Abstract
Penalized logistic regression machines (PLRMs) have recently been shown to give good performance on isolated word speech re cognition. In this paper, we extend this framework to continuous speech recognition. We present two approaches that both make use of the output from an HMM Viterbi recognizer. The first approach performs probabilistic prediction with PLRM on the segments obtained from the HMM Viterbi recognizer. The resulting subwords and subword probabilities are combined to form a sentence and a sentence probability, respectively. In the second approach, an N-best list generated by the HMM Viterbi recognizer is rescored using PLRM. Experiments on the Aurora2 connected digits database show that both approaches outperform the baseline HMM Viterbi recognizer