Audio Strips Network (ASNet) and Amalgamated Audio Features (A2F): A Synergistic Approach for Audio Source Separation

Abstract

Audio source separation refers to the procedure of decomposing a mixed audio signal into its constituent components. This technique enables numerous applications, including creative music production, educational tools, karaoke, transcription, and music analysis. Despite the recent success of deep learning-based source separation techniques, these techniques often do not perform very accurately and do not provide high-quality separation of sources when many contain complex combinations in their mixtures. Source separation techniques generally rely on temporal or spectral features for analysis, which does not fully capture the complex dynamics of audio signals. To address these limitations, this paper proposes the amalgamated audio features (A2F), a hybrid representation combining temporal and spectral features. Then, the audio strip network (ASNet) is proposed, a novel framework designed to achieve clean and precise separation of individual audio sources with enhanced performance. ASNet utilizes A2F to separate sources more effectively. The model is trained and evaluated on the MUSDB, DSD100, and MUSDB18-HQ datasets, a benchmark for music source separation, and its standard measures, such as the signal-to-distortion ratio (SDR) and signal-to-interference ratio (SIR), are used to examine performance. ASNet achieves enhanced separation performance with SDR values of 12.63 for drums, 11.42 for vocal, 12.01 for bass, and 11.14 for other, and SIR values of 9.57 for drums, 9.61 for vocal, 9.66 for bass, and 9.67 for other. This advancement benefits musicians through high-quality remixing and creativity while aiding researchers in improving deep learning and hybrid audio processing models.