EFFICIENT AUDIO SOURCE SEPARATION USING MEL-SPECTROGRAMS

Abstract

Audio source separation deals with extracting a source of audio from a mixture, for example vocals from a musical recording. Recent strides have been made in the release of the Open-Unmix GitHub project in September of 2019 to provide new researchers with a framework to hit the ground running with state-of-the-art techniques. The base architecture uses a 3 layer bidirectional LSTM to complete a pixel-wise regression problem to estimate masks for each source'sspectrogram. The theses explores the idea of replacing the spectrograms in this process with mel-spectrograms which have achieved marginally better results in other audio problems such as speech recognition. A novel inverse function to convert from mel-spectrogram to spectrogram is provided that runs exponentially faster than the best available function with similar accuracy. We found that the results documented by the Open-Unmix project were reproducible and that the mel-spectrogram model did not provide an improvement.