publisher

keywords

sound classification applications/use cases

algorithms

preprocessing

Feature Extraction

steps/process

Denoising techniques

Tools used

Context awareness in sound classification

design/model/framework

setup images

field tests

challenges/limitations

Accuracy Levels

Other Information

Classifies

formula

Algorithm/Flowchart

Architecture

Pseudocode

Network/Component Diagram

HONG R. et. al

Video Accessibility Enhancement for Hearing-Impaired Users

Accessibility, dynamic captioning, hearing impairment

Dynamic Captioning

Viola Jones, Haar feature based cascade mouth detector

script location, script-speech alignment, and voice volume estimation

videos along with scripts but can be extended to process general videos without scripts

dynamic captioning put scripts at suitable positions to help the hearing-impaired audience better recognize the speakers

script location, script-speech alignment, and voice volume estimation

Video Feeds

better tracking of the scripts and perceive the moods that are conveyed by the variation of volume

Focus more on dynamic captioning rather than the user interface

Speech in video to dynamic caption

Y-Gaussian distance, linear representation

Y-accessibility enhancement & script-speech alignment

face mapping

Wang W. et. al

A Smartphone-based Digital Hearing Aid to Mitigate Hearing Loss at Specific Frequencies

Digital hearing aids, smartphone, sound classification

Hearing Loss of certain frequencies among elderly

GMM Classifier, WOLA (Weighted Over-Lap Add) filter bank

speech processing in the frequency domain and sound classification to classify input sounds into speech and speech with noise categories.

WOLA filter banks then split the sound up into different frequency bands, which are then amplified (reduced) by the amplification in the specific frequency ranges at which
the user’s hearing is impaired. Finally, the WOLA synthesis filter bank reconstructs the acoustic signal from the amplified sub-band signals, which is sent to the receiver for play
out.

Smartphones

script location, script-speech alignment, and voice volume estimation

acoustic signals

frequency domain processing is currently a bit slow due to computational complexity

Audio Frequencies

hearing aid app (Application and Storage)

Bountourakis V. et. al

Machine Learning Algorithms for Environmental Sound Recognition: Towards Soundscape Semantics

Environmental Sound Recognition, audio classification, semantic audio analysis, computer audition, feature extraction, feature selection, machine learning algorithms

Comparison between algorithms for sound classification

• k-Nearest Neighbors (k-NN)
• Naive Bayes
• Support Vector Machines (SVM)
• C4.5 algorithm (decision tree)
• Logistic Regression
• Artificial Neural Networks (ANN)

stationary (frequency-based) feature extraction and non-stationary (time frequency based) feature extraction

database, segmentation, feature extraction, feature selection, classification, evaluation

Environmental Sounds

database, segmentation, feature extraction, feature selection, classification, evaluation

Sound Signals

the highest classification rates were achieved by k-NN with feature set 3 (85.8%), ANN with feature set 2 and use of PCA (86.95%) and SVM with feature set 2 and use of PCA (85.41%).

airplanes, alarms, applause, birds, dogs, footsteps, motorcycles, rain, rivers, sea waves, thunders, wind.

Bragg D. et. al

A Personalizable Mobile Sound Detector App Design for Deaf and Hard-of-Hearing Users

Sound detection, accessibility, deaf, hard-of-hearing

deaf and hard-of-hearing people to be notified about sounds around them

Smartphones, mobile application

Accesibility

Sound Signals

87 participants (51 female, 36 male). 50 were deaf, and 37 were hard-of-hearing. Ages ranged 18-99 (mean 42, std dev 17).

app design to be usable for deaf and hard-of-hearing users recording training examples of sound

No participants used apps to monitor sounds outside of the study

participants revealed they wanted classifications of dropping items, walking/running behind, moving carts, fire drill, printer, conversations, and baby sound.

vehicles passing by, children having bad dreams, smoke and carbon monoxide detectors, app
pliances making unusual noises, water running, socializing,
something dropping on the floor, gunshots, conversations,
and distinguishing between multiple sources with similar frequency range.

