Mel-scale Frequency Cepstral Coefficients Research Articles

A time-delay neural network for ship-radiated noise recognition based on residual block and attention mechanism

Ship-radiated noise recognition remains a challenging issue in underwater acoustic signal processing for a variety of reasons. The ocean ambient noise influences the signal quality, and the scarcity of samples limits the performance of deep learning algorithms. In this study, a self-attention mechanism and residual block-based time-delay neural network (SRTDNN) is proposed to improve the recognition performance of ship-radiated noise. The noise samples are first segmented and transformed into three-dimensional Mel-scale frequency cepstral coefficients (3D-MFCC), which combine static and dynamic characteristics of targets through the concatenation of MFCC, Delta MFCC and Delta-Delta MFCC. Subsequently, 3D-MFCC are fed as input to the SRTDNN. The SRTDNN extracts time-domain features of inputs through the TDNN layers and then assigns different weights to the frame and channel dimensions of the feature maps through two attention mechanisms. The Res Connections and Res2Block are interspersed to facilitate convergence. During training, a time-domain and frequency-domain data augmentation pipeline is designed to alleviate the lack of samples. Experiments with public and real sea experiment datasets are conducted to compare the SRTDNN with acknowledged high-quality recognition methods. The results show that the SRTDNN has a significant improvement in recognition performance. Furthermore, the ablation tests demonstrate the effectiveness of the proposed method.

Research on lung sound classification model based on dual-channel CNN-LSTM algorithm

ulmonary diseases have a significant impact on human health and life safety, and abnormalities in the lungs are a direct response to lung diseases. Establishing an effective lung sound classification model that can assist in diagnosis is of great significance for electronic auscultation.In addressing the issue of lung sound signal classification, this study introduces a deep learning classification model based on a dual-channel CNN-LSTM algorithm. Initially, Mel-scale Frequency Cepstral Coefficients (MFCC) are employed for feature extraction from the dataset, transforming lung sound signals into Mel spectrograms. On this foundation, a dual-channel algorithm classification model is constructed, with parallel Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) modules. The CNN module is designed to capture spatial dimension features of the input data, while the LSTM module focuses on temporal dimension features. These two feature sets are fused together, enabling the model to classify lung sounds and thereby assisting in diagnosing pulmonary diseases for healthcare practitioners. This experiment used the ICBHI2017 Challenge Lungs dataset and obtained 5054 pieces of data through data augmentation and sampling techniques.The results show that the accuracy, recall, and F1 score of this model reach 99.01%, 99.13%, and 0.9915, respectively, significantly superior to other models, highlighting its practical application value.

Application of multimodal perception scenario construction based on IoT technology in university music teaching

In the contemporary landscape of diversified talent cultivation, enhancing education through intelligent means and expediting the process of talent development stand as paramount pursuits. Within the domain of instrumental music education, beyond merely listening to student performances, it becomes imperative to assess their movements, thus furnishing additional insights to fuel their subsequent growth. This article introduces a novel multimodal information fusion evaluation approach, combining sound information and movement data to address the challenge of evaluating students’ learning status in college music instruction. The proposed framework leverages Internet of Things (IoT) technology, utilizing strategically positioned microphones and cameras within the local area network to accomplish data acquisition. Sound feature extraction is accomplished through the employment of Mel-scale frequency cepstral coefficients (MFCC), while the OpenPose framework in deep learning and convolutional neural networks (CNN) is harnessed to extract action features during students’ performances. Subsequently, the fusion of feature layers is achieved through CNN, culminating in the evaluation of students’ academic efficacy, facilitated by a fully connected network (FCN) and an activation function. In comparison to evaluations conducted by the teacher in the class, this approach achieves an impressive accuracy of 95.7% across the three categories of Excellent, Good, and Failed students’ evaluation processes. This breakthrough offers novel insights for the future of music teaching and interactive class evaluations while expanding the horizons of multimodal information fusion methods’ applications.

Heart Murmur Classification Using a Capsule Neural Network.

