A New HMM Adaptation Approach for the Case of a Hands-free Speech Input in Reverberant Rooms

A new method is presented for adapting the HMMs of a speech recognition system to the condition of a hands-free speech input in a room environment. The reverberation in a room usually has a bad effect on the performance of a recognition system. Reverberation causes an artificial extension of acoustic excitations what gets visible as so called reverberation tail when looking at the envelope of the short-term energy over the whole frequency range or in subbands. The approach is based on the assumption that the acoustic excitation of a speech segment, as modeled by an HMM state, will be seen as attenuated versions at successive HMM states. Adding this attenuated excitations in the spectral domain at each HMM state leads to a considerable improvement of the recognition performance. Furthermore a new approach is presented to adapt the Delta parameters that are usually taken as additional acoustic features. The efficiency of both new techniques has been proved by some experiments on isolated and connected word recognition with the TIDigits speech data base.