Deep Learning for Lossless Audio Compression

Abstract

Audio and speech compression techniques are used to reduce the storage of these data in the required space and the transmission rate of these data in the communication and network systems. In this paper, the researchers exploit neural networks and artificial intelligence to compress audio signals. The researchers investigated compression ratios of 8, 4, 2, and 1 (no compression), and then chose the highest ratio of 8. The compromising choice is based on the best SNR of the recovered audio signal and the required time for implementation. The researchers tested 119 different audio files from the standard BBC audio library. The duration of these files is about 1000 seconds. The average SNR was 26.33 dB, and the mean square error was -52.58 dB. To reduce the running time, the epochs were 30, the hidden layers were 64 to 128, the quantization level was 1, the dimensions were 15 to 20, and each second of the input signal needed 100 seconds to be compressed. The input audio signal files were single-channel mono audio, and the stereo multi-channel audio files were reformatted to mono single-channel. According to the results, the proposal process accomplished good audio compression, while the other parameters were acceptable.

Keywords

• Deep learning
• Audio compression
• Variational autoencoders
• Compression ratio