The healthcare industry has made significant progress in the diagnosis of heart conditions due to the use of intelligent detection systems such as electrocardiograms, cardiac ultrasounds, and abnormal sound diagnostics that use artificial intelligence (AI) technology, such as convolutional neural networks (CNNs). Over the past few decades, methods for automated segmentation and classification of heart sounds have been widely studied. In many cases, both experimental and clinical data require electrocardiography (ECG)-labeled phonocardiograms (PCGs) or several feature extraction techniques from the mel-scale frequency cepstral coefficient (MFCC) spectrum of heart sounds to achieve better identification results with AI methods. Without good feature extraction techniques, the CNN may face challenges in classifying the MFCC spectrum of heart sounds. To overcome these limitations, we propose a capsule neural network (CapsNet), which can utilize iterative dynamic routing methods to obtain good combinations for layers in the translational equivariance of MFCC spectrum features, thereby improving the prediction accuracy of heart murmur classification. The 2016 PhysioNet heart sound database was used for training and validating the prediction performance of CapsNet and other CNNs. Then, we collected our own dataset of clinical auscultation scenarios for fine-tuning hyperparameters and testing results. CapsNet demonstrated its feasibility by achieving validation accuracies of 90.29% and 91.67% on the test dataset.

Heart sound classification based on equal scale frequency cepstral coefficients and deep learning.

Heart diseases represent a serious medical condition that can be fatal. Therefore, it is critical to investigate the measures of its early prevention. The Mel-scale frequency cepstral coefficients (MFCC) feature has been widely used in the early diagnosis of heart abnormity and achieved promising results. During feature extraction, the Mel-scale triangular overlapping filter set is applied, which makes the frequency response more in line with the human auditory property. However, the frequency of the heart sound signals has no specific relationship with the human auditory system, which may not be suitable for processing of heart sound signals. To overcome this issue and obtain a more objective feature that can better adapt to practical use, in this work, we propose an equal scale frequency cepstral coefficients (EFCC) feature based on replacing the Mel-scale filter set with a set of equally spaced triangular overlapping filters. We further designed classifiers combining convolutional neural network (CNN), recurrent neural network (RNN) and random forest (RF) layers, which can extract both the spatial and temporal information of the input features. We evaluated the proposed algorithm on our database and the PhysioNet Computational Cardiology (CinC) 2016 Challenge Database. Results from ten-fold cross-validation reveal that the EFCC-based features show considerably better performance and robustness than the MFCC-based features on the task of classifying heart sounds from novel patients. Our algorithm can be further used in wearable medical devices to monitor the heart status of patients in real time with high precision, which is of great clinical importance.

Bearing fault diagnosis based on inverted Mel-scale frequency cepstral coefficients and deformable convolution networks

In the real-time test fault diagnosis algorithm based on deep learning, it is difficult to guarantee that the training and testing data come from the same time series distribution. Inconsistent distribution will lead to a decline in diagnostic performance. In addition, the convolutional neural network is limited by the fixed shape of its convolution kernel, which makes it difficult to fully extract the spatial constraint features between fault data. To solve the above problems, this paper proposes a bearing fault diagnosis method based on inverted Mel-scale frequency cepstrum coefficients and deformable convolution networks. The core of traditional Mel-scale frequency cepstrum coefficients is to construct a non-uniformly distributed frequency-domain filter bank. It is characterized by the dense distribution of low-frequency regions and the sparse distribution of high-frequency regions. Considering that the features that can well characterize fault information are concentrated in the high-frequency part, we reconstruct the traditional Mel-scale frequency cepstrum coefficients filter bank and propose a feature extraction method of inverted Mel-scale frequency cepstrum coefficients. This method can obtain the frequency-domain characteristics of bearing vibration signals, highlight the fault information contained in the high-frequency region, and reduce the influence of time series distribution inconsistency between training samples and testing samples on the diagnosis accuracy. In order to further improve the spatial discrimination between different fault categories, the deformable convolution networks model is introduced to extract the spatial distribution information of fault features and improve the accuracy of fault diagnosis. Finally, two public data sets and data from an experimental platform verify that the method can achieve high-precision fault diagnosis, and that inverted Mel-scale Frequency cepstrum coefficients are effective in solving the problem of inconsistent distribution.

Emotion recognition based on speech signals by combiningempirical mode decomposition and deep neural network

This paper proposes a novel method for speech emotion recognition. Empirical mode decomposition (EMD) isapplied in this paper for the extraction of emotional features from speeches, and a deep neural network (DNN)is used to classify speech emotions. This paper enhances the emotional components in speech signals by usingEMD with acoustic feature Mel-Scale Frequency Cepstral Coefficients (MFCCs) to improve the recognition ratesof emotions from speeches using the classifier DNN. In this paper, EMD is first used to decompose the speechsignals, which contain emotional components into multiple intrinsic mode functions (IMFs), and then emotionalfeatures are derived from the IMFs and are calculated using MFCC. Then, the emotional features are used to trainthe DNN model. Finally, a trained model that could recognize the emotional signals is then used to identify emotionsin speeches. Experimental results reveal that the proposed method is effective.

