustic features include frequency, intensity, direction, and distance.

[0082] In some implementations, the external acoustic signal is analyzed based on a convolutional neural network model, including:

[0083] The external acoustic signal is input into a trained convolutional neural network model to identify specific acoustic events in the external acoustic signal and the specific acoustic features corresponding to the specific acoustic events;

[0084] The training process of the convolutional neural network model is as follows:

[0085] An audio training set is formed by acquiring a collection of audio signals and noise audio with specific acoustic events;

[0086] Convert the audio signals in the audio training set into a spectrogram atlas;

[0087] The spectrogram set is input into a convolutional neural network model to train the model. The convolutional neural network model includes an input layer, a convolutional layer, a pooling layer, an activation function layer, a fully connected layer, and an output layer. The input layer receives the spectrogram as input. The convolutional layer uses multiple convolutional kernels to extract local features from the audio. The activation function introduces non-linearity. The fully connected layer maps the features extracted by the convolutional layer to the output dimension required for the classification task. The output layer converts the output into a probability distribution representing the likelihood of different sound events.

[0088] Analyzing external acoustic signals using a convolutional neural network (CNN) model requires pre-training the model, which generally includes data acquisition, model building, and model training. In the data acquisition phase, a large number of sounds and noises (wind noise, rain noise, etc.) from specific acoustic events (sirens, horns, ambulance sounds, train crossing sounds, tire explosions, etc.) are collected beforehand, covering scenes at different distances and under varying environmental noise levels as much as possible. The collected audio samples are then labeled with their categories, such as siren = 1, horn = 2, ambulance sound = 3, train crossing sound = 4, tire explosion sound = 5, etc. The acquired audio data is then divided into training, validation, and test sets in a 7:2:1 ratio.

[0089] The acquired audio data is transformed from a one-dimensional time-domain signal into a two-dimensional feature map (spectral atlas) while suppressing noise and enhancing the target sound features. Before converting the time-domain signal into a two-dimensional feature map, preprocessing of the audio data in the spectrogram atlas is included. This includes: uniform sampling rate processing: all audio samples are standardized to a fixed sampling rate, such as 16kHz, to balance feature accuracy and computational power while avoiding feature distortion caused by sampling rate differences; duration standardization processing: excessively long audio clips, such as those exceeding 20 seconds, are trimmed, retaining the core segments containing the target sound; excessively short audio clips are padded, for example, with silence to a fixed duration, such as 2 seconds, to ensure consistent input dimensions; and noise reduction processing, using spectral subtraction, Wiener filtering, and other methods to remove environmental noise and enhance the temporal features of the target sound.

[0090] Mel spectrograms are used to transform the preprocessed one-dimensional time-domain signal into a two-dimensional spectrogram. Specifically, this involves: segmenting the audio into frames to avoid information loss, for example, frame lengths of 20-30 ms and frame shifts of 5-10 ms; performing a Fast Fourier Transform on each frame to obtain the frequency domain amplitude spectrum; mapping the frequency axis to a Mel scale to simulate the human ear's sensitivity to low frequencies and insensitivity to high frequencies; and taking a logarithmic compression of the dynamic range to finally obtain a two-dimensional feature map, for example, 128×128, i.e., 128 Mel frequency points × 128 frames.

[0091] After acquiring the two-dimensional feature map, data augmentation is performed to improve the model's generalization ability and avoid overfitting. Data augmentation generates "virtual samples," allowing the neural network model to learn target features from more scenarios. Examples, but not limited to, include temporal augmentation, such as random cropping to extract audio segments; time stretching (speed ±10%, without changing pitch); and volume scaling (amplitude ±15%). Frequency domain augmentation includes frequency masking (randomly obscuring 1-5 frequency points in the spectrogram); temporal masking (randomly obscuring 1-5 frames); and adding low-amplitude Gaussian noise to simulate slight interference. Scene augmentation involves mixing background noise and superimposing the sound of a specific acoustic event with environmental noise.

