Elevator remote alarm intercom noise reduction method and system

Abstract

The present application relates to elevator security technology field, especially to a kind of elevator remote alarm talkback noise reduction method and system, its method includes to noisy speech signal is handled, extracts 40-dimensional speech feature, the 40-dimensional speech feature includes 20 mel frequency cepstrum coefficient, first derivative and second derivative according to order take the first 10 mel frequency cepstrum coefficient;The 40-dimensional speech feature is input to the RNN model that is trained in advance, the RNN model outputs 20 gain coefficients;Based on the 20 gain coefficients dynamically adjusts the coefficient of 20 first-order IIR band-pass filter, the noisy speech signal is parallel through the 20 first-order IIR band-pass filter, the output of each first-order IIR band-pass filter is added, and the noise reduction speech signal is obtained.The present application combines the comprehensive noise reduction mode of neural network and traditional signal processing, can solve the technical problems of poor effect, low efficiency or high cost in the existing elevator remote alarm talkback noise reduction method, and realizes efficient noise reduction on low-cost hardware.

Description

Technical Field

[0001] This invention relates to the field of elevator security technology, and in particular to a method and system for reducing noise in elevator remote alarm intercom. Background Technology

[0002] The elevator five-way intercom system is a key security device that enables communication between the elevator car, elevator control room, and management center. Its core function is to ensure that trapped people can clearly communicate with the outside world in emergency situations such as elevator entrapment. GB / T24475-2023 Elevator Remote Alarm System clearly stipulates that every elevator must be equipped with this system, and specifies its usage and installation location.

[0003] To ensure clear intercom communication, existing technologies mainly employ three noise reduction methods, but all have significant drawbacks:

[0004] The first method is to use discrete LC components to build a bandpass filter circuit for analog signal noise reduction. This method is simple but the noise reduction effect is poor.

[0005] The second method is to use traditional digital filters for frequency domain filtering. However, because it does not take into account the characteristics of human voice, it is difficult to effectively distinguish between noise and human voice.

[0006] The third method is to directly input the voice signal into the neural network model for full processing. Although the effect is better, it requires a massive amount of data and a high-performance intercom terminal, resulting in excessively high costs.

[0007] Therefore, there is an urgent need for a noise reduction method for elevator remote alarm intercom that takes into account noise reduction effect, computing efficiency and cost. Summary of the Invention

[0008] One objective of this invention is to propose a noise reduction method for elevator remote alarm intercom, which combines neural networks and traditional signal processing to solve the technical problems of poor performance, low efficiency, or high cost in existing elevator remote alarm intercom noise reduction methods, and achieves efficient noise reduction on low-cost hardware.

[0009] Another objective of this invention is to provide an elevator remote alarm intercom noise reduction system, which employs the elevator remote alarm intercom noise reduction method described above.

[0010] To achieve this objective, the present invention adopts the following technical solution:

[0011] A method for noise reduction in elevator remote alarm intercom includes the following steps:

[0012] S1. Feature extraction step: Process the noisy speech signal and extract 40-dimensional speech features. The 40-dimensional speech features include 20 Mel-frequency cepstral coefficients. Take the first and second derivatives of the first 10 Mel-frequency cepstral coefficients in sequence.

[0013] S2, Dynamic Gain Output Step: Input the 40-dimensional speech features into a pre-trained RNN model, and the RNN model outputs 20 gain coefficients;

[0014] S3. Filtering process: Based on the 20 gain coefficients, dynamically adjust the coefficients of the 20 first-order IIR bandpass filters. After the noisy speech signal passes through the 20 first-order IIR bandpass filters in parallel, add the outputs of each first-order IIR bandpass filter to obtain the denoised speech signal.

[0015] Preferably, in the feature extraction step, the extraction process of the Mel-frequency cepstral coefficients includes:

[0016] S11. A 50% overlap framing strategy is used to process noisy speech signals: 320 sampling points are collected in the current 20ms of each frame and merged with the 320 sampling points of the previous frame to form 640 sampling points.

[0017] S12. Perform FFT transformation on the 640 sampling points to obtain the amplitude spectrum, and multiply and accumulate the amplitude spectrum with 20 triangular filter groups in sequence to obtain 20 energy values.

[0018] S13. Perform DCT transformation on the 20 energy values ​​and take the first 20 coefficients as Mel-Cepstral Coefficients.