Voice Activity Detection Optimized by Adaptive Attention Span Transformer

Voice Activity Detection (VAD) is a widely used technique for separating vocal regions from audio signals, with applications in voice language coding, noise reduction, and other domains. While various strategies have been proposed to improve VAD performance, such as ACAM, DCU-10, and Tr-VAD, these approaches often suffer from common limitations, including being unsuitable for long audio and being time-consuming. To address these issues, we propose a new method called AAT-VAD, which integrates an adaptive width attention learning mechanism into the classic transformer framework. Our approach involves extracting Mel-scale Frequency Cepstral Coefficients (MFCC) from the Mel scale frequency domain, adding a masking function to each transformer attention head, and inputting the features processed by the transformer encoder layer into the classifier. Experimental results indicate that our method achieves a 12.8% higher F1-score than DCU-10 and a 0.6% higher F1-score than Tr-VAD under different noise interferences. Furthermore, the average detection cost function (DCF) value of our method is only 14.3% of DCU-10 and 92.4% of Tr-VAD, and the test time of AAT-VAD is only 37.4% of that of Tr-VAD for the same noisy vocal mixed audio.

A Binaural MFCC-CNN Sound Quality Model of High-Speed Train

The high-speed train (HST) is one of the most important transport tools in China, and the sound quality of its interior noise affects passengers’ comfort. This paper proposes a HST sound quality model. The model combines Mel-scale frequency cepstral coefficients (MFCCs), the most popular spectral-based input parameter in deep learning models, with convolutional neural networks (CNNs) to evaluate the sound quality of HSTs. Meanwhile, two input channels are applied to simulate binaural hearing so that the different sound signals can be processed separately. The binaural MFCC-CNN model achieves an accuracy of 96.2% and outperforms the traditional shallow neural network model because it considers the time-varying characteristics of noise. The MFCC features are capable of capturing the characteristics of noise and improving the accuracy of sound quality evaluations. Besides, the results suggest that the time and level differences in sound signals are important factors affecting sound quality at low annoyance levels. The proposed model is expected to optimize the comfort of the interior acoustic environment of HSTs.

Analysis of Rehabilitation Occupational Therapy Techniques Based on Instrumental Music Chinese Tonal Language Spectrogram Analysis.

This paper provides an in-depth analysis of timbre-speech spectrograms in instrumental music, designs a model analysis of rehabilitation occupational therapy techniques based on the analysis of timbre-speech spectrograms in instrumental music, and tests the models for comparison. Starting from the mechanism of human articulation, this paper models the process of human expression as a time-varying linear system consisting of excitation, vocal tract, and radiation models. The system's overall architecture is designed according to the characteristics of Chinese speech and everyday speech rehabilitation theory (HSL theory). The dual judgment of temporal threshold and short-time average energy realized the phonetic length training. Tone and clear tone training were achieved by linear predictive coding technique (LPC) and autocorrelation function. Using the DTW technique, isolated word speech recognition was achieved by extracting Mel-scale Frequency Cepstral Coefficients (MFCC) parameters of speech signals. The system designs corresponding training scenes for each training module according to the extracted speech parameters, combines the multimedia speech spectrogram motion situation with the speech parameters, and finally presents the training content as a speech spectrogram, and evaluates the training results through human-machine interaction to stimulate the interest of rehabilitation therapy and realize the speech rehabilitation training of patients. After analyzing the pre- and post-test data, it was found that the p-values of all three groups were <0.05, which was judged to be significantly different. Also, all subjects changed their behavioral data during the treatment. Therefore, it was concluded that the music therapy technique could improve the patients' active gaze communication ability, verbal command ability, and active question-answering ability after summarizing the data, i.e., the hypothesis of this experiment is valid. Therefore, it is believed that the technique of timbre-speech spectrogram analysis in instrumental music can achieve the effect of rehabilitation therapy to a certain extent.

