| This research presents a new technique for adapting Hidden Markov Model (HMM) automatic speech recognition (ASR) systems when there is a mismatch between training and testing conditions. In standard ASR systems, each observation vector is assumed completely reliable, and implicitly weighted equally during decoding. The new method, referred to as the Uncertain Observation technique, leverages the knowledge that all feature extraction has an inherent amount of uncertainty. In noisy environments, for example, different regions in the time frequency plane have vastly different local signal to noise ