Fourth European Conference on Speech Communication and Technology

Madrid, Spain
September 18-21, 1995

Signal Conditioned Minimum Error Rate Training

Wu Chou, Mazin G. Rahim, Eric Buhrke

AT&T Bell Laboratories, Murray Hill, NJ, USA

In this paper, a new approach, signal conditioned minimum error training, is proposed, where signal conditioning and minimum string error rate training are integrated into one process. The signal conditioning in this approach is based on hierarchical signal bias removal (HSBR), a novel extension of the signal bias removal algorithm. The HSBR is applied in conjunction with minimum string error rate training. In contrast to using a fixed codebook, the HSBR codebook used in our approach is derived from HMM parameters and updated with the HMMs during the process of minimum error rate training. As such, both HSBR signal conditioning and string model based minimum error rate training are based on the same set of HMMs. Experiments are performed on a connected digit database collected from the telephone network. The database covers various analog and digital network channels, different regional areas and a variety of telephone handsets. It is found that the proposed approach of signal conditioned minimum error rate training can lead to a significant reduction in recognition error rate. Based on a sub-word model consisting of various inter-word and intra-word context dependent (head-body-tail) model units, a 47% word error rate reduction is obtained through the proposed approach comparing with the model obtained from the conventional maximum likelihood (ML) training. This corresponds to an additional 27% word error rate reduction comparing with the inter-word context dependent sub-word model obtained from minimum error rate training where signal conditioning is not incorporated.

Full Paper

Bibliographic reference.  Chou, Wu / Rahim, Mazin G. / Buhrke, Eric (1995): "Signal conditioned minimum error rate training", In EUROSPEECH-1995, 495-498.