Real-Time Vehicle Sound Detection System Based on Depthwise Separable Convolution Neural Network and Spectrogram Augmentation

This paper proposes a lightweight model combined with data augmentation for vehicle detection in an intelligent sensor system. Vehicle detection can be considered as a binary classification problem, vehicle or non-vehicle. Deep neural networks have shown high accuracy in audio classification, and convolution neural networks are widely used for audio feature extraction and audio classification. However, the performance of deep neural networks is highly dependent on the availability of large quantities of training data. Recordings such as tracked vehicles are limited, and data augmentation techniques can be applied to improve the overall detection accuracy. In our case, spectrogram augmentation is applied on the mel spectrogram before extracting the Mel-scale Frequency Cepstral Coefficients (MFCC) features to improve the robustness of the system. Then depthwise separable convolution is applied to the CNN network for model compression and migrated to the hardware platform of the intelligent sensor system. The proposed approach is evaluated on a dataset recorded in the field using intelligent sensor systems with microphones. The final frame-level accuracy achieved was 94.64% for the test recordings and 34% of the parameters were reduced after compression.

Jarvis: AI Virtual Assistant

"Jarvis" was the protagonist of Tony's Stark starring in the movie Iron Man. Unlike the first jokes when Jarvis was Stark's servant, Jarvis' movie version is a smart computer that communicates critical issues, monitors his family, and helps build and organize his hero's suit. In this project, Jarvis is a Digital Health Assistant who uses social media i.e., voice that creates two-way communication between people and his personnel Computer, performing various tasks such as calling and chatting on WhatsApp, taking notes, locating, etc. In our project, we use the word just as communication means that Jarvis is a Speech recognition program. The concept of speech technology combines two technologies: Synthesizer and vision. Speech synthesizer captures input and generates audio streams as output this is achieved using the pyttsx3 library accompanied by Speech function. Speech recognition. On the other hand, he did the opposite. It treats audio streaming as input and thus converts it into text writing and this output uses the Speech Recognition library accompanied by the TakeCommand function. The word is a symbol of endless knowledge. Direct analysis and integration of complex voice signal is due to the vast amount of information contained in the signal. Therefore, digital signal processes such as Feature Visibility and Feature Matching are introduced to represent the voice signal. For this project, we are directly using a speech engine that uses an output element such as Mel scaled frequency cepstral. The mel scale frequency cepstral coefficients (MFCCs) obtained from the fourier transform and filter bank analysis are probably the most widely used in high-level speech recognition systems. We aim to create more jobs that can help people in their daily lives and reduce their efforts as in this project we have introduced Chrome and Youtube Automation. In our test, we checked that all of this functionality works well. We test this on 2 speakers (1 female and 1 male) to find the exact motive for the speculation.

Speech Emotion Recognition via Multi-Level Attention Network

Aiming to improve the performance of human speech emotion recognition (SER), the existing work has made great progress based on the popular mel-scale frequency cepstral coefficient (MFCC). However, the existing work rarely pays attention to the low-level emotion related features in MFCC, such as the underlying interactive relations. In this letter, we propose a novel multi-level attention network (MLAnet), which contains a multi-scale low-level feature (MLF) extractor and a multi-unit attention (MUA) module. Within the MLF extractor, we minimize the task-irrelevant information which harms the performance of SER by applying the attention mechanism. Since the features extracted by the MLF extractor contain rich domain-specific emotion information, we further present a MUA module to simultaneously weight the features in terms of time, frequency and channel dimensions. In this way, the discriminative emotion features in different dimensions can be extracted by corresponding weighting blocks. Experimental results on two benchmark datasets demonstrate that the proposed method outperforms other state-of-the-art approaches.

Emotion Recognition Based on Speech Signals by Combining Empirical Mode Decomposition and Deep Neural Network

This paper proposes a novel method for speech emotion recognition. Empirical mode decomposition (EMD) is applied in this paper for the extraction of emotional features from speeches, and a deep neural network (DNN) is used to classify speech emotions. This paper enhances the emotional components in speech signals by using EMD with acoustic feature Mel-Scale Frequency Cepstral Coefficients (MFCCs) to improve the recognition rates of emotions from speeches using the classifier DNN. In this paper, EMD is first used to decompose the speech signals, which contain emotional components into multiple intrinsic mode functions (IMFs), and then emotional features are derived from the IMFs and are calculated using MFCC. Then, the emotional features are used to train the DNN model. Finally, a trained model that could recognize the emotional signals is then used to identify emotions in speeches. Experimental results reveal that the proposed method is effective.

