HMM ADAPTATION FOR TELEPHONE APPLICATIONS

The presence of background noise and the frequency response of a transmission line like in telephone applications have a major influence on the performance of speech recognition systems. An approach is presented in this paper to cope with both effects. It is based on an estimation of the stationary noise spectrum and an estimation of the mismatch between the frequency responses present during training and during recognition. These estimations are used in combination with the PMC scheme [1] to adapt the whole word HMMs for a speaker independent recognition of connected words. A considerable improvement can be achieved on recognizing distorted speech data. The technique is also used as part of a complete speech dialogue system over the telephone network where it could also proof its beneficial usability