Voice communication method and device based on artificial intelligence

Abstract

The invention discloses a voice communication method and device based on artificial intelligence, and the method comprises the steps: detecting that the communication device is in an audio output state, and picking up a first sound signal; recognizing the first sound signal based on a trained human voice recognition model; if the target sound signal exists in the first sound signal, detecting whether the distance between the target sound signal and the communication device is smaller than or equal to a preset distance; if yes, determining the frequency range of the first sound signal, and waking up the corresponding primary noise reduction node; the first-level noise reduction nodes are adopted to filter the first sound signal, a signal component corresponding to the frequency range of the target sound signal is extracted, and output signals of all the awakened first-level noise reduction nodes are combined into a second sound signal; and if the residual noise signal exists in the second sound signal, waking up a second-stage noise reduction node, filtering the second sound signal by using the second-stage noise reduction node, removing the residual noise signal, and outputting a third sound signal.

Description

Technical Field

[0001] The present invention relates to the field of voice communication technology, and in particular to a voice communication method and a communication device based on artificial intelligence. Background Art

[0002] With the development of society and the acceleration of urbanization, noise pollution has become an increasingly serious problem. Noise not only interferes with people's daily lives, such as rest, work and study, but long-term exposure to high-decibel noise environments can also lead to health problems such as hearing loss and increased psychological stress. Especially in the field of voice communication, the presence of background noise seriously affects call quality and user experience.

[0003] At present, traditional noise suppression technology mainly relies on hardware-level noise filters or simple software algorithms, which are often ineffective in dealing with noise in complex environments. For example, in public places, busy traffic areas or industrial production environments, traditional technologies have difficulty effectively distinguishing target voice signals from background noise, resulting in poor call quality. Summary of the invention

[0004] In order to solve the problems existing in the prior art, the present invention provides a voice communication method and a communication device based on artificial intelligence.

[0005] The technical solution adopted by the present invention is:

[0006] The first aspect of the present application provides a voice communication method based on artificial intelligence, which is applied to a communication device, including: detecting that the communication device is in an audio output state and picking up a first sound signal; identifying the first sound signal based on a trained human voice recognition model; if there is a target sound signal in the first sound signal, detecting whether the distance between the target sound signal and the communication device is less than or equal to a preset distance; if the distance between the user's first sound signal and the communication device is less than or equal to the preset distance; detecting the frequency component in the first sound signal and determining the frequency range of the first sound signal; based on the frequency range of the first sound signal, waking up the corresponding first-level noise reduction node; using the first-level noise reduction node to filter the first sound signal, extracting the signal component corresponding to the frequency range of the target sound signal, and merging the output signals of all awakened first-level noise reduction nodes into a second sound signal; detecting whether there is a residual noise signal in the second sound signal, if there is a residual noise signal in the second sound signal, waking up the second-level noise reduction node, using the second-level noise reduction node to filter the second sound signal, removing the residual noise signal, and outputting a third sound signal.

[0007] Preferably, the detecting the frequency component in the first sound signal, determining the frequency range of the first sound signal, and waking up the corresponding first-level noise reduction node based on the frequency range of the first sound signal includes the following:

[0008] Detecting the frequency components of the noise signal and the frequency components of the target sound signal in the first sound signal, and determining the frequency range of the noise signal and the frequency range of the target sound signal;

[0009] Based on the frequency range of the noise signal and the frequency range of the target sound signal, a corresponding first-level noise reduction node is awakened, wherein the first-level noise reduction node is provided with a filter and a first noise reduction network model.

[0010] Preferably, filtering the first sound signal using a first-level noise reduction node to extract a signal component corresponding to the frequency range of the target sound signal, and merging the output signals of all awakened first-level noise reduction nodes into a second sound signal includes:

[0011] Calculate the overlap between the frequency range of the noise signal and the frequency range of the target sound signal. If the overlap is less than or equal to a preset overlap threshold, use a filter to filter the noise signal and the target sound signal, extract the signal component corresponding to the frequency range of the target sound signal, and combine the output signals of all awakened first-level noise reduction nodes into a second sound signal.

[0012] If the overlap is greater than the preset overlap threshold, the first noise reduction network model is used to filter the noise signal and the target sound signal, extract the signal components corresponding to the frequency range of the target sound signal, and merge the output signals of all awakened first-level noise reduction nodes into the second sound signal.

[0013] Preferably, the filtering of the first sound signal by the first-level noise reduction node, extracting the signal component corresponding to the frequency range of the target sound signal, and merging the output signals of all awakened first-level noise reduction nodes into the second sound signal comprises the following contents:

[0014] The frequency range of the target user's voice signal and the frequency range of the noise signal are combined and divided into a plurality of signal frequency sub-ranges;

[0015] The first-level noise reduction node includes a central node and multiple sub-nodes, wherein the central node is used to send instructions to the multiple sub-nodes to determine the signal frequency sub-range corresponding to each sub-node for filtering;

[0016] The filtered output signals of all the child nodes are combined into a second sound signal.

[0017] Preferably, the detecting whether there is a residual noise signal in the second sound signal comprises:

[0018] performing spectrum analysis on the second sound signal to determine a noise level;

[0019] If the noise level exceeds a preset noise level threshold, it is considered that there is a residual noise signal; otherwise, it is considered that there is no residual noise signal and a second sound signal is output.

[0020] Preferably, if there is a residual noise signal in the second sound signal, waking up the secondary noise reduction node, using the secondary noise reduction node to filter the second sound signal, removing the residual noise signal, and outputting the third sound signal includes:

[0021] If there is a residual noise signal in the second sound signal, the secondary noise reduction node is awakened, wherein a second noise reduction network model is provided in the secondary noise reduction node, and the second sound signal is filtered using the second noise reduction network model to remove the residual noise signal and output a third sound signal.

[0022] Preferably, using a primary noise reduction node to filter the first sound signal, extracting a signal component corresponding to the frequency range of the target sound signal, and merging the output signals of all awakened primary noise reduction nodes into a second sound signal includes the following contents:

[0023] Perform short-time energy and short-time zero-crossing rate analysis on the noise signal to determine whether the noise signal is a steady-state noise signal or a non-steady-state noise signal. If the noise signal is a steady-state noise signal, calculate the overlap between the frequency range of the noise signal and the frequency range of the target sound signal. If the overlap is less than or equal to a preset overlap threshold, use a filter to filter the noise signal and the target sound signal to extract the signal component corresponding to the frequency range of the target sound signal, and merge the output signals of all awakened first-level noise reduction nodes into a second sound signal.

[0024] If the noise signal is a non-steady-state noise signal or the overlap is greater than a preset overlap threshold, the first noise reduction network model is used to filter the noise signal and the target sound signal, extract the signal components corresponding to the frequency range of the target sound signal, and merge the output signals of all awakened first-level noise reduction nodes into a second sound signal.

[0025] Preferably, detecting whether there is a residual noise signal in the second sound signal, if there is a residual noise signal in the second sound signal, waking up the secondary noise reduction node, using the secondary noise reduction node to filter the second sound signal, after removing the residual noise signal, before outputting the third sound signal, it also includes:

[0026] converting the second sound signal into a first sound digital signal;

[0027] The decibel level of the first sound digital signal is analyzed, and if the decibel level of the first sound digital signal does not meet a preset threshold range, the first sound digital signal is adjusted by a volume adjustment algorithm to generate a third sound signal.

[0028] The second aspect of the present application provides a voice communication device based on artificial intelligence, comprising:

[0029] A sound detection unit, the sound detection unit is used to detect that the communication device is in an audio output state and pick up a first sound signal;

[0030] A sound recognition unit, the sound recognition unit is used to recognize the first sound signal based on a trained human voice recognition model;

[0031] a distance detection unit, wherein if the first sound signal contains a target sound signal, the distance detection unit is used to detect whether the distance between the target sound signal and the communication device is less than or equal to a preset distance;

[0032] A frequency detection unit, if the distance between the user's first sound signal and the communication device is less than or equal to a preset distance; the frequency detection unit is used to detect the frequency component in the first sound signal and determine the frequency range of the first sound signal;

[0033] A wake-up unit, configured to wake up a corresponding first-level noise reduction node based on a frequency range of the first sound signal;

[0034] A primary filtering unit, the primary filtering unit is used to filter the first sound signal using a primary noise reduction node, extract a signal component corresponding to a frequency range of the target sound signal, and combine output signals of all awakened primary noise reduction nodes into a second sound signal;

[0035] a residual noise detection unit, the residual noise detection unit being used to detect whether there is a residual noise signal in the second sound signal;

[0036] If there is a residual noise signal in the second sound signal, the awakening unit is also used to wake up the secondary noise reduction node;

[0037] The secondary filtering unit uses the secondary noise reduction node to filter the second sound signal to remove residual noise signals.

[0038] The output unit is used to output the third sound signal or the second sound signal.

[0039] The beneficial effects of the present invention are at least one of the following:

[0040] By introducing an artificial intelligence-based human voice recognition model and a hierarchical noise reduction mechanism, the present invention can more accurately identify and extract the target voice signal while effectively filtering out background noise. This refined processing method significantly improves the call quality in complex environments.

[0041] In the hierarchical noise reduction mechanism, the corresponding noise reduction node is awakened only when the target sound signal is detected. This on-demand resource allocation method can significantly reduce power consumption during standby and idle states. At the same time, the corresponding noise reduction node is awakened based on the frequency range of the sound signal, which can accurately process the parts related to the target sound signal, reduce the processing of irrelevant signals, and thus save computing resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 The figure is a schematic diagram of the method flow of the first embodiment of the present invention.

[0043] Figure 2 This is a system block diagram of Embodiment 2 of the present invention. DETAILED DESCRIPTION

[0044] The embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0045] Embodiment 1

[0046] The first aspect of the present application provides a voice communication method based on artificial intelligence, which is applied to a communication device, such as Figure 1 The following steps are shown:

[0047] Step 1: Detect that the communication device is in an audio output state and pick up a first sound signal.

[0048] For reference, by checking the state of the audio output port or the internal audio processing module of the communication device, it is determined whether the communication device is in the audio output state. For example, it can be detected whether the earmuff is playing audio or whether the microphone is receiving audio input.

[0049] Select a built-in or external microphone as the device for picking up sound signals. The microphone should have good sensitivity and frequency response characteristics to ensure that it can accurately capture sound signals. After determining that the communication device is in the audio output state, activate the microphone to start picking up sound signals. For example, a microphone array can be used to pick up sound signals from different directions to improve the accuracy of sound positioning.

[0050] Step 2: Recognize the first sound signal based on the trained human voice recognition model.

[0051] For reference, the trained human voice speech recognition model can be a common deep learning-based model in the prior art, such as a convolutional neural network (CNN), a recurrent neural network (RNN), or a long short-term memory network (LSTM). For example, a model based on a combination of a convolutional neural network (CNN) and a long short-term memory network (LSTM) includes the following structure:

[0052] Input Layer: The input layer receives audio features.

[0053] Convolutional layer: The convolutional layer is used to extract local features, usually using multiple convolution kernels for multi-scale feature extraction. Each convolution kernel can capture different feature patterns. The convolutional layer is usually followed by a pooling layer to reduce the feature dimension and reduce the amount of computation.

[0054] LSTM layer: LSTM layer is used to capture long-term dependencies, which is very important for understanding speech signals. LSTM units are able to remember long-term information while ignoring short-term irrelevant information. Bidirectional LSTM can be used to consider both past and future contextual information.

[0055] Fully connected layer: The fully connected layer is used to map the output of the LSTM layer to the classification label. The output layer usually uses the softmax function for probability distribution output.

[0056] Output layer: The output layer gives the result of human voice speech recognition, for example, outputs a probability value indicating the probability that the signal belongs to human voice.

[0057] Step 3: If the first sound signal contains a target sound signal, detecting whether the distance between the target sound signal and the communication device is less than or equal to a preset distance.

[0058] For reference, it is determined whether the target sound signal exists according to the result of recognizing the first sound signal based on the trained human voice recognition model in step 2. If the probability value output by the model is higher than the preset threshold, it is considered that the target sound signal exists, and a variety of methods can be used to detect the distance between the target sound signal and the communication device, such as based on the change of sound intensity, based on the time difference of a multi-microphone array, etc.

[0059] For example: Based on the change of sound intensity: Estimate the distance between the sound source and the communication device by detecting the intensity of the sound signal. Sound intensity is usually inversely proportional to the distance. The distance between the target sound signal and the communication device is calculated based on the relationship formula between sound intensity and distance. Another example: Time difference detection based on a multi-microphone array: If a microphone array is used, the time difference between the sound reaching different microphones can be used to estimate the location of the sound source. The time difference can be determined by comparing the signals received by different microphones. Position calculation: Calculate the approximate location of the sound source based on the time difference and the speed of sound. Calculate the distance between the target sound signal and the communication device through methods such as triangulation.

[0060] The calculated distance between the target sound signal and the communication device is compared with a preset distance threshold, and if the distance between the target sound signal and the communication device is less than or equal to the preset distance threshold, the subsequent steps are continued. If the distance between the target sound signal and the communication device is greater than the preset distance threshold, a prompt is given to inform the user to move closer to the communication device to improve the sound quality.

[0061] Step 4: If the distance between the user's first sound signal and the communication device is less than or equal to the preset distance, detect the frequency component in the first sound signal to determine the frequency range of the first sound signal.

[0062] For reference, according to the result of detecting the distance between the target sound signal and the communication device in step 3, it is determined whether the distance between the target sound signal and the communication device is less than or equal to the preset distance threshold. If the distance is less than or equal to the preset distance threshold, proceed to step 4.

[0063] Detecting the frequency components in the first sound signal and determining the frequency range of the first sound signal may use spectrum analysis tools, such as Fast Fourier Transform (FFT) or other spectrum analysis methods, to analyze the spectrum of the first sound signal and determine the frequency components and distribution of the first sound signal.

[0064] Specifically, the main frequency components in the first sound signal can be identified from the results of the spectrum analysis. The main frequency components are usually the frequency band with the largest signal energy. The frequency range of the first sound signal is determined based on the main frequency components. For example, if the main frequency components are concentrated between 100 Hz and 3 kHz, then this frequency band can be used as the frequency range of the first sound signal.

[0065] Step 5: wake up the corresponding first-level noise reduction node based on the frequency range of the first sound signal.

[0066] For reference, the first-level noise reduction node can be composed of a central node and multiple sub-nodes. The central node is responsible for sending instructions to the sub-nodes to determine the signal frequency sub-range that each sub-node needs to process, and the sub-node performs specific signal processing tasks according to the instructions of the central node. The frequency range of the first sound signal is divided into several signal frequency sub-ranges.

[0067] For example, the frequency range of 100 Hz to 3 kHz can be divided into several sub-ranges, such as 100 Hz to 1 kHz, 1 kHz to 2 kHz, and 2 kHz to 3 kHz. The central node sends instructions to each sub-node to clarify the frequency range that each sub-node needs to process.

[0068] For example, a command is sent to subnode 1 to specify that it is responsible for the frequency range of 100Hz to 1kHz; a command is sent to subnode 2 to specify that it is responsible for the frequency range of 1kHz to 2kHz; a command is sent to subnode 3 to specify that it is responsible for the frequency range of 2kHz to 3kHz. According to the received commands, each subnode starts to perform the specified noise reduction task. For example, subnode 1 starts to process signals in the frequency range of 100Hz to 1kHz. Subnode 2 starts to process signals in the frequency range of 1kHz to 2kHz. Subnode 3 starts to process signals in the frequency range of 2kHz to 3kHz.

[0069] Step 6: Use the first-level noise reduction node to filter the first sound signal, extract the signal component corresponding to the frequency range of the target sound signal, and merge the output signals of all awakened first-level noise reduction nodes into the second sound signal.

[0070] For reference, each sub-node filters the first sound signal according to the received instruction.

[0071] For example, subnode 1 processes signals in the frequency range of 100Hz to 1kHz; subnode 2 processes signals in the frequency range of 1kHz to 2kHz; and subnode 3 processes signals in the frequency range of 2kHz to 3kHz. Each subnode can use technologies such as filters and noise reduction network models to process signals in the corresponding frequency range. For example, subnode 1 uses a bandpass filter to process signals in the frequency range of 100Hz to 1kHz; subnode 2 uses an adaptive filter to process signals in the frequency range of 1kHz to 2kHz; and subnode 3 uses a deep learning-based noise reduction model to process signals in the frequency range of 2kHz to 3kHz. After processing each subnode, the signal components corresponding to the frequency range of the target sound signal are extracted.

[0072] For example, after processing, sub-node 1 extracts the target sound signal components in the frequency range of 100Hz to 1kHz; after processing, sub-node 2 extracts the target sound signal components in the frequency range of 1kHz to 2kHz; after processing, sub-node 3 extracts the target sound signal components in the frequency range of 2kHz to 3kHz.

[0073] Since the sub-nodes process signals in different frequency ranges respectively, the output signals of the sub-nodes can be combined in the frequency domain. For example, the output signals of sub-nodes 1, 2, and 3 can be combined in the order of frequency ranges. This ensures that the second sound signal contains signal components in the entire frequency range from 100 Hz to 3 kHz.

[0074] In a possible implementation, detecting the frequency component in the first sound signal, determining the frequency range of the first sound signal, and waking up the corresponding first-level noise reduction node based on the frequency range of the first sound signal includes the following:

[0075] Detecting the frequency components of the noise signal and the frequency components of the target sound signal in the first sound signal, and determining the frequency range of the noise signal and the frequency range of the target sound signal;

[0076] Based on the frequency range of the noise signal and the frequency range of the target sound signal, a corresponding first-level noise reduction node is awakened, wherein the first-level noise reduction node is provided with a filter and a first noise reduction network model.

[0077] The filter may be a bandpass filter, and the first noise reduction network model may adopt an adaptive filter + deep residual network (ResNet) commonly used in the prior art, including an input layer: receiving preprocessed audio features: for example, linear spectrum features (Linear Spectrogram). Linear spectrum features: by converting the audio signal into a frequency domain representation, a two-dimensional matrix is ​​obtained, in which each column represents the spectrum at a time point. The linear spectrum feature retains more original signal information and is a good basis for subsequent processing.

[0078] Adaptive filter layer: Use adaptive filter algorithms (such as LMS or RLS algorithms) to adjust filter parameters according to real-time input signals. The goal is to remove the main noise components and retain the target sound signal.

[0079] Deep residual network layer: contains multiple residual blocks, each residual block includes two convolutional layers and a skip connection. Each convolutional layer is usually followed by a batch normalization layer and an activation function (such as ReLU). The design of the residual block enables the network to train deeper layers and avoid the gradient vanishing problem. Fully connected layer: used to map the output of the residual network layer to the denoised audio signal. The output layer usually uses a linear activation function to directly output the denoised audio signal. Output layer: outputs the denoised audio signal.

[0080] Considering that in actual application scenarios, background noise and target sound signals often overlap in certain frequency ranges, this makes it difficult for traditional single filters to effectively separate the target sound signals.

[0081] Traditional filters may not handle this situation well, especially when the overlap is high, and simple filters may mistakenly remove part of the target sound signal as noise.

[0082] Therefore, in a possible implementation, the first sound signal is filtered using a first-level noise reduction node to extract a signal component corresponding to the frequency range of the target sound signal, and the output signals of all awakened first-level noise reduction nodes are combined into a second sound signal, including:

[0083] Calculate the overlap between the frequency range of the noise signal and the frequency range of the target sound signal. If the overlap is less than or equal to a preset overlap threshold, use a filter to filter the noise signal and the target sound signal, extract the signal component corresponding to the frequency range of the target sound signal, and combine the output signals of all awakened first-level noise reduction nodes into a second sound signal.

[0084] If the overlap is greater than the preset overlap threshold, the first noise reduction network model is used to filter the noise signal and the target sound signal, extract the signal components corresponding to the frequency range of the target sound signal, and merge the output signals of all awakened first-level noise reduction nodes into the second sound signal.

[0085] In this embodiment, the appropriate noise reduction strategy is dynamically selected by calculating the overlap between the frequency range of the noise signal and the frequency range of the target sound signal.

[0086] When the overlap is low, a filter is used for processing, which can remove background noise without affecting the target sound signal.

[0087] When the overlap is high, the first noise reduction network model is used for processing, using the powerful capabilities of the deep learning model to distinguish and retain the target sound signal while removing background noise.

[0088] For low overlap situations, using filters can save computational resources because the computational effort of filters is relatively small.

[0089] For cases with high overlap, although the use of a denoising network model will increase the amount of computation, it is only used when necessary and can still effectively control the consumption of computing resources overall.

[0090] Considering that in actual applications, user voice signals and background noise often contain multiple frequency components, these components may be intertwined and difficult to be effectively separated by a single filter. Traditional single filters are difficult to handle the situation where multiple frequency components overlap, which can easily cause false filtering or missed filtering, resulting in damage to the target sound signal or failure to effectively remove the background noise.

[0091] Therefore, in a possible implementation, the filtering of the first sound signal by the first-level noise reduction node, extracting the signal component corresponding to the frequency range of the target sound signal, and merging the output signals of all awakened first-level noise reduction nodes into the second sound signal includes the following contents:

[0092] The frequency range of the target user's voice signal and the frequency range of the noise signal are combined and divided into a plurality of signal frequency sub-ranges;

[0093] The first-level noise reduction node includes a central node and multiple sub-nodes, wherein the central node is used to send instructions to the multiple sub-nodes to determine the signal frequency sub-range corresponding to each sub-node for filtering;

[0094] The filtered output signals of all the child nodes are combined into a second sound signal.

[0095] In this embodiment, the frequency range of the target user's voice signal and the frequency range of the noise signal are combined and divided into several signal frequency sub-ranges. The signal components in each sub-range are more concentrated. Each sub-node specializes in processing the signal in a signal frequency sub-range, which can perform noise reduction processing more accurately and avoid the problem of false filtering or leakage filtering that may occur in traditional methods. Through fine division and targeted processing, the effective components of the target user's voice signal can be retained to the maximum extent, and the clarity of the voice can be improved. The central node can dynamically assign tasks to different sub-nodes according to actual conditions, avoiding the waste of computing resources and improving resource utilization.

[0096] In a possible implementation, a first-level noise reduction node is used to filter the first sound signal, a signal component corresponding to the frequency range of the target sound signal is extracted, and the output signals of all awakened first-level noise reduction nodes are combined into a second sound signal, including the following contents:

[0097] Perform short-time energy and short-time zero-crossing rate analysis on the noise signal to determine whether the noise signal is a steady-state noise signal or a non-steady-state noise signal. If the noise signal is a steady-state noise signal, calculate the overlap between the frequency range of the noise signal and the frequency range of the target sound signal. If the overlap is less than or equal to a preset overlap threshold, use a filter to filter the noise signal and the target sound signal to extract the signal component corresponding to the frequency range of the target sound signal, and merge the output signals of all awakened first-level noise reduction nodes into a second sound signal.

[0098] If the noise signal is a non-steady-state noise signal or the overlap is greater than a preset overlap threshold, the first noise reduction network model is used to filter the noise signal and the target sound signal, extract the signal components corresponding to the frequency range of the target sound signal, and merge the output signals of all awakened first-level noise reduction nodes into a second sound signal.

[0099] In this embodiment, steady-state noise (such as constant background noise) has relatively stable frequency components and can be effectively processed by a simple filter. Non-steady-state noise (such as sudden, intermittent noise) has variable frequency components and is difficult to be effectively removed by a simple filter.

[0100] Through the analysis of short-time energy and short-time zero-crossing rate, steady-state noise and non-steady-state noise can be distinguished, so as to select the appropriate noise reduction strategy. For steady-state noise, the use of filters can effectively remove the noise signal while retaining the integrity of the target sound signal. For non-steady-state noise, the use of the first noise reduction network model can better handle the changing noise components and ensure the quality of the target sound signal.

[0101] For steady-state noise, using filters can save computational resources because the computational effort of filters is relatively small.

[0102] For non-steady-state noise, although the use of a denoising network model will increase the amount of calculation to a certain extent, since it is only used when necessary, it can still effectively control the consumption of computing resources overall.

[0103] Step 7, detect whether there is a residual noise signal in the second sound signal. If there is a residual noise signal in the second sound signal, wake up the secondary noise reduction node, use the secondary noise reduction node to filter the second sound signal, remove the residual noise signal, and output a third sound signal. If there is no residual noise signal in the second sound signal, output the second sound signal.

[0104] For reference, in a possible implementation, the detecting whether there is a residual noise signal in the second sound signal includes: performing spectral analysis on the second sound signal to determine the noise level; if the noise level exceeds a preset noise level threshold, it is considered that there is a residual noise signal; otherwise, it is considered that there is no residual noise signal and the second sound signal is output.

[0105] If there is a residual noise signal in the second sound signal, waking up the secondary noise reduction node, using the secondary noise reduction node to filter the second sound signal, removing the residual noise signal, and outputting the third sound signal includes:

[0106] If there is a residual noise signal in the second sound signal, the secondary noise reduction node is awakened, wherein a second noise reduction network model is provided in the secondary noise reduction node, and the second sound signal is filtered using the second noise reduction network model to remove the residual noise signal and output a third sound signal.

[0107] Among them, the second denoising network model can adopt the Transformer model based on the attention mechanism commonly used in the prior art, including an input layer: receiving preprocessed audio features, such as Mel-frequency cepstral coefficients (MFCCs). Encoder layer: comprising multiple encoder blocks, each encoder block including a multi-head self-attention mechanism and a feedforward neural network. Decoder layer: comprising multiple decoder blocks, each decoder block including a multi-head self-attention mechanism, an encoder-decoder attention mechanism and a feedforward neural network. Fully connected layer: mapping the output of the decoder layer to a classification label, and the output layer using a softmax function for probability distribution output. Output layer: outputting the denoised result, such as outputting a clean audio signal.

[0108] In this embodiment, by using a human voice recognition model based on deep learning, the target sound signal and background noise can be distinguished more accurately, thereby improving the accuracy of speech recognition. By combining a convolutional neural network (CNN) and a long short-term memory network (LSTM), the model can not only capture local features, but also process long-term series dependencies, further improving the recognition effect.

[0109] Through the staged processing of the primary and secondary noise reduction nodes, various types of noise can be effectively removed, including steady-state noise and non-steady-state noise. According to the frequency components of the first sound signal, the signal is divided into multiple frequency sub-ranges and processed separately to ensure that the signal in each frequency sub-range is targeted for noise reduction. According to the type of noise signal (steady-state or non-steady-state) and the degree of overlap, the filter or noise reduction network model is intelligently selected for processing, which improves the flexibility and effectiveness of the processing.

[0110] Save computing resources and adopt a dynamic wake-up mechanism to wake up the corresponding noise reduction node only when necessary to avoid unnecessary waste of computing resources. Determine the noise level through spectrum analysis and decide whether further processing is needed based on the noise level threshold, reducing unnecessary computing burden.

[0111] By detecting the distance between the target sound signal and the communication device and prompting the user to move closer when the distance is too far, the quality of the voice signal is guaranteed and the user experience is improved. The signals processed by each sub-node are combined into the second sound signal to ensure the integrity and coherence of the signal.

[0112] In a possible implementation, detecting whether there is a residual noise signal in the second sound signal, and if there is a residual noise signal in the second sound signal, waking up the secondary noise reduction node, using the secondary noise reduction node to filter the second sound signal, after removing the residual noise signal, before outputting the third sound signal, further includes:

[0113] converting the second sound signal into a first sound digital signal;

[0114] The decibel level of the first sound digital signal is analyzed, and if the decibel level of the first sound digital signal does not meet a preset threshold range, the first sound digital signal is adjusted by a volume adjustment algorithm to generate a third sound signal.

[0115] Specifically, to analyze the decibel level of the first sound digital signal, it is necessary to first sample the first sound digital signal at a preset frequency and obtain the length of the first sound digital signal, and then calculate the decibel value based on the sampling point and the length of the first sound digital signal:

[0116] The average decibel value D of the first sound digital signal is calculated using the following formula:

[0117]

[0118] Among them, x i is the i-th sampling point of the first sound digital signal, and N is the length of the first sound digital signal.

[0119] The volume adjustment algorithm includes:

[0120]

[0121] Where G(D) is the gain adjustment function, D min is the minimum value of the threshold range, such as 40 decibels, D max It is the maximum value of the threshold range, such as 60 decibels.

[0122] If the signal decibel value D is lower than the minimum decibel value D min , then according to the formula Calculate the gain to ensure that the signal does not fall below D min ;

[0123] If the decibel value D of the signal is between the minimum decibel value and the maximum decibel value, the gain is 1, that is, no gain adjustment is performed.

[0124] If the signal's decibel value D exceeds the maximum decibel value D max , then according to the formula Calculates the gain to reduce the signal in decibels.

[0125] The integrated gain adjustment function G(D) is used to amplify or reduce the signal. The formula is as follows:

[0126] y i =G(D)·x i

[0127] Among them, y i is the i-th sampling point of the enhanced signal.

[0128] The decibel value D' of the enhanced third sound signal can be calculated using the same formula:

[0129]

[0130] Embodiment 2

[0131] like Figure 2As shown, a voice communication device based on artificial intelligence includes: a sound detection unit, the sound detection unit is used to detect that the communication device is in an audio output state and pick up a first sound signal. A sound recognition unit, the sound recognition unit is used to recognize the first sound signal based on a trained human voice recognition model. A distance detection unit, if there is a target sound signal in the first sound signal, the distance detection unit is used to detect whether the distance between the target sound signal and the communication device is less than or equal to a preset distance. A frequency detection unit, if the distance between the user's first sound signal and the communication device is less than or equal to a preset distance; the frequency detection unit is used to detect the frequency component in the first sound signal and determine the frequency range of the first sound signal. A wake-up unit, the wake-up unit is used to wake up the corresponding first-level noise reduction node based on the frequency range of the first sound signal. A first-level filtering unit, the first-level filtering unit is used to filter the first sound signal using the first-level noise reduction node, extract the signal component corresponding to the frequency range of the target sound signal, and merge the output signals of all awakened first-level noise reduction nodes into a second sound signal. A residual noise detection unit, the residual noise detection unit is used to detect whether there is a residual noise signal in the second sound signal. If there is a residual noise signal in the second sound signal, the awakening unit is also used to awaken the secondary noise reduction node. The secondary filtering unit uses the secondary noise reduction node to filter the second sound signal to remove the residual noise signal. The output unit is used to output the third sound signal or the second sound signal.

[0132] The communication device of this embodiment can be an intercom device or earmuffs. Taking the intercom device as an example, 10 intercom devices form a group, including one host and nine slaves. When the host speaks, all nine slaves can clearly hear the speech. The host is responsible for receiving the user's voice input and sending the processed voice signal to all slaves. It has a high-quality microphone, an audio processing module and a wireless communication module.

[0133] The slave receives the voice signal from the host and plays it out. It has an audio playback module and a wireless communication module. Choose a suitable wireless communication protocol, such as Wi-Fi, Bluetooth 5.0, or a dedicated digital radio communication protocol (such as Zigbee, LoRa). Give priority to protocols that support multicast or broadcast, so that the host can cover all slaves with one transmission. A star topology is adopted, with the host as the central node and nine slaves as peripheral nodes. A stable wireless connection is established between the host and each slave. To ensure that all slaves can receive voice signals synchronously, time synchronization is required between the host and the slave. This can be achieved using NTP (Network Time Protocol) or other time synchronization protocols. Use UDP (User Datagram Protocol) or multicast protocol to send voice data packets to ensure that all slaves can receive data at the same time. If TCP (Transmission Control Protocol) is used, it is necessary to ensure that all slaves play synchronously after receiving the complete data packet.

[0134] The above-mentioned embodiments only express the specific implementation of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A voice communication method based on artificial intelligence, applied to a communication device, characterized in that: include: detecting that the communication device is in an audio output state, and picking up a first sound signal; Recognize the first sound signal based on a trained human voice recognition model; If the first sound signal contains a target sound signal, detecting whether the distance between the target sound signal and the communication device is less than or equal to a preset distance; If the distance between the user's first sound signal and the communication device is less than or equal to the preset distance; then detecting the frequency component in the first sound signal to determine the frequency range of the first sound signal; Based on the frequency range of the first sound signal, waking up the corresponding first-level noise reduction node; The first sound signal is filtered using a first-level noise reduction node to extract a signal component corresponding to the frequency range of the target sound signal, and the output signals of all awakened first-level noise reduction nodes are combined into a second sound signal; Detect whether there is a residual noise signal in the second sound signal. If there is a residual noise signal in the second sound signal, wake up the secondary noise reduction node, use the secondary noise reduction node to filter the second sound signal, remove the residual noise signal, and output a third sound signal.

2. The artificial intelligence-based voice communication method according to claim 1, characterized in that: The detecting the frequency component in the first sound signal, determining the frequency range of the first sound signal, and waking up the corresponding first-level noise reduction node based on the frequency range of the first sound signal includes the following: Detecting the frequency components of the noise signal and the frequency components of the target sound signal in the first sound signal, and determining the frequency range of the noise signal and the frequency range of the target sound signal; Based on the frequency range of the noise signal and the frequency range of the target sound signal, a corresponding first-level noise reduction node is awakened, wherein the first-level noise reduction node is provided with a filter and a first noise reduction network model.

3. The artificial intelligence-based voice communication method according to claim 2, characterized in that: The first sound signal is filtered using a first-level noise reduction node to extract a signal component corresponding to the frequency range of the target sound signal, and the output signals of all awakened first-level noise reduction nodes are combined into a second sound signal, including: Calculate the overlap between the frequency range of the noise signal and the frequency range of the target sound signal. If the overlap is less than or equal to a preset overlap threshold, use a filter to filter the noise signal and the target sound signal, extract the signal component corresponding to the frequency range of the target sound signal, and combine the output signals of all awakened first-level noise reduction nodes into a second sound signal. If the overlap is greater than the preset overlap threshold, the first noise reduction network model is used to filter the noise signal and the target sound signal, extract the signal components corresponding to the frequency range of the target sound signal, and merge the output signals of all awakened first-level noise reduction nodes into the second sound signal.