Research on Scattering Transform of Urban Sound Events Detection Based on Self-Attention Mechanism

Urban sound event detection can automatically preload relevant information for the robot to ensure that it can be competent for various scene activity tasks. Aiming at the limitations of timbre similarity and scene recognition limited by audio collection devices, a fusion model based on self-attention mechanism is proposed in this paper. The model consists of scattering transform and self-attention model. Scattering transform computes modulation spectrum coefficients of multiple orders, through cascades of wavelet convolutions and modulus operators, and it is learnable compared with Mel-scale Frequency Cepstral Coefficients (MFCC), and can be used to better restore the semantic features of some sound scenes with similar timbres. Transformer has an outstanding effect on Natural Language Processing (NLP) with its self-attention mechanism. In this paper, the self-attention mechanism in its encoder is used in the model, mainly to make the feature granularity consistent to refine the features. in addition, Focal Loss function is adopted in the model to curb the problem of sample distribution imbalance. The datasets Google-Command and ESC-50 are used to supplement the scene categories of dataset UrbanSound8K. The model parameters of the learnable filters that performed well on the dataset UrbanSound8K were preserved to fine-tune the other two datasets with insufficient data volume and more target categories. The length of slice duration are further explored the on the model. Experimental results show that the model can achieve better performance in a large range of scene models.

Hardware Accelerator Design of DCT Algorithm With Unique-Group Cosine Coefficients for Mel-Scale Frequency Cepstral Coefficients

This study presents a compact <i>L</i>-points discrete cosine transform (DCT) hardware accelerator for <i>M</i>-points Mel-scale Frequency Cepstral Coefficients (MFCC). The main contributions of this work can be summarized as 1) proposing an algorithm with lower complexity; 2) achieving higher accuracy performance; 3) implementing a low-cost accelerator with a unique group of cosine coefficients. For algorithm derivation, the proposed method converts the original formula into the type IV of discrete cosine transform (DCT-IV) with a preprocessing procedure. The kernel computation of DCT-IV can be further derived into the same cosine multiplication with the proposed preprocessing. Therefore, a total of (<i>M</i>-1) (<i>L</i>-1) additions, (<i>M</i>-1) <i>L</i> multiplications, and <i>L</i> coefficients are required for the computation. Compared with Jo <i>et al</i>.&#x2019;s algorithm, the proposed method respectively reduces the number of additions and multiplications by 42.32 % and 41.67 %. Instead, the number of coefficients is increased by 33.33 %. Moreover, the proposed algorithm exhibits a higher peak signal-to-noise ratio (PSNR) value which is achieved at 90.1dB with a 16-bit coefficient word length. For hardware realization, the FPGA implementation results show that it can operate at a clock rate of 135.85 MHz and requires only 113 combinational elements, 87 registers, 3 DSP multipliers, 64&#x00D7;16 bits RAM and 32&#x00D7;16 bits ROM. Overall, it would be a good choice for integrating MFCC applications in the future.

Dynamic mode decomposition: A feature extraction technique based hidden Markov model for detection of Mysticetes' vocalisations

The detection and classification of Mysticetes' vocalisations have evoked the attention of researchers over the years because of their relevance to the marine ecosystem. These vocalisations are gathered employing different passive acoustic monitoring techniques. The vocalisation datasets are accumulated over a period; thus, they are large and cannot be easily analysed manually. Consequently, efficient machine learning (ML) tools such as the hidden Markov models (HMMs) are used extensively to automatically detect and classify these huge vocalisation datasets. As with most ML tools, the detection efficiency of the HMMs depend on the adopted feature extraction technique. Feature extraction techniques such as the Mel-scale frequency cepstral coefficient (MFCC), linear predictive coefficient (LPC), and empirical mode decomposition (EMD) have been adopted with the HMMs to detect different Mysticetes' vocalisations. This article introduces the method of dynamic mode decomposition (DMD) as a feature extraction technique that can be adopted with the HMMs for the detection of Mysticetes' vocalisations. The DMD has emerged as a robust tool for analysing the dynamics of non-stationary and non-linear signals. It is a completely data-driven tool that decomposes a signal over a certain period of time into relevant modes. Here, these spatial-temporal modes are reconstructed mathematically to form reliable feature vectors of the decomposed vocalisation signals. The performance of the proposed DMD-HMM detection technique is demonstrated using the acoustic datasets of two different Mysticetes' vocalisations: Humpback whale songs and Inshore Bryde's whale short pulse calls. In both species, the proposed DMD-HMM exhibits superior sensitivity and false discovery rate performances as compared to the MFCC-HMM, LPC-HMM, and EMD-HMM detection methods. Likewise, this proposed DMD-HMM detection method can be extended to other Mysticetes' that produce characteristics sounds.