Users train the system using the sounds at home

Kurnaz S. & Aljabery M.

Predict the type of hearing aid of audiology patients using data mining techniques

audiology, National Health System, audiograms, BTE, ITE, Machine Learning, Data Mining, Hearing Aid.

Choice of hearing aid

AdaBoost classifier, Random forests classification, Logistic Regression, Orange Canvas Modeler

Hearing aid choice

ML Model

Li M. et. al

Environmental Noise Classification Using Convolution Neural Networks

Environmental noise; Convolution Neural Network (CNN); Short-Time Fourier Transform (STFT); Log Mel-Frequency Spectral Coefficients (MFSCs); Tensorflow

Short-Time Fourier Transform (STFT

Environment

Log Mel-Frequency Spectral Coefficients

Y ML, STFT

Alsouda Y. et. al.

IoT-based Urban Noise Identification Using Machine Learning: Performance of SVM, KNN, Bagging, and Random Forest

urban noise; smart cities; support vector machine (SVM); k-nearest neighbors (KNN); bootstrap aggregation (Bagging); random forest; mel-frequency cepstral coefficients (MFCC); internet of things (IoT).

classification of environmental sounds

SVM, KNN, Bagging, Random Forest

mel-frequency cepstral coefficients (MFCC)

Feature extraction, model training, classifier, prediction

Raspberry pi, Microphone hat

Environment

Feature extraction, model training, classifier, prediction

high noise identification accuracy that is in the range 88% – 94%. E

Classifier SVM KNN Bagging Random Forest Accuracy [%]
93.87 93.88 87.81 89.91

quietness, silence, car horn, children playing, gunshot, jackhammer, siren, and street music

K-Nearest Neighbors

ML, MFCC

Wang . et. al.

Privacy-aware environmental sound classification for indoor human activity recognition

Smart Buildings, Privacy-aware Environmental Sound Recognition, Voice Bands Stripping, Internet Of Things, Computational Efficiency, Web Crawling, Mel Frequency Cepstral Coefficients, Linear Predictive Cepstral Coefficients, Support Vector Machine

indoor environmental sound classification

Decision tree, Random Forest, Mixed Gaussian, Naive Bayes, SVM(Linear & RBF kernel), Artificial Neural Network

Environment

ML, Feature extraction

Inik O. & Seker H

Convolutional Neural Networks for the Classification of Environmental Sounds

Environmental sound classification (ESC), Deep Learning, Convo�lutional Neural Networks (CNN), Urbansound8k

classification of environmental sounds

Intel®Core™ i9-7900X 3.30GHz×20 processor, 64 GB Ram and 2 x GeForce RTX2080Ti graphic card. Matlab R2020a 64bit (win64)

Environment

e air conditioner, car horn, children playing, dog bark, drilling, engine idling, gun shot, jackhammer, siren, and street music

Sigtia S. et. al.

Automatic Environmental Sound Recognition: Performance Versus Computational Cost

Automatic environmental sound recognition, computational auditory scene analysis, deep learning, machine learning.

classification of environmental sounds

Gaussian Mixture Models, SVM, DNN, RNN

Mel-frequency cepstral coefficient (MFCC)

Baby Cry Data Set, Smoke Alarm Data Set

Environment

Deep Neural Networks yield the best ratio of sound classification accuracy across a range of computational costs, while Gaussian Mixture Models offer a reasonable accuracy at a consis�tently small cost, and Support Vector Machines stand between both in terms of compromise between accuracy and computational cost.

smoke alarms and baby cries

Gaussian Mixture Models, SVM, DNN (Feed forward) RNN

Laar V. & Vries. B

A Probabilistic Modeling Approach to Hearing Loss Compensation

Hearing aids, hearing loss compensation, probabilistic modeling, factor graphs, message passing, machine learning

Hearing Aid (HA) algorithms tuning, probabilistic modeling approach
to the design of HA algorithms

Bayes factor (BF)

signal processing (SP), parameter estimation (PE) and model comparison (MC) tasks evaluation

Speech Understanding

performance evaluation, signal processing

hearting aid signal processing

hearing aid agent

Salehi H. et. al.