[0019] Preferably, in S12, the center frequencies of the 20 triangular filter banks are determined based on the Mel scale, and the conversion formula between the Mel scale and frequency is:

[0020] ;

[0021] ;

[0022] in, For frequency The corresponding Mel value, The value is Mel;

[0023] The center frequencies of the 20 triangular filters are, in order, 110Hz, 212Hz, 327Hz, 457Hz, 603Hz, 767Hz, 952Hz, 1160Hz, 1394Hz, 1658Hz, 1955Hz, 2290Hz, 2667Hz, 3091Hz, 3568Hz, 4106Hz, 4712Hz, 5395Hz, 6162Hz, and 7026Hz.

[0024] Preferably, in the feature extraction step, the method for calculating the first derivative of the first 10 Mel-frequency cepstral coefficients is as follows:

[0025] ;

[0026] The method for calculating the second derivative of the first 10 Mel-frequency cepstral coefficients is as follows:

[0027] ;

[0028] in, is the nth Mel-frequency cepstral coefficient.

[0029] Preferably, in the dynamic gain output step, the structure of the RNN model is as follows:

[0030] The input layer is sequentially connected to the first GRU layer, the second GRU layer, the third GRU layer, and the fully connected layer. The output of the fully connected layer is processed by the HardSigmoid activation function to obtain 20 gain coefficients.

[0031] The input layer is 40-dimensional, the number of neurons in the first GRU layer, the second GRU layer, and the third GRU layer are 64, 48, and 96 respectively, the fully connected layer has 20 neurons, and the total weight of the RNN model is approximately 120K.

[0032] Preferably, in the dynamic gain output step, the training method of the RNN model includes:

[0033] Dataset preparation: Obtain 2500 clean speech samples, of which 2000 are from the Microsoft customizable speech dataset and 500 are from audio recordings inside an elevator car; add noise to each clean speech sample to generate corresponding noisy speech, resulting in 2500 sets of data, each set including clean speech and noisy speech;

[0034] Training data generation: Each data set is divided into frames, each frame is 40ms and has 50% overlap. 40-dimensional speech features are extracted from each frame as input. 20 energy values ​​obtained by processing clean speech through a triangular filter bank are used as the target gain output to construct the training dataset.

[0035] Model training: The RNN model structure is trained using the training dataset to obtain a trained RNN model.

[0036] Preferably, in the filtering process, the parameters of the first-order IIR bandpass filter satisfy the following: the center frequency of each first-order IIR bandpass filter is the same as the center frequency of the 20 triangular filters; the low-frequency -3dB cutoff frequency is the midpoint between the center frequency of the current first-order IIR bandpass filter and the center frequency of the previous first-order IIR bandpass filter; the high-frequency -3dB cutoff frequency is the midpoint between the center frequency of the current first-order IIR bandpass filter and the center frequency of the next first-order IIR bandpass filter.

[0037] The filtering formula for the first-order IIR bandpass filter is as follows:

[0038] y[i]=b[0]·x[i]+b[1]·x[i-1]+b[2]·x[i-2]-a[1]·y[i-1]-a[2]·y[i-2];

[0039] Where x[i] is the input signal, y[i] is the output signal, and a and b are the filter coefficients;

[0040] The gain coefficient is dynamically updated by adjusting the b coefficient of the first-order IIR bandpass filter:

[0041] currentb[n] = b[n]·g;

[0042] Where g is the gain coefficient.

[0043] An elevator remote alarm intercom noise reduction system includes an input unit, a noise reduction processing unit, and an output unit;

[0044] The input unit is used to collect speech in real time, and the real-time speech frame is 20ms / frame, with 320 sampling points per frame and a sampling rate of 16000Hz.

[0045] The noise reduction processing unit is a noise reduction method for elevator remote alarm intercom as described above, which is used to perform feature extraction and filtering processing on real-time acquired voice combined with the previous frame data, with a processing time of ≤6ms per frame.

[0046] The output unit is used to output the noise-reduced speech.

[0047] One of the above technical solutions has the following beneficial effects:

[0048] 1. Significant noise reduction effect: Compared with the first LC discrete element filtering method in the background technology, it captures speech features by Mel-frequency cepstral coefficients and combines them with RNN model to intelligently distinguish between human voice and noise, which greatly improves the noise reduction accuracy; compared with the second traditional digital filtering method in the background technology, it introduces dynamic speech feature analysis to solve the problem of "noise and human voice in the same frequency band cannot be distinguished".

