计算机科学
声音(地理)
人工智能
深度学习
语音识别
地质学
海洋学
作者
Gandra Shreya,Dipanshu Bhardwaj,Prashant A. Athavale
出处
期刊:
[Institute of Electrical and Electronics Engineers]
日期:2024-08-09
卷期号:: 1-6
标识
DOI:10.1109/nmitcon62075.2024.10699074
摘要
Human brains constantly interpret sound impulses to offer pertinent contextual information, so humans can easily recognise various environmental noises. For this purpose, many machine-learning and deep-learning techniques like artificial neural networks (ANNs), and convolutional neural networks (CNNs) have been developed specifically for training on spectral features. The present research uses Mel-frequency cepstral coefficients (MFCC), chromagram and Log-mel spectrograms to train the deep CNN model. Data augmentation is performed to address data shortages and enhance accuracy with unseen data. Using several combination augmentation strategies, current research achieved an accuracy of ${9 9. 0 2 \%}$ by extracting log-mel spectrograms and training them on a 6-layer CNN architecture with a Leaky ReLU activation function. Out of all the suggested models, the proposed model performs the best and is most similar to pre-trained models like ResNet and DenseNet, which need millions of parameters to train, compared to the suggested model’s less than a million parameters. Features in one and two dimensions are generated for CNN1d, CNN2d, and ANN models’ training.
科研通智能强力驱动
Strongly Powered by AbleSci AI