[0092] In the model building phase, a convolutional neural network model is constructed, comprising an input layer, convolutional layers, pooling layers, activation function layers, fully connected layers, and an output layer. The input layer serves as the model's data entry point, receiving the standardized two-dimensional spectrum and converting it into a Tensor format that the convolutional neural network model can process. The convolutional layers perform sliding convolution operations on the two-dimensional spectrum using multiple convolutional kernels of different sizes, such as 3×3 and 5×5, to extract local time-frequency features of the sound, such as peak textures and frequency continuity patterns in the spectrogram, corresponding to key information such as harmonics, rhythm, and tone. The activation function layer performs non-linear transformations on the feature maps output by the convolutional layers, enabling the model to learn complex feature mapping relationships. The activation function can be... Activation function, formula is The pooling layer is used to "downsample" the activated feature map, reducing its dimensionality while preserving core features. This reduces the number of parameters and computational cost, and improves the model's robustness to interference; for example, slight shifts in the spectrogram do not affect feature recognition. Max pooling is preferred, taking the maximum value within a local window (e.g., 2×2) as the output, preserving the most significant features in the feature map (corresponding to the strongest time-frequency signal in the sound). In some implementations, convolutional layers, activation function layers, and pooling layers can be combined to form a feature extraction block, which can be stacked multiple times (e.g., 3-5 blocks) to achieve hierarchical feature extraction from simple to complex. The fully connected layer flattens the multidimensional feature map output by the pooling layer into a one-dimensional feature vector, and then maps the high-dimensional features to the output dimension corresponding to the "number of sound time categories." The output layer converts the "raw score" output by the fully connected layer into a probability value between 0 and 1, intuitively representing the model's prediction confidence for each sound event category.

[0093] In some implementations, the external acoustic signal is analyzed based on a recurrent neural network model, including:

[0094] The external acoustic signal is input into a trained recurrent neural network model to identify specific acoustic events in the external acoustic signal and the specific acoustic features corresponding to the specific acoustic events;

[0095] The training process of the recurrent neural network model is as follows:

[0096] An audio training set is formed by acquiring a collection of audio signals and noise audio with specific acoustic events;

[0097] Convert the audio signals in the audio training set into a spectrogram atlas;

[0098] The spectrogram set is input into a recurrent neural network model to train the model. The recurrent neural network model includes an input layer, several hidden layers, a fully connected layer, and an output layer. The hidden layers are used to extract temporal features. The input layer is used to receive the spectrogram as input. The fully connected layer establishes the association between the temporal features and the sound time category. The output layer is used to convert the output into a probability distribution to represent the probability of different sound events.

[0099] The main difference between using a recurrent neural network (RNN) model and the convolutional neural network (CNN) model described above lies in the construction of the RNN model, which will be discussed below. The input layer of the RNN model receives the temporal spectrogram sequence. The hidden layer uses a forget gate to determine how much historical temporal feature to retain. The input gate is used to update the current temporal feature, and the output gate outputs the current feature to capture long-term temporal correlations. The first layer of several hidden layers is set to True, outputting temporal features for the next layer to continue capturing; the second layer is set to False, outputting global temporal features for classification by the fully connected layer. The fully connected layer maps the global temporal features extracted by the hidden layers to the classification dimension, establishing a correlation between temporal features and sound time categories. The output layer uses Softmax activation to transform the original score into a probability distribution.

[0100] In some modified embodiments, it is also possible to identify specific acoustic events in the external acoustic signal and specific acoustic features corresponding to the specific acoustic events by combining convolutional neural network models with recurrent neural network models.

[0101] refer to Figure 3 In step S104 above, reconstructing a reconstructed time-domain wave with consistent characteristics based on the specific acoustic event and the corresponding specific acoustic features further includes:

[0102] S1041: Determine the corresponding basic waveform template based on the specific acoustic event;

[0103] S1042: Inject the specific acoustic features into the basic waveform template, and adjust the waveform of the basic waveform template to generate a reconstructed acoustic waveform;

[0104] S1043: Modulate the reconstructed acoustic waveform based on the modulation function to generate a time-domain waveform corresponding to the specific acoustic event.

[0105] In step S1041 above, the basic waveform template is preferably extracted through classification and reuse. For example, a template library for specific acoustic events can be pre-rewarded, and the corresponding template in the template library can be directly called as the basic waveform template based on the recognition result. In some modified embodiments, the basic waveform template is obtained by direct extraction, such as by denoising and smoothing the time-domain signal of the original specific acoustic event, removing instantaneous noise and random fluctuations, and retaining the fundamental frequency and timing structure as the basic waveform template.