[0049] 2. Balance between computational efficiency and cost: Avoiding the high computational requirements of the third pure neural network solution in the background technology, the hybrid architecture of "feature extraction + lightweight RNN + traditional filter" reduces computational complexity while ensuring noise reduction effect, without the need for high-performance hardware support, significantly reducing the cost of the intercom terminal.

[0050] 3. Adapt to elevator scenario requirements: Specifically addresses the noise issue in elevator five-way intercom systems, and the results meet the communication clarity requirements of "GB / T24475-2023 Elevator Remote Alarm System", ensuring effective communication in emergency situations such as entrapment.

[0051] 4. Strong dynamic adaptability: By adjusting the filter parameters in real time through the gain coefficient, it can cope with different intensities and types of noise during elevator operation (such as motor operation, car vibration, etc.) and maintain stable noise reduction performance in complex environments. Attached Figure Description

[0052] Figure 1 This is an overall flowchart of a method for reducing noise in a remote alarm intercom system for elevators.

[0053] Figure 2 This is a flowchart illustrating the feature extraction steps in a noise reduction method for elevator remote alarm intercom.

[0054] Figure 3 This is a noise reduction method for elevator remote alarm intercom, which includes the center frequencies of 20 triangular filter banks and their corresponding frequency maps in the frequency dimension during the feature extraction step.

[0055] Figure 4 This is a detailed structural diagram of the RNN model within the dynamic gain output step of a noise reduction method for elevator remote alarm intercom.

[0056] Figure 5 This is a flowchart of the training process of the RNN model in the dynamic gain output step of a noise reduction method for elevator remote alarm intercom.

[0057] Figure 6 This is a frequency diagram of 20 first-order IIR bandpass filters in the filtering process of a noise reduction method for elevator remote alarm intercom.

[0058] Figure 7 This is an overall flowchart of an elevator remote alarm intercom noise reduction system;

[0059] Figure 8 This is a spectrum comparison chart of un-noise-reduced voice and noise-reduced voice after applying this elevator remote alarm intercom noise reduction system. Detailed Implementation

[0060] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0061] A method for noise reduction in elevator remote alarm intercom includes the following steps:

[0062] S1. Feature extraction step: Process the noisy speech signal and extract 40-dimensional speech features. The 40-dimensional speech features include 20 Mel-frequency cepstral coefficients. Take the first and second derivatives of the first 10 Mel-frequency cepstral coefficients in sequence.

[0063] S2, Dynamic Gain Output Step: Input the 40-dimensional speech features into a pre-trained RNN model, and the RNN model outputs 20 gain coefficients;

[0064] S3. Filtering process: Based on the 20 gain coefficients, dynamically adjust the coefficients of the 20 first-order IIR bandpass filters. After the noisy speech signal passes through the 20 first-order IIR bandpass filters in parallel, add the outputs of each first-order IIR bandpass filter to obtain the denoised speech signal.

[0065] like Figure 1-6 As shown, this elevator remote alarm intercom noise reduction method achieves noise reduction through a three-level collaborative mechanism of "feature extraction - dynamic gain - adaptive filtering". The specific principle is as follows:

[0066] S1. Feature Extraction Steps: For noisy speech signals in elevator cars, firstly, 20 Mel-frequency cepstral coefficients are extracted to capture the static spectral features of the speech; then, the first 10 Mel-frequency cepstral coefficients are selected in sequence, and the dynamic time-varying features of the speech are supplemented by calculating the first derivative (reflecting the rate of change of the feature over time) and the second derivative (reflecting the acceleration of the rate of change), ultimately forming a 40-dimensional speech feature vector to comprehensively characterize the difference between human voice and noise.

[0067] S2, Dynamic Gain Output Step: The 40-dimensional speech features are input into a pre-trained RNN model. The model learns the characteristic patterns of speech and noise in the elevator scene and outputs 20 gain coefficients. These coefficients dynamically represent the "voice confidence" in different frequency bands—frequency bands with a high proportion of voice correspond to high gain, and frequency bands with a high proportion of noise correspond to low gain, achieving accurate differentiation between speech and noise frequency bands.