Learning-Based Reference-Free Speech Quality Measures for Hearing Aid Applications

Hearing aids, speech quality, perceptual linear prediction, gammatone filterbank energies, reference-free quality assessment, support vector regression, machine learning.

Speech quality of hearing aids

Speech Understanding

A group of 18 HI listeners were recruited to provide the speech quality ratings

Linear prediction

Feature extraction

Demir F. et. al

A New Deep CNN Model for Environmental Sound Classification

Environmental sound classification, spectrogram images, CNN model, deep features

Environmental sound classification

spectrogram method converts the signals into time frequency images or loudness of a signal over time at different frequencies existing in a specific waveform
deep feature extraction

DCASE-2017 ASC and the UrbanSound8K datasets

Environment

air conditioner, car horn, children, dog bark drilling, engine idling, gun shot, jack�hammer, siren, and street music

STFT, CNN, Accuracy

ML, CNN, KNN

Ridha. A & Shehieb W.

Assistive Technology for Hearing-Impaired and Deaf Students Utilizing Augmented Reality

Assistive technology; Augmented Reality; Deaf; Education; Hearing-Impairment; Machine Learning.

augmented reality glasses that will assist students in their educational journey with real-time transcribing, speech emotion recognition, sound indications features, as well as classroom assistive tools.

AR Glasses

Environmental Sounds

live transcription feature that uses Google Cloud services, additionally storing the transcribed lectures for future reference in the classroom tools feature, that can also be shared among other students, making it a platform that can be used for communication between students

Car Horn, Siren, Gunshots, Broken Glass

PCB Schematic

Melati A.& Karyono K.

ANDROID BASED SOUND DETECTION APPLICATION FOR HEARING-IMPAIRED USING ADABOOSTM1 CLASSIFIER WITH REPTREE WEAKLEARNER

sound detection for hearing-impaired, machine learning, AdaBoostM1, REPTree, Android

help the hearing-impaired people to detect sound around them and to recognize the sound

AdaBoostM1 functioning as a classifier and REPTree as weak learner

indoor sounds and the second database is outdoor sounds with a total of 23 sounds

Environment

Low Accuracy, propose better approach

baby crying x beep x broom sweeps x door creaking x door slam x door bell x foot step x hairdryer x knocking door x ringing x water runs x whistle
airplanes
x applause
x birds chirp
x car honk
x crowded
x dog bark
x engine start
x screaming
x thunder
x train
x wind blowing

AdaBoostM, Bagging

Chen C. et. al.

Audio-Based Early Warning System of Sound Events on the Road for Improving the Safety of Hearing-Impaired People

Android application, warning, audio detection, machine learning

Road Safety for hearing impared

urbansound 8k

CNN is effective for environment sounds classification tasks by appropriate parameter settings and feature sets

Car-approaching, Car-horn, Children-playing, Dog-barking, Gun-shot, Construction, Siren, Engine-idling

Bhat G. et. al.

Automated machine learning based speech classification for hearing aid applications and its real-time implementation on smartphone

Automated Machine Learning, AutoML, Voice Activity Detection (VAD), Hearing aid devices (HADs), smartphone, real-time

speech classification

AutoML based VAD, CNN

Speech Understanding

Signal model and training feature

Healy E. & Yoho S.

Difficulty understanding speech in noise by the hearing impaired: Underlying causes and technological solutions

poor speech understanding

single-microphone algorithm to extract speech from noise, DNN

Speech Understanding

Groups of 10 NH and 10 HI subjects heard IEEE sentences in unprocessed speech-plus-noise conditions and corresponding algorithm-processed conditions. In this study, multi-talker babble and cafeteria noise, each at two SNRs, were employed.

Jatturas C. et. al.