Detection of Bryde's whale short pulse calls using time domain features with hidden Markov models

Passive acoustic monitoring (PAM) is generally usedto extract acoustic signals produced by cetaceans. However, the large data volume from the PAM process is better analyzed using an automated technique such as the hidden Markovmodels (HMM). In this paper, the HMM is used as a detection and classification technique due to its robustness and low time complexity. Nonetheless, certain parameters, such as the choice of features to be extracted from the signal, the frame duration, and the number of states affect the performance of the model. Theresults show that HMM exhibits best performances as the number of states increases with short frame duration. However, increasing the number of states creates more computational complexity in the model. The inshore Bryde's whales produce short pulse calls with distinct signal features, which are observable in the time-domain. Hence, a time-domain feature vector is utilized to reduce the complexity of the HMM. Simulation results also show that average power as a time-domain feature vector provides the best performance compared to other feature vectors for detecting the short pulse call of inshore Bryde's whales based on the HMM technique. More so, the extracted features such as the average power, mean, and zero-crossing rate, are combined to form a single 3-dimensional vector (PaMZ). The PaMZ-HMM shows improved performance and reduced complexity over existing feature extraction techniques such as Mel-scale frequency cepstral coefficients (MFCC) and linear predictive coding (LPC). Thus, making the PaMZ-HMM suitable for real-time detection.

Implementation of Machine Learning for Fault Classification on Vehicle Power Transmission System

This research presents the implementation of machine learning (ML) for fault classification and diagnosis on vehicle power transmission system (VPTS). Machine learning method can be used to classify their independent diagnostic components for each fault characteristic states. Under the internet of vehicle (IoV) demands, early prediction system is necessary to notify the drivers or clouding services how the vehicle maintenance and the driving safety degree. The acoustic sensors can be carried out to realize a real-time diagnostic system for automobile engine and chassis transmission system. This method is to acquire the dynamics acoustic signals of the vehicle through the data acquisition device (DAQ). These acoustic features is firstly filtered by Mel-scale frequency cepstral coefficient (MFCC) to determine the each characteristic states of the vehicle engine and the chassis parts. Next, support vector machine (SVM), multilayer perceptron (MLP), deep neural networks (DNN),principal component analysis (PCA), ${k}$ -nearest neighbor (${k}$ -NN), and decision tree (DT) several classifier algorithms are applied to implement the feature classification of fault causes for stability and higher accuracy of VPTS. And dimension reduction model is compared and applied in proposed ML algorithms by an PCA algorithm. All training model datasets are carried out in Matlab and Python pytorch platform by using Nvidia graphics processing unit (GPU) processors, they are evaluated and discussed. The effectiveness on the filtered feature database in the experiments is classified by means of this research proposed schemes. The expected experimental results of the classification and identification with respect to different fifteen VPTS conditions are obtained and inferred.

A 22nm, 10.8 μ W/15.1 μ W Dual Computing Modes High Power-Performance-Area Efficiency Domained Background Noise Aware Keyword- Spotting Processor

This paper proposes a high power-performance-area efficient background noise aware keyword-spotting (KWS) processor based on an optimized binarized weight network (BWN). To reduce the power consumption while maintaining the system recognition accuracy for different background noise, the KWS processor with a SNR prediction module can be adaptively configured to use dual computing modes (standard computing mode and approximate computing mode) for both high recognition accuracy under high background noise and ultra-low power consumption under low background noise. The mel-scale frequency cepstral coefficients (MFCC) module is optimized with approximate computing technologies, which can reduce the power consumption by up to $3.1\times $ and $5.7\times $ for high/low background noise, respectively. Based on the evaluation of the architecture design space exploration, an ultra-low power BWN accelerator with low voltage, area and leakage power and using precision self-adaptive approximate computing units was proposed. Evaluated under 22nm process technology, this work can support up to 10 keywords real time recognition with power consumption of $15.1~\mu \text{W}$ for high background noise and $10.8~\mu \text{W}$ for low background noise. Compared to the state-of-the-art KWS architectures, our work can achieve ultra-low power consumption (about $1.7\times $ reduced), while maintaining high system capability and adaptability.