[0068] S3. Filtering Process: Twenty first-order IIR bandpass filters are used in parallel to process the noisy speech, each corresponding to a specific frequency band. The filter parameters are adjusted in real time based on the gain coefficients of the RNN output, allowing the signal to pass through high-gain frequency bands (dominated by human voice) and attenuating low-gain frequency bands (dominated by noise). Finally, the outputs of all filters are superimposed to reconstruct the clear, denoised speech.

[0069] To further explain, the extraction process of the Mel-frequency cepstral coefficients in the feature extraction step includes:

[0070] S11. A 50% overlap framing strategy is used to process noisy speech signals: 320 sampling points are collected in the current 20ms of each frame and merged with the 320 sampling points of the previous frame to form 640 sampling points.

[0071] S12. Perform FFT transformation on the 640 sampling points to obtain the amplitude spectrum, and multiply and accumulate the amplitude spectrum with 20 triangular filter groups in sequence to obtain 20 energy values.

[0072] S13. Perform DCT transformation on the 20 energy values ​​and take the first 20 coefficients as Mel-Cepstral Coefficients.

[0073] In S12, the center frequencies of the 20 triangular filter banks are determined based on the Mel scale, and the conversion formula between the Mel scale and frequency is:

[0074] ;

[0075] ;

[0076] in, For frequency The corresponding Mel value, The value is Mel;

[0077] The center frequencies of the 20 triangular filters are, in order, 110Hz, 212Hz, 327Hz, 457Hz, 603Hz, 767Hz, 952Hz, 1160Hz, 1394Hz, 1658Hz, 1955Hz, 2290Hz, 2667Hz, 3091Hz, 3568Hz, 4106Hz, 4712Hz, 5395Hz, 6162Hz, and 7026Hz.

[0078] like Figure 2 As shown, in the feature extraction step, the extraction process of Mel-frequency cepstral coefficients (MFCC) is based on the characteristics of human auditory perception and precise signal processing logic. The core working principle is as follows:

[0079] Step S11 employs a 50% overlap framing strategy to process the noisy speech signal: 320 sampling points are collected for the current 20ms in each frame (based on a 16000Hz sampling rate, i.e., 16000 sampling points / second × 0.02 seconds = 320 sampling points), and these are combined with the 320 sampling points from the previous frame to form a processing unit of 640 sampling points (corresponding to a 40ms duration). This design avoids speech information breaks at frame boundaries through inter-frame overlap, while ensuring that each frame contains sufficient short-time spectral features, providing complete basic data for subsequent spectral analysis.

[0080] In step S12, a Fast Fourier Transform (FFT) is performed on 640 sampling points to convert the time-domain signal into a frequency-domain amplitude spectrum, obtaining the amplitude values ​​of 640 frequency points (frequency point interval is 16000Hz / 640=25Hz), which fully reflects the frequency distribution of the signal in the range of 20Hz-8000Hz.

[0081] Next, based on the nonlinear perception characteristics of human hearing (linear sensitivity below 1kHz, logarithmic sensitivity above 1kHz), 8000Hz is mapped to 2840.0mel using the Mel transform formula, and the center frequencies of 20 triangular filters are distributed at intervals of 135.2mel (2840 / 21). Each filter has a response of 1 at its center frequency, which linearly decays to 0 towards the center frequencies of adjacent filters, forming a "focused" frequency band isolation.

[0082] Finally, the amplitude spectrum is multiplied and accumulated with 20 triangular filters in sequence to extract the energy value of the frequency band covered by each filter, and finally 20 energy values ​​are obtained, so as to achieve precise focusing on the dominant frequency band of human voice (110Hz-7026Hz) and suppress irrelevant frequency band energy of elevator mechanical noise (such as high-frequency motor noise).

[0083] Step S13 performs Discrete Cosine Transform (DCT) on the energy values ​​of the 20 frequency bands. This orthogonal transformation removes the correlation between energy values, preserving the low-frequency components that reflect the speech spectral envelope. The first 20 coefficients are then used as Mel-frequency Cepstral Coefficients (MFCCs). MFCCs not only contain the static spectral features of speech but also provide the foundation for subsequent calculations of the first and second derivatives (dynamic features), thus fully characterizing the time-frequency properties of speech.

[0084] In summary, the extraction process of Mel-frequency cepstral coefficients (MFCC) adopts a collaborative logic of "framing to preserve integrity → filtering to improve effectiveness → transformation to reduce redundancy". It not only inherits the high efficiency of traditional signal processing, but also incorporates the intelligent feature extraction concept of simulating human ear perception. It solves the problems of "poor analog filtering effect, lack of perceptual adaptation of pure digital filtering, and high cost of pure neural network" in existing technologies, and provides a feature extraction solution for elevator remote alarm intercom system that combines noise reduction effect, real-time performance and low cost.

