Hands-free speech recognition using blind source separation post-processed by two-stage spectral subtraction

This paper proposes hands-free speech recognition using blind source separation (BSS) post-processed by two-stage spectral subtraction (2S-SS). The BSS using independent component analysis (ICA) estimates a target signal and jammer signals. The 2S-SS removes its residual crosstalk components and suppresses spatially-distributed noise not separated by BSS. In large vocabulary continuous speech recognition (LVCSR) evaluation, utterances by other speakers and computer-room noise were used as a jammer signal and a spatially-distributed noise source, respectively. In all noisy environments, it was confirmed that the proposed method outperformed the BSS with single-channel spectral subtraction (1SS).