[0106] In step S1042 above, the base template is adjusted based on the specific acoustic features. This includes, but is not limited to, adjusting the amplitude of the base waveform template based on the amplitude in the specific acoustic features, so that the amplitude of the base waveform template matches the amplitude of the specific acoustic features. The harmonic frequency distribution in the specific acoustic features is added to the base waveform template to restore the true spectral structure. Timing detail features in the specific acoustic features are extracted, and the timing parameters of the base waveform template are fine-tuned.

[0107] In some implementations, an acoustic processing strategy library is pre-established. Upon identifying specific acoustic events and characteristics, corresponding processing parameters are retrieved from this library. For example, for a rapidly approaching siren from directly behind, frequency domain equalization or filtering is used to appropriately boost high-frequency components to enhance warning effectiveness, and spatial sound effect processing is applied to the signal to make it sound like it's coming from behind. In other implementations, spatial sound effect processing is performed using binaural rendering. A large amount of FRTF data is pre-stored in an HRTF database, including HRTF filters for different directions such as "0° behind," "30° behind," and "60° behind." Each rear direction corresponds to a unique set of filtering parameters used to simulate the ITD, ILD, and frequency response of that direction. After identifying the direction of the sound source, the corresponding HRTF filter is directly invoked. The original sound signal (mono / multi-channel) is input into the HRTF filters of the left and right ears respectively. The left earphone signal passes through the "rear left ear HRTF," simulating the delay and frequency attenuation of sound waves diffracting from the rear to the left ear; the right earphone signal passes through the "rear right ear HRTF," simulating the corresponding physical characteristics. Through the HRTF algorithm, the mono or multi-channel signal is processed to adapt to the physiological characteristics of the human ear, presenting a 360° spatial sense and achieving "rear positioning" using only the left and right channels.

[0108] In some implementations, spatial sound effects processing is performed using Ambisonics 3D spatial audio technology. During the encoding stage, for "directly rear sound sources," the sound field information of the sound source signal is mapped to the "rear channel" of Ambisonics (e.g., the W (omnidirectional), X (front / back), Y (left / right), and Z (up / down) channels of first-order Ambisonics, with the X channel set to a negative value to indicate the rear). Encoding parameters are preset in the acoustic processing strategy library and are directly invoked upon recognition of specific events. During the decoding stage, if the playback device is a multi-channel speaker array, the Ambisonics rear channel signal is allocated to the rear speakers during decoding, allowing the sound signal to be played directly through the rear speakers, physically realizing "sound coming from the rear." If the playback device is a stereo headphone, Ambisonics converts the signal to binaural rendering and decoding, combined with the HRTF algorithm, to convert the rear channel signal into differential signals for the left and right ears, restoring the rear positioning.

[0109] In some implementations, spatial audio processing is performed using Dolby Surround Sound technology. First, object audio mapping is performed, with coordinates preset in a policy library for each location. Once the sound of a specific event is detected, the sound of that event is marked and assigned corresponding spatial coordinates. During multi-channel decoding and playback, the audio signal is distributed to the speakers in the corresponding locations for playback.

[0110] In step S1043 above, the reconstructed acoustic waveform generated in step S1042 is optimized by a modulation function to make the reconstructed acoustic waveform more closely resemble the dynamic characteristics of a real acoustic event. The modulation process includes, but is not limited to, amplitude modulation, frequency modulation, and / or phase modulation.

[0111] In step S1045 above, it is preferable to use the same loudspeaker to play the reverse noise reduction wave and the reconstructed time-domain wave, and the driving signal of the same loudspeaker = reverse noise reduction wave + α × reconstructed time-domain wave, where α is the gain coefficient related to the specific acoustic event.