[0085] Specifically, the center frequencies of the 20 triangular filter banks and their corresponding frequencies in the frequency dimension are shown in the table below. Figure 3 As shown.

[0086]

[0087] To further clarify, in the feature extraction step, the method for calculating the first derivative of the first 10 Mel-frequency cepstral coefficients is as follows:

[0088] ;

[0089] The method for calculating the second derivative of the first 10 Mel-frequency cepstral coefficients is as follows:

[0090] ;

[0091] in, is the nth Mel-frequency cepstral coefficient.

[0092] Specifically, the calculation method for the first derivative is based on the characteristic that "short-term dynamic changes in speech signals contain key semantic information." It quantifies the rate of change of speech features by using the difference between the first 10 MFCCs of adjacent frames (frame (i+1) and frame i). The first 10 MFCCs are chosen because they contain the main energy and spectral envelope information of the speech signal, and their dynamic changes better reflect the time-varying features of human voice such as pitch and speech rate. Higher-order MFCCs, on the other hand, are mostly detail noise and contribute less to dynamic semantics.

[0093] The second derivative is calculated by using the difference between the first derivatives of adjacent frames to quantify the changing trends of dynamic features (such as the rate of speech acceleration / deceleration and the rate of pitch increase / decrease), further characterizing the nonlinear time-varying properties of the speech signal. Since the first derivative has already filtered out some noise, the second derivative can focus more on the subtle dynamic changes in human voice (such as changes in speech rhythm caused by emotional fluctuations).

[0094] In summary, the first and second derivative calculation steps, through progressive feature extraction of "dynamic rate of change → acceleration of change," complement the static MFCC, overcoming the limitation of traditional feature extraction that only focuses on single-frame information. In elevator remote alarm scenarios, this design can capture the dynamic semantics of trapped personnel's speech (such as changes in speech rate and tone during calls for help) and suppress mechanical noise through dynamic feature differences, providing a crucial basis for the subsequent RNN model to output accurate gain coefficients, ultimately achieving the technical goal of "low-cost hardware + high noise reduction effect."

[0095] To further explain, in the dynamic gain output step, the structure of the RNN model is as follows:

[0096] The input layer is sequentially connected to the first GRU layer, the second GRU layer, the third GRU layer, and the fully connected layer. The output of the fully connected layer is processed by the HardSigmoid activation function to obtain 20 gain coefficients.

[0097] The input layer is 40-dimensional, the number of neurons in the first GRU layer, the second GRU layer, and the third GRU layer are 64, 48, and 96 respectively, the fully connected layer has 20 neurons, and the total weight of the RNN model is approximately 120K.

[0098] like Figure 4As shown, the RNN model achieves dynamic analysis of speech features and output of gain coefficients through a collaborative architecture of multi-layer GRU networks and fully connected layers. The specific principle is as follows:

[0099] The input layer receives 40-dimensional speech features (20 Mel-frequency cepstral coefficients, and the first and second derivatives of the first 10 Mel-frequency cepstral coefficients). This dimension directly matches the output of the feature extraction step, ensuring that the static spectral features (MFCC) and dynamic time-varying features (first and second derivatives) of the speech are completely input into the model, providing comprehensive time-frequency information for subsequent feature learning.

[0100] GRU (Gated Recurrent Unit), a variant of recurrent neural networks, has the advantage of capturing temporal dependencies in sequential data, making it suitable for processing dynamic temporal signals such as speech. The model employs a three-layer GRU progressive processing: the first GRU layer (64 neurons): initially encodes the 40-dimensional features of the input, extracts basic temporal features, and filters redundant information; the second GRU layer (48 neurons): further compresses and abstracts the output features of the upper layer, focusing on the core differences between speech and noise (such as the periodic changes in human voice and the irregular fluctuations in elevator mechanical noise); the third GRU layer (96 neurons): expands features through dimensionality enhancement, deeply mining higher-order temporal dependencies (such as dynamic features like speech rhythm and intonation changes), enhancing the ability to recognize speech features in complex elevator environments.