Feature-based and Deep Learning-based Classification of Environmental Sound

comparison techniques for environmental sound classification

SVM, MLP, Deep Learning

Urban Sound 8k, Scikit-learn and Tensorflow

Environment

Air cond., Children Playing, Engine Idling, Siren, and Street Music.,

STFT, NN, SVM

Saleem N. et. al.

Machine Learning Approach for Improving the Intelligibility of Noisy Speech

Machine learning, speech enhancement, intelligibility, time-frequency masking, deep neural networks

Intelligibility of Noisy Speech

Speech Understanding

Jatturas C. et. al.

Recurrent Neural Networks for Environmental Sound Recognition using Scikit-learn and Tensorflow

Environmental sound classification

MLP, SVM

Urban Sound 8k,

Environment

deep neural network models outperform both MLP and SVM with PCA

Car-approaching, Car-horn, Children-playing, Dog-barking, Gun-shot, Construction, Siren, Engine-idling

STFT, SVM

Davis. N & Suresh. K

Environmental Sound Classification using Deep Convolutional Neural Networks and Data

Environmental sound classification

Conditional Neural Network

Time Stretchnig, pitch shifting, Dynamic Range Compression, Background Noise, Linear Prediction Ceptal Coefficients (LPCC)

Urbansound 8K

Environment

air conditioner, car horns, children playing, dog bark, drilling, engine idling, gunshot, jackhammers, siren and street music

Chu. S et. al

Environmental Sound Recognition With Time–Frequency Audio Features

Terms—Audio classification, auditory scene recognition, data representation, feature extraction, feature selection, matching pursuit, Mel-frequency cepstral coefficient (MFCC).

Environmental sound classification

Environment

that Restaurant, Casino,Train, Rain, and Street ambulance

short time energy, zero crossing rate, signal decomposition

Chu. S et. al

WHERE AM I? SCENE RECOGNITION FOR MOBILE ROBOTS USING AUDIO FEATURES

Environmental sound classification

Environment

Ullo. S et. al

Hybrid Computerized Method for Environmental Sound Classification

Environmental sound classification, Optimal allocation sampling, spectrogram, convolu�tional neural network, classification techniques

Environmental sound classification

AlexNet and Visual Geometry Group (VGG)-16 networks
decision tree (fine, medium, coarse kernel),
k-nearest neighbor (fine, medium, cosine, cubic, coarse and weighted kernel), support vector machine,
linear discriminant analysis, bagged tree and softmax classifiers

short-time Fourier transform (STFT)

Deep Feature Extraction

ESC-10, a ten-class environmental sound dataset,
The experiments have
been carried out on MATLAB (2018R). A computer with
8 GB RAM, intel i7 third generation processor of 3.4 GHz,
64 bit memory has been used.

Environment

AlexNet (FC-6) for fine kernel using a decision tree is 89.9%

90.1%, 95.8%, 94.7%, 87.9%, 95.6%, and 92.4% is obtained with a decision tree, k-neared neighbor, support vector machine, linear discriminant analysis, bagged tree and softmax classifier respectively

The methods proposed until now by the researchers have been limited in terms of performance. Hence an effective and robust method is required to classify environmental signals accurately. In the present work, authors aim to propose a method in which the dimension of data is reduced by OAS. The reduced data are then used to be transformed into images by STFT. Several features have been extracted from the spectrograms by using two pre-trained CNNs.

Classes from dataset

STFT, Sample Size,

Zhang X. et. al

Dilated Convolution Neural Network with LeakyReLU for Environmental Sound Classification

Environmental sound classification ; Dilated Convolution Neural Network; Leaky Rectified Linear Unit; Activation Function

Environmental sound classification

a dilated CNN-based ESC (D-CNN-ESC)

transforming acoustic waves to low level feature vectors following commonly used method

UrbanSound8K, ESC50, and CICESE

Environment

proposed D�CNN-ESC system outperforms the state-of-the-art ESC results obtained by very deep CNN-ESC system on UrbanSound8K dataset, the absolute error of our method is about 10% less than that of compared method.

All classes in the 3 datasets

Han B. & Hwang E.

