Recent advances in the multi-stream HMM/ANN hybrid approach to noise robust ASR

摘要: Abstract In this article we review several successful extensions to the standard hidden-Markov-model/artificial neural network (HMM/ANN) hybrid, which have recently made important contributions field of noise robust automatic speech recognition. The first extension hybrid was “multi-band hybrid”, in a separate ANN is trained on each frequency sub-band, followed by some form weighted combination state posterior probability outputs prior decoding. However, due inaccurate assumption sub-band independence, system usually gives degraded performance, except case narrow-band noise. All systems overcome independence and give improved performance noise, while also improving or not significantly degrading with clean speech. “all-combinations multi-band” trains for combination. This, however, typically requires large number ANNs. multi-stream” an expert every just small complementary data streams. Multiple posteriors using maximum a-posteriori (MAP) weighting rise further strategy hypothesis level MAP selection. An alternative exploiting classification capacity ANNs “tandem hybrid” approach one more classifiers are multi-condition generate discriminative features input ASR system. “multi-stream tandem feature streams, permitting multi-stream fusion. “narrow-band particularly narrow sub-bands. This robustness noises seen during training. Of presented, all provide generic models multi-modal Test results presented discussed.