[0101] The output of the third GRU layer is connected to a fully connected layer with 20 neurons. A linear transformation maps the high-order features to a 20-dimensional vector, which is then processed by the HardSigmoid activation function to limit the output value to the range of 0-1, ultimately yielding 20 gain coefficients. The HardSigmoid function is more efficient than the traditional Sigmoid function and can stably output a reasonable gain range, ensuring the effectiveness of subsequent filter adjustments (a gain approaching 1 indicates that the corresponding frequency band is dominated by human voice, while a gain approaching 0 indicates that it is dominated by noise).

[0102] To further explain, in the dynamic gain output step, the training method of the RNN model includes:

[0103] Dataset preparation: Obtain 2500 clean speech samples, of which 2000 are from the Microsoft customizable speech dataset and 500 are from audio recordings inside an elevator car; add noise to each clean speech sample to generate corresponding noisy speech, resulting in 2500 sets of data, each set including clean speech and noisy speech;

[0104] Training data generation: Each data set is divided into frames, each frame is 40ms and has 50% overlap. 40-dimensional speech features are extracted from each frame as input. 20 energy values ​​obtained by processing clean speech through a triangular filter bank are used as the target gain output to construct the training dataset.

[0105] Model training: The RNN model structure is trained using the training dataset to obtain a trained RNN model.

[0106] like Figure 5 As shown, the RNN model is trained on a dataset specifically designed for elevator car scenarios: it uses 2000 general clean speech samples and 500 clean speech samples recorded inside an elevator car. Noisy speech is generated by adding elevator environmental noise (such as motor noise and friction noise), establishing a mapping relationship between "noisy speech features and clean speech energy." During training, a multi-layer GRU learns the temporal differences between speech and noise, optimizing the weight parameters (approximately 120K in total), ultimately achieving real-time analysis of input features and accurate gain output.

[0107] To further clarify, in the filtering process, the parameters of the first-order IIR bandpass filter satisfy the following: the center frequency of each first-order IIR bandpass filter is the same as the center frequency of the 20 triangular filters; the low-frequency -3dB cutoff frequency is the midpoint between the center frequency of the current first-order IIR bandpass filter and the center frequency of the previous first-order IIR bandpass filter; the high-frequency -3dB cutoff frequency is the midpoint between the center frequency of the current first-order IIR bandpass filter and the center frequency of the next first-order IIR bandpass filter.

[0108] The filtering formula for the first-order IIR bandpass filter is as follows:

[0109] y[i]=b[0]·x[i]+b[1]·x[i-1]+b[2]·x[i-2]-a[1]·y[i-1]-a[2]·y[i-2];

[0110] Where x[i] is the input signal, y[i] is the output signal, and a and b are the filter coefficients;

[0111] The gain coefficient is dynamically updated by adjusting the b coefficient of the first-order IIR bandpass filter:

[0112] currentb[n] = b[n]·g;

[0113] Where g is the gain coefficient.

[0114] In the filtering process, the first-order IIR bandpass filter achieves targeted noise reduction through precise parameter design and dynamic adjustment mechanisms. The center frequency of each first-order IIR bandpass filter is completely consistent with the center frequencies of the 20 triangular filters in the feature extraction stage, corresponding sequentially to the frequency bands of 110Hz, 212Hz, and up to 7026Hz, ensuring precise matching with the main distribution range of human voices and achieving frequency band coordination between feature extraction and filtering. To avoid frequency band overlap or gaps, the low-frequency -3dB cutoff frequency of the first-order IIR bandpass filter is set to the midpoint between the current and the previous first-order IIR bandpass filter's center frequencies, and the high-frequency -3dB cutoff frequency is set to the midpoint between the current and the next first-order IIR bandpass filter's center frequencies, ensuring complete coverage of signals in the 20Hz-8000Hz range without redundant processing.

[0115] Specifically, the center frequencies, low-frequency -3dB cutoff frequencies, and high-frequency -3dB cutoff frequencies of the 20 first-order IIR bandpass filters are shown in the table below. Figure 6 As shown.

[0116]

[0117] The filtering process follows the filtering formula of a second-order IIR structure. Stable filtering is achieved by weighting the current input signal and the historical output signal. The coefficient 'a' ensures the stability and attenuation characteristics of the filter, while the coefficient 'b' determines the amplitude-frequency response gain, providing key parameters for dynamic adjustment.