ENVIRONMENTAL SOUND CLASSIFICATION BASED ON FEATURE COLLABORATION

Environmental sound recognition, discrete chirplet transform, discrete curvelet transform, discrete Hilbert transform, feature extraction

Environmental sound classification

We then applied equal-loudness level contours to each frame, to ensure that the signal more accurately represented human sound perception, and we eliminated the silence signal from the start and end points of each frame.

For traditional features, we collected mel-frequency cepstal coefficients (MFCC), zero-crossing rate (ZCR), spectral centroid (SC), spectral spread (SS), spectral flatness (SF), and spectral flux (SFX).

Environment

CDFs and ATFs are more effective than TFs for classification. Furthermore, when combined with TFs, they achieved the maximum accuracy.

three types of features: traditional features (TFs), change detection features (CDFs), and acoustic texture features (ATFs)

Street, road, talking, raining,bar, car

Hilbert transform, discrete chirplet transform

ML, Feature extraction

Wang J. et. al

Environmental Sound Classification Using Hybrid SVM/KNN Classifier and MPEG-7 Audio Low-Level Descriptor

Environmental sound classification

Hybrid SVM/KNN

Audio Spectrum Centroid, Audio Spectrum Spread, Audio Spectrum Flatness

Environment

the proposed hybrid SVM/KNN classifier outperforms the HMM classifier in MPEG-7 sound recognition tool

male speech (50), female speech (50), cough (50), laughing (49), screaming (26), dog barking (50), cat mewing (45), frog wailing (50), piano (40), glass breaking (34), gun shooting (33), and knocking (50). There are totally 527 sound files in our database

SVM, KNN, Feature extraction (Audio Spectrum Centroid, spread and flatness)

ENVIRONMENTAL SOUND CLASSIFICATION WITH CONVOLUTIONAL NEURAL NETWORKS

environmental sound, convolutional neu�ral networks, classification

Environmental sound classification

ESC-50 and ESC-10, UrbanSound 8k

Environment

publicly available datasets of environmen�tal recordings are still very limited - both in number and in size1

UrbanSound8K dataset (LP - 73.1%, US - 73.7%)

Classes from dataset

Salamon J. & Belo J.

Deep Convolutional Neural Networks and Data Augmentation for Environmental Sound Classification

Environmental sound classification, deep convo�lutional neural networks, deep learning, urban sound datase

Environmental sound classification

Environment

Wang J. et. al

Gabor-Based Nonuniform Scale-Frequency Map for Environmental Sound Classification in Home Automation

Environmental sound classification, feature extraction, Gabor function, home automation, matching pursuit (MP), nonuniform scale-frequency map

Environmental sound classification

Gabor Dictionary Based on Critical Frequency Bands
Nonuniform Scale-Frequency Map
Dimensional Reduction of Scale-Frequency Maps Using Principal
Component Analysis and Linear Discriminant Analysis

Environment

proposed feature is more appro�priate for practical use, especially environmental sound classification, since the proposed method has higher robustness against noise.

nonuniform scale frequency classifier

Nayak D. et. al.

Machine Learning Models for the Hearing Impairment Prediction in Workers Exposed to Complex Industrial Noise: A Pilot Study

Complex noise exposure, Hearing impairment, Machine learning, Noise-induced hearing loss

Hearing Impairment Prediction in Workers Exposed to Complex Industrial Noise

Environment

Data sets were collected from 1,644 workers exposed to complex noises in 53 workshops of 17 factories in the Zhejiang province of China

78.6 and 80.1

Tokozume Y. & Harada T.

LEARNING ENVIRONMENTAL SOUNDS WITH END-TO-END CONVOLUTIONAL NEURAL NETWORK

Environmental sound classification, convolu�tional neural network, end-to-end system, feature learning

We refer to our CNN as EnvNet

Urbansound 8K, YorNoise

Environment

local binary pattern

ScienceDirect1

Nossier S. et. al.