[0112] refer to Figure 4According to another aspect of this application, an active noise cancellation system 200 for an in-vehicle smart speaker is further provided. The active noise cancellation system 200 includes: an acoustic signal acquisition module 201, a noise cancellation module 202, an acoustic reconstruction module 203, and an acoustic signal playback module 204. The acoustic signal acquisition module 201 includes an external acquisition unit 2011 and an internal acquisition unit 2012. The external acquisition unit 2011 is used to acquire external acoustic signals from outside the vehicle, and the internal acquisition unit 2012 is used to acquire internal acoustic signals from inside the vehicle cabin. The noise cancellation module 202 is used to generate a reverse noise cancellation wave based on the external acoustic signals, and the reverse noise cancellation wave is used to cancel out the external acoustic signals. The acoustic reconstruction module 203 includes a feature extraction unit 2031 and a time-domain reconstruction unit 2032. The feature extraction unit 2031 is used to identify specific acoustic events in the external acoustic signal and specific acoustic features corresponding to the specific acoustic events. The time-domain reconstruction unit 2032 is used to reconstruct a reconstructed time-domain wave with consistent characteristics based on the specific acoustic events and corresponding specific acoustic features. The acoustic signal playback module 204 is connected to the noise reduction module 202 and the acoustic reconstruction module 203, and is used to convert the reverse noise reduction wave and the reconstructed time-domain wave into sound signals. The active noise reduction system 200 of the vehicle-mounted smart speaker is a system corresponding to the active noise reduction method 100 of the vehicle-mounted smart speaker provided in the above embodiment, and is used to execute the active noise reduction method 100 of the vehicle-mounted smart speaker described in the above embodiment. The scheme of the active noise reduction system 200 of the vehicle-mounted smart speaker will not be described again here, but please refer to the description of the active noise reduction method 100 of the vehicle-mounted smart speaker above.

[0113] According to another aspect of this application, a computer-readable storage medium is further provided, storing a computer program that can be loaded by a processor and execute the active noise reduction method for in-vehicle intelligent sound field described in the above embodiments.

[0114] Those skilled in the art should understand that the embodiments of the present invention described above and shown in the accompanying drawings are merely examples and do not limit the invention. The advantages of the present invention have been fully and effectively realized. The functional and structural principles of the present invention have been demonstrated and explained in the embodiments; any variations or modifications can be made to the implementation of the present invention without departing from these principles.

Claims

1. An active noise reduction method for in-vehicle intelligent sound field, characterized in that, The method comprises: acquiring an external acoustic signal outside a vehicle; generating a reverse noise reduction wave based on the external acoustic signal, the reverse noise reduction wave being used to offset the external acoustic signal; identifying a specific acoustic event in the external acoustic signal and a specific acoustic feature corresponding to the specific acoustic event; reconstructing a reconstructed time-domain wave with consistent features based on the specific acoustic event and the specific acoustic feature; playing the reconstructed time-domain wave and the reverse noise reduction wave in the vehicle cabin through a loudspeaker.

2. The method of claim 1, wherein, In the step of playing the reconstructed time-domain wave and the reverse noise reduction wave in the vehicle cabin through a loudspeaker: the reverse noise reduction wave and the reconstructed time-domain wave are played using the same loudspeaker, and the driving signal of the same loudspeaker = reverse noise reduction wave + α × reconstructed time-domain wave, where α is a gain coefficient related to the specific acoustic event, and the specific acoustic event includes a siren, a whistle, an ambulance sound, a train crossing sound, and a tire explosion sound, and the specific acoustic feature includes frequency, intensity, direction, and distance.

3. The method of claim 2, wherein, The identification of the specific acoustic event in the external acoustic signal and the specific acoustic feature corresponding to the specific acoustic event comprises analyzing the external acoustic signal based on a convolutional neural network model, comprising: inputting the external acoustic signal into a trained convolutional neural network model to identify the specific acoustic event in the external acoustic signal and the specific acoustic feature corresponding to the specific acoustic event; the training process of the convolutional neural network model is as follows: acquiring a set of audio signals with specific acoustic events and noise audio to form an audio training set; converting the audio signals in the audio training set into a set of spectrograms; inputting the set of spectrograms into a convolutional neural network model to realize the training of the convolutional neural network model, wherein the convolutional neural network model comprises an input layer, a convolutional layer, a pooling layer, an activation function layer, a fully connected layer, and an output layer, the input layer is used to receive a spectrogram as input, the convolutional layer is used to extract local features in the audio using a plurality of convolutional kernels, the activation function is used to introduce nonlinearity, the fully connected layer is used to map the features extracted by the convolutional layer to the output dimension required by the classification task, and the output layer is used to convert the output into a probability distribution representing the possibility of different sound events.

