Multi-Style Reverberation Models and Efficient Model Adaptation for Robust Distant-Talking Speech Recognition with REMOS

Inhalt:
To further increase the flexibility of the REMOS (REverberation MOdeling for Speech recognition) concept for distant-talking speech recognition, multi-style reverberation models (RVMs) trained on data from different rooms as well as simplified RVMs are analyzed in this contribution. If the multi-style probability density functions (pdfs) used for score calculation are adapted to the reverberation conditions of the current room, a remarkable improvement in recognition performance can be achieved. Evaluations of a very efficient reverberation model adaptation scheme by connected digit recognition experiments show that REMOS can be adjusted to different reverberation conditions with minimal effort.