[0118] Meanwhile, the 20 gain coefficients g (range 0-1) output by the RNN model are applied in real time to the b coefficients of the corresponding filters, i.e., the b coefficients of a single first-order IIR bandpass filter are updated in real time: when the gain approaches 1, the b coefficients increase to enhance the signal in the dominant human voice frequency band; when the gain approaches 0, the b coefficients decrease to attenuate the noise frequency band. This dynamic adjustment can flexibly adapt to changes in noise such as motor operation and friction in the elevator car, accurately preserving the human voice signal. It not only solves the problem of poor adaptability of traditional fixed filters, but also achieves efficient processing of 20ms audio per frame ≤6ms on an 800MHz CPU through a lightweight structure, without the need for high-computing hardware support, perfectly balancing noise reduction effect, real-time performance and cost requirements.

[0119] An elevator remote alarm intercom noise reduction system includes an input unit, a noise reduction processing unit, and an output unit;

[0120] The input unit is used to collect speech in real time, and the real-time speech frame is 20ms / frame, with 320 sampling points per frame and a sampling rate of 16000Hz.

[0121] The noise reduction processing unit is a noise reduction method for elevator remote alarm intercom as described above, which is used to perform feature extraction and filtering processing on real-time acquired voice combined with the previous frame data, with a processing time of ≤6ms per frame.

[0122] The output unit is used to output the noise-reduced speech.

[0123] The present invention will be described in detail below with reference to specific steps:

[0124] like Figure 7 As shown, the startup and parameter configuration of this elevator remote alarm intercom noise reduction system are as follows: After the system is powered on, the microphone sampling rate is set to 16000Hz, the frame acquisition length is 20ms, and there are 320 sampling points per frame.

[0125] Real-time feature extraction: For each frame of the current speech, 320 sampling points are collected and combined with the 320 sampling points of the previous frame to form 640 sampling points;

[0126] An FFT transformation was performed on 640 sampling points to obtain the amplitude spectrum;

[0127] The amplitude spectrum is multiplied and summed with 20 triangular filter banks to obtain 20 energy values;

[0128] Perform DCT transformation on the energy value and take the first 20 coefficients as Mel-Cepstral coefficients;

[0129] Calculate the first and second derivatives of the first 10 Mel-Cepstral coefficients to form a 40-dimensional feature.

[0130] Dynamic gain calculation: Input 40-dimensional features into the trained RNN model, and the model outputs 20 gain coefficients, ranging from 0 to 1.

[0131] Real-time filtering: Update the b coefficients of 20 first-order IIR bandpass filters according to the gain coefficients; input the noisy voice signal into the 20 filters in parallel, calculate the output of each filter using the updated b coefficients according to the filtering formula; sum the corresponding points of the 20 output signals, scale them to 16-bit signed numbers and send them to the other end of the intercom.

[0132] Termination condition: When the intercom stop command is triggered, the noise reduction process ends and the noise-reduced voice is output.

[0133] Combination Figure 8 As shown, compared to undenoised speech, this invention, through the above steps, combines the advantages of neural network algorithms (good performance and strong adaptability) with the high efficiency of traditional signal processing methods, and achieves efficient noise reduction for elevator remote alarm intercom on low-cost hardware, meeting the requirements of "GB / T24475-2023 Elevator Remote Alarm System".

