Keyword-spotting in noisy continuous speech using word pattern vector subabstraction and noise immunity learning

Abstract

Noise immunity learning, previously proposed by the authors (1991) for isolated word recognition in noisy environments, is extended to keyword spotting in noisy continuous speech. The powerful features of the noise immunity keyword-spotting method are keyword spotting based on the multiple similarity (MS) method for reliable keyword detection, noise immunity learning for greater robustness in recognition of spontaneous or noisy speech, and word pattern vector subabstraction to represent noisy keyword patterns from different viewpoints. Integrating the spotting results obtained by different kinds of subabstracted word pattern vectors significantly improved the performance of the keyword spotting. A system to spot 30 keywords currently runs in real-time on a workstation with two accelerators. The spotted keywords are fed into a keyword sequence LR parser for spontaneous speech understanding. >