Enhanced smart hearing aid using deep neural networks

SCIENCE DIRECT

Deep learning; Dropout; Noise of interest awareness; Smart hearing aid; Speech enhancement

Smart hearing aid

Hearing aid

ScienceDirect2

Abdoli et. al.

End-to-end environmental sound classification using a 1D convolutional neural network

SCIENCE DIRECT

Convolutional neural network Environmental sound classification Deep learning Gammatone filterbank

Environemntal sound classification using 1D CNN

Environment

Feature extraction, MSE

ScienceDirect3

Mushtaq Z & Su S.F

Environmental sound classification using a regularized deep convolutional neural network with data augmentation

SCIENCE DIRECT

Data augmentation Environmental sound classification Regularization Deep convolutional neural network Urbansound8k

Environmental Sound Classification

Environment

ScienceDirect4

Chen Y. et. al.

Environmental sound classification with dilated convolutions

SCIENCE DIRECT

Sound information retrieval Environmental sound classification Dilated convolutions

Sound signal retrieval

Sound retrieval

ReLU, Softmax value, cross entropy, CNN

ScienceDirect6

Demir F et. al.

A new pyramidal concatenated CNN approach for environmental sound classification

SCIENCE DIRECT

Sound classification Deep learning SVM STFT CNN

Environment sound classification

Deep learning CNN

Short Time Frontier Transform

VCGNet 16, VCGNet 19, DenseNet 201

Urbansound 8K, ESC - 10, ESC -50

Environment

94.8, 81.4, 78.1

Short Time Fourier Transform (STFT)

ScienceDirect7

Mushtaq Z et. al.

Spectral images based environmental sound classification using CNN with meaningful data augmentation

SCIENCE DIRECT

Environmental sound classification Convolutional neural network Spectrogram Data augmentation Transfer learning

pproach of spectral images based on environmental sound classification using Convolutional Neural Networks (CNN) with meaningful data augmentation

ESC -10, ESC -50, Urbansound 8K

Environment

99.04, 99.49, 97.57

ML, Transfer Learning

ScienceDirect8

Ahmad S et. al.

Environmental sound classification using optimum allocation sampling based empirical mode decomposition

SCIENCE DIRECT

Environmental sound classification Optimum allocation sampling Empirical mode decomposition Multi-class least squares support vector machine Extreme learning machine

Automatic environmet sound classification

Optimum allocation sampling

ESC - 10

Environment

87.25, 77.61

dog bark, rain, sea waves, baby cry, clock tick, person sneeze, helicopter, chainsaw, rooster, and fire crackling

Empirical Mode Decomposition , Feature extraction (Approximate Entropy, Permutation Entropy, Log-energy entropy, Zero Crossing Rate) SVM, NN

ScienceDirect9

Mushtaq Z. & Su S.

Environmental sound classification using a regularized deep convolutional neural network with data augmentation

SCIENCE DIRECT

Data augmentation Environmental sound classification Regularization Deep convolutional neural network Urbansound8k ESC-10 ESC-50

Mel-spectrogram (Mel), Mel-Frequency Cep�stral Coefficient (MFCC) and Log-Mel by using DCNN

ESC-10 ESC-50 US8K

94.9 89.2 95.3

Springer1

Medhat F. et. al.

Masked Conditional Neural Networks for Environmental Sound Classification

Conditional Neural Networks � CLNN � Masked Conditional Neural Networks � MCLNN � Restricted Boltzmann Machine, RBM � Conditional Restricted Boltzmann Machine � CRBM � Deep Belief Nets � Environmental Sound Recognition � ESR � YorNois

Environmental sound classification

Conditional Neural Network

Urbansound 8K, YorNoise

Environment

air conditioner, car horns, children playing, dog bark, drilling, engine idling, gunshot, jackhammers, siren and street music

CNN, Feature extraction

Springer2

Zhang Z. et.al.