4. The method of claim 2, wherein, The identification of the specific acoustic event in the external acoustic signal and the specific acoustic feature corresponding to the specific acoustic event comprises analyzing the external acoustic signal based on a recurrent neural network model, comprising: inputting the external acoustic signal into a trained recurrent neural network model to identify the specific acoustic event in the external acoustic signal and the specific acoustic feature corresponding to the specific acoustic event; the training process of the recurrent neural network model is as follows: acquiring a set of audio signals with specific acoustic events and noise audio to form an audio training set; converting the audio signals in the audio training set into a set of spectrograms; inputting the spectrum set into a recurrent neural network model to realize training of the recurrent neural network model, wherein the recurrent neural network model comprises an input layer, a plurality of hidden layers, a full connection layer and an output layer, the hidden layers are used to extract time sequence features, the input layer is used to receive a spectrum as an input, the full connection layer establishes an association between the time sequence features and a sound time category, and the output layer is used to convert an output into a probability distribution representing a possibility of different sound events.

5. The method of claim 2, wherein, Based on the specific acoustic event and the corresponding specific acoustic feature, a step of generating a reconstructed acoustic signal includes: determining a corresponding basic waveform template based on the specific acoustic event; injecting the specific acoustic feature into the basic waveform template to adjust the waveform of the basic waveform template to generate a reconstructed acoustic waveform; modulating the reconstructed acoustic waveform based on a modulation function to generate a time-domain waveform corresponding to the specific acoustic event.

6. The method of claim 1, wherein, Based on the external acoustic signal, a step of generating a reverse noise reduction wave used to offset the external acoustic signal includes: converting the external acoustic signal into a frequency domain feature to obtain a frequency distribution, an amplitude and a phase of the external acoustic signal; performing 180° inversion on the phase of each frequency component to generate a reverse frequency domain signal, and adjusting the amplitude of the reverse frequency domain signal based on the amplitude of the external acoustic signal; converting the reverse frequency domain signal into a reverse time domain signal to generate the reverse noise reduction wave.

7. An active noise reduction system for a car-mounted intelligent speaker, characterized in that, It includes: an acoustic signal acquisition module, which includes an external acquisition unit configured to acquire an external acoustic signal outside a vehicle; a noise reduction module configured to generate a reverse noise reduction wave based on the external acoustic signal, the reverse noise reduction wave being used to offset the external acoustic signal; an acoustic reconstruction module, which includes a feature extraction unit and a time-domain reconstruction unit, the feature extraction unit being configured to identify a specific acoustic event in the external acoustic signal and a specific acoustic feature corresponding to the specific acoustic event; the time-domain reconstruction unit being configured to reconstruct a reconstructed time-domain wave with consistent features based on the specific acoustic event and the corresponding specific acoustic feature; an acoustic signal playing module connected with the noise reduction module and the acoustic reconstruction module, configured to convert the reverse noise reduction wave and the reconstructed time-domain wave into a sound signal.

8. The ANC system of a car smart speaker according to claim 7, wherein, The acoustic signal playing module is configured to play the reverse noise reduction wave and the reconstructed time-domain wave using the same speaker, and a driving signal of the same speaker = reverse noise reduction wave + α × reconstructed time-domain wave, where α is a gain coefficient related to the specific acoustic event, and the specific acoustic event includes a siren, a whistle, an ambulance sound, a train crossing sound and a tire explosion sound, and the specific acoustic feature includes a frequency, an intensity, an azimuth and a distance.

9. The ANC system of a car smart speaker according to claim 7, wherein, The feature extraction unit is adapted to identify the specific acoustic event in the external acoustic signal and the specific acoustic feature corresponding to the specific acoustic event through a convolutional neural network model or a recurrent neural network model.

10. A computer-readable storage medium, characterized in that, A computer program is stored, which can be loaded and executed by the processor to perform the active noise reduction method of the in-vehicle intelligent sound field as claimed in any one of claims 1 to 6.

Images