4. The artificial intelligence-based voice communication method according to claim 1, characterized in that: The method of filtering the first sound signal using the first-level noise reduction node, extracting the signal component corresponding to the frequency range of the target sound signal, and combining the output signals of all awakened first-level noise reduction nodes into the second sound signal includes the following contents: The frequency range of the target user's voice signal and the frequency range of the noise signal are combined and divided into a plurality of signal frequency sub-ranges; The first-level noise reduction node includes a central node and multiple sub-nodes, wherein the central node is used to send instructions to the multiple sub-nodes to determine the signal frequency sub-range corresponding to each sub-node for filtering; The filtered output signals of all the child nodes are combined into a second sound signal.

5. The artificial intelligence-based voice communication method according to claim 1, characterized in that: The detecting whether there is a residual noise signal in the second sound signal comprises: performing spectrum analysis on the second sound signal to determine a noise level; If the noise level exceeds a preset noise level threshold, it is considered that there is a residual noise signal; otherwise, it is considered that there is no residual noise signal and a second sound signal is output.

6. The artificial intelligence-based voice communication method according to claim 5, characterized in that: If there is a residual noise signal in the second sound signal, waking up the secondary noise reduction node, using the secondary noise reduction node to filter the second sound signal, removing the residual noise signal, and outputting the third sound signal includes: If there is a residual noise signal in the second sound signal, the secondary noise reduction node is awakened, wherein a second noise reduction network model is provided in the secondary noise reduction node, and the second sound signal is filtered using the second noise reduction network model to remove the residual noise signal and output a third sound signal.

7. The artificial intelligence-based voice communication method according to claim 2, characterized in that: The first sound signal is filtered using a first-level noise reduction node to extract a signal component corresponding to the frequency range of the target sound signal, and the output signals of all awakened first-level noise reduction nodes are combined into a second sound signal including the following contents: Perform short-time energy and short-time zero-crossing rate analysis on the noise signal to determine whether the noise signal is a steady-state noise signal or a non-steady-state noise signal. If the noise signal is a steady-state noise signal, calculate the overlap between the frequency range of the noise signal and the frequency range of the target sound signal. If the overlap is less than or equal to a preset overlap threshold, use a filter to filter the noise signal and the target sound signal to extract the signal component corresponding to the frequency range of the target sound signal, and merge the output signals of all awakened first-level noise reduction nodes into a second sound signal. If the noise signal is a non-steady-state noise signal or the overlap is greater than a preset overlap threshold, the first noise reduction network model is used to filter the noise signal and the target sound signal, extract the signal components corresponding to the frequency range of the target sound signal, and merge the output signals of all awakened first-level noise reduction nodes into a second sound signal.

8. The artificial intelligence-based voice communication method according to claim 1, characterized in that: Detecting whether there is a residual noise signal in the second sound signal, if there is a residual noise signal in the second sound signal, waking up the secondary noise reduction node, using the secondary noise reduction node to filter the second sound signal, after removing the residual noise signal, before outputting the third sound signal, it also includes: converting the second sound signal into a first sound digital signal; The decibel level of the first sound digital signal is analyzed, and if the decibel level of the first sound digital signal does not meet a preset threshold range, the first sound digital signal is adjusted by a volume adjustment algorithm to generate a third sound signal.

9. A voice communication device based on artificial intelligence, characterized in that: include, A sound detection unit, the sound detection unit is used to detect that the communication device is in an audio output state and pick up a first sound signal; A sound recognition unit, the sound recognition unit is used to recognize the first sound signal based on a trained human voice recognition model; a distance detection unit, wherein if the first sound signal contains a target sound signal, the distance detection unit is used to detect whether the distance between the target sound signal and the communication device is less than or equal to a preset distance; The frequency detection unit, if the distance between the user's first sound signal and the communication device is less than or equal to a preset distance; The frequency detection unit is used to detect the frequency component in the first sound signal and determine the frequency range of the first sound signal; A wake-up unit, configured to wake up a corresponding first-level noise reduction node based on a frequency range of the first sound signal; A primary filtering unit, the primary filtering unit is used to filter the first sound signal using a primary noise reduction node, extract a signal component corresponding to a frequency range of the target sound signal, and combine output signals of all awakened primary noise reduction nodes into a second sound signal; a residual noise detection unit, the residual noise detection unit being used to detect whether there is a residual noise signal in the second sound signal; If there is a residual noise signal in the second sound signal, the awakening unit is also used to wake up the secondary noise reduction node; A secondary filtering unit, wherein the secondary filtering unit uses the secondary noise reduction node to filter the second sound signal to remove residual noise signals; The output unit is used to output the third sound signal or the second sound signal.

Images