[0134] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0135] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for noise reduction in elevator remote alarm intercom, characterized in that, Includes the following steps: S1. Feature extraction step: Process the noisy speech signal and extract 40-dimensional speech features. The 40-dimensional speech features include 20 Mel-frequency cepstral coefficients. Take the first and second derivatives of the first 10 Mel-frequency cepstral coefficients in sequence. The extraction process of the Mel-frequency cepstral coefficients includes: S11. A 50% overlap framing strategy is used to process noisy speech signals: 320 sampling points are collected in the current 20ms of each frame and merged with the 320 sampling points of the previous frame to form 640 sampling points. S12. Perform FFT transformation on the 640 sampling points to obtain the amplitude spectrum. Multiply and accumulate the amplitude spectrum with 20 triangular filter banks in sequence to obtain 20 energy values. The center frequencies of the 20 triangular filters are 110Hz, 212Hz, 327Hz, 457Hz, 603Hz, 767Hz, 952Hz, 1160Hz, 1394Hz, 1658Hz, 1955Hz, 2290Hz, 2667Hz, 3091Hz, 3568Hz, 4106Hz, 4712Hz, 5395Hz, 6162Hz, and 7026Hz in the frequency dimension. S13. Perform DCT transformation on the 20 energy values ​​and take the first 20 coefficients as Mel-Cepstral Coefficients; S2, Dynamic Gain Output Step: Input the 40-dimensional speech features into a pre-trained RNN model, and the RNN model outputs 20 gain coefficients; The structure of the RNN model is as follows: the input layer is sequentially connected to the first GRU layer, the second GRU layer, the third GRU layer and the fully connected layer. The output of the fully connected layer is processed by the HardSigmoid activation function to obtain 20 gain coefficients. The input layer is 40-dimensional, the number of neurons in the first GRU layer, the second GRU layer, and the third GRU layer are 64, 48, and 96 respectively, the fully connected layer has 20 neurons, and the total weight of the RNN model is 120K. The training dataset for the RNN model contains 2,500 clean speech samples, of which 2,000 are from the Microsoft Customizable Speech Dataset and 500 are from audio recordings inside an elevator car. Noise is added to each clean speech sample to generate corresponding noisy speech, resulting in 2,500 sets of data, each set including both clean and noisy speech. S3. Filtering process: Based on the 20 gain coefficients, dynamically adjust the coefficients of the 20 first-order IIR bandpass filters. After the noisy speech signal passes through the 20 first-order IIR bandpass filters in parallel, add the outputs of each first-order IIR bandpass filter to obtain the denoised speech signal. The parameters of the first-order IIR bandpass filter satisfy the following: the center frequency of each first-order IIR bandpass filter is the same as the center frequency of the 20 triangular filters; the low-frequency -3dB cutoff frequency is the midpoint between the center frequency of the current first-order IIR bandpass filter and the center frequency of the previous first-order IIR bandpass filter; the high-frequency -3dB cutoff frequency is the midpoint between the center frequency of the current first-order IIR bandpass filter and the center frequency of the next first-order IIR bandpass filter. The filtering formula for the first-order IIR bandpass filter is as follows: y[i]=b[0]·x[i]+b[1]·x[i-1]+b[2]·x[i-2]-a[1]·y[i-1]-a[2]·y[i-2]; Where x[i] is the input signal, y[i] is the output signal, and a and b are the filter coefficients; The gain coefficient is dynamically updated by adjusting the b coefficient of the first-order IIR bandpass filter: currentb[n] = b[n]·g; Where g is the gain coefficient.

2. The method for noise reduction in elevator remote alarm intercom according to claim 1, characterized in that, In S12, the center frequencies of the 20 triangular filter banks are determined based on the Mel scale, and the conversion formula between the Mel scale and frequency is: ； ； in, For frequency The corresponding Mel value, This is the Mel value.

3. The method for noise reduction in elevator remote alarm intercom according to claim 1, characterized in that, In the feature extraction step, the first derivative of the first 10 Mel-frequency cepstral coefficients is calculated as follows: ； The method for calculating the second derivative of the first 10 Mel-frequency cepstral coefficients is as follows: ； in, is the nth Mel-frequency cepstral coefficient.

4. The method for noise reduction in elevator remote alarm intercom according to claim 1, characterized in that, In the dynamic gain output step, the training method of the RNN model includes: Dataset preparation: Obtain 2500 clean speech samples, of which 2000 are from the Microsoft customizable speech dataset and 500 are from audio recordings inside an elevator car; add noise to each clean speech sample to generate corresponding noisy speech, resulting in 2500 sets of data, each set including clean speech and noisy speech; Training data generation: Each data set is divided into frames, each frame is 40ms and has 50% overlap. 40-dimensional speech features are extracted from each frame as input. 20 energy values ​​obtained by processing clean speech through a triangular filter bank are used as the target gain output to construct the training dataset. Model training: The RNN model structure is trained using the training dataset to obtain a trained RNN model.

5. A remote alarm intercom noise reduction system for elevators, characterized in that, It includes an input unit, a noise reduction processing unit, and an output unit; The input unit is used to collect speech in real time, and the real-time speech frame is 20ms / frame, with 320 sampling points per frame and a sampling rate of 16000Hz. The noise reduction processing unit is a noise reduction method for elevator remote alarm intercom as described in any one of claims 1-4, used to perform feature extraction and filtering processing on real-time acquired voice combined with the previous frame data, with a processing time of ≤6ms per frame. The output unit is used to output the noise-reduced speech.

Images