Deep Convolutional Neural Network with Mixup for Environmental Sound Classification

Environmental sound classification Convolutional neural network · Mixup

ESC-10 dataset is a subset of 10 classes (400 samples), UrbanSound8K dataset is a collection of 8732 short (up to 4 s) audio
clips of urban sound areas

Environment

91.7 83.9 83.7

dog bark, rain, sea waves, baby cry, clock tick, person sneeze, helicopter, chainsaw, rooster, fire crackling,air conditioner, car horn, children playing,
dog bark, drilling, engine idling, gun shot, jackhammer, siren, and street music

we propose a novel CNN as our ESC system model inspired by VGG Net, . In order to achieve a better performance for
our system on ESC, the effect of mixup hyper-parameter α is further explored.
Figure 5 shows the change of accuracy with different α ranging from [0.1, 0.5].
We see that when α = 0.2, the best accuracy is achieved on all three datasets.

Generating training data

INTERSPEECH1

Sailor B. et. al.

Unsupervised Filterbank Learning Using Convolutional Restricted Boltzmann Machine for Environmental Sound Classification

INTERSPEECH

Unsupervised Filterbank Learning, ConvRBM, Sound Classification, CNN

Environmental sound classification

supervised Convolutional Neural Network (CNN)

ESC -50 dataset

Environment

Sound/Audio Signal

proposed ConvRBM-BANK outperform EnvNET [18] even without the system combination. this shows the significance of unsupervised generative training using ConvRBM

ConvRBM-BANK performs significantly better than CNN with FBEs

INTERSPEECH2

Sharma J. et. al.

Environment Sound Classification using Multiple Feature Channels and Attention based Deep Convolutional Neural Networ

INTERSPEECH

Convolutional Neural Networks, Attention, Multiple Feature Channels, Environment Sound Classification

Environmental sound classification

Mel-Frequency Cepstral Coeffi�cients (MFCC), Gammatone Frequency Cepstral Coefficients (GFCC), Constant Q-transform (CQT) and Chromagram

ESC-10 ESC-50 US8K

Environment

94.75(ESC-10) 87.45(ESC-50) 97.52(US8k)

We stop at 128 features, which pro�duces the best results, to avoid increasing the complexity of the model.

eprint aRXIV

Mohaimenuzzaman Md. et. al.

Environmental Sound Classification on the Edge: A Pipeline for Deep Acoustic Networks on Extremely Resource-Constrained Devices

eprint aRXIV

Deep Learning, Audio Classification, Environmental Sound Classification, Acoustics, Intelligent Sound Recognition, Micro-Controller, IoT, Edge-AI, ESC-50

ACDNet, which produces above state-of-the-art accuracy on ESC-10 (96.65%) and ESC-50 (87.1%), we describe the com�pression pipeline and show that it allows us to achieve 97.22% size reduction and 97.28%

ACDNet is implemented in PyTorch version 1.7.1 and Wavio audio library is used to process the audio files. ESC-10 ESC-50 US8K

Environment

While limitations of the programming environment have restricted the accuracy of our current test deployment on a physical MCU, we have conclusively shown that 81.5% accuracy is achievable on such a resource-impoverished device, close to the state-of-the-art and above human performance.

it is likely that the performance can be improved by using quantization-aware training and pruning. Secondly, we would like to try the SpArSe approach for further optimisations now that we have developed Micro-ACDNet as a suitable starting point for its optimisations.

ACDNet, which produces above state-of-the-art accuracy on ESC-10 (96.65%) and ESC-50 (87.1%), we describe the com�pression pipeline and show that it allows us to achieve 97.22% size reduction and 97.28%

Training sample, Learning rate, prunning process

Hybrid prunning

Dempter-Shafer Evidence Theory

crossover fitting, user preference matching, populating diversity

Presentation, application, storage

Speech Enhancement, CNN

soundscaping, source mixing and source, modeling, STFT, posterior distribution

soundscaping, source mixing and source modeling

acoustic feedback signal, microphone, feedback and error signals, computational complexity

Hearing aid sructure/curcuit

Sampling frequency, Speech Quality Perception Evaluation

ML, Feature extraction, Evaluation process

denoising autoencoder networks