H-Semantics: A Hybrid Approach to Singing Voice Separation
Stratis Sofianos, Aladdin M. Ariyaeeinia, Richard Polfreman, Reza Sotudeh · ePrints Soton (University of Southampton) · 2012
Separating the singing voice from accompanying instruments is important in music information-retrieval systems, since it allows for such applications as melody extraction, lyrics recognition, and singer identity. The authors investigate effective methods for unsupervised separation of the singing voice, called H-Semantics (Hybrid Singing Extraction through Multiband Amplitude Enhanced Thresholding and Independent Component Subtraction). The proposed method adds time-domain separation to the previous work that was based on frequency-domain cepstral methods. The results indicate separation of approximately 8.5 dB signal-to-distortion ratio over the baseline