Multi-stream acoustic model adaptation for noisy speech recognition

In this paper, a multi-stream-based model adaptation method is proposed for speech recognition in noisy or real environments. The proposed scheme comes from our experience about audio-visual model adaptation. At first, an acoustic feature vector is divided into several vectors (e.g. static, first-order and second-order dynamic vectors), namely streams. While adaptation, a stream performing relatively high recognition performance is updated for the stream only. Alternatively, a stream having less recognition power is adapted using all the streams that are superior to the stream. In order to evaluate the proposed technique, recognition experiments were conducted using every streams, and then adaptation experiments were also investigated for various types of combination of streams.