LAUSR

Abstract Automatic sound recognition (ASR) is a remarkable field of research in recent years. The ability to automatically recognize sound events through computers in a complex audio environment is very useful for machine hearing, acoustic surveillance and multimedia retrieval applications. On the other hand, ASR task become highly difficult as the ambient noise levels increase and many traditional methods show very weak performance under noise. Recent studies has shown that spectrogram image features (SIF) have high performance under noise, while success rates in clean conditions are relatively lower than in the state-of-the-art approaches. In this study, after converting highly overlapped spectrograms into linear quantized images and reducing dimensions by applying various image resizing methods, feature extraction and classification are performed with convolutional neural networks (CNN), which have very high performance in image classification. In the mismatched case, the proposed method achieves a performance improvement of 4.5%, which is equivalent to a relative error reduction of 63.4%, with a classification success of 97.4%, while the multicondition training method achieves an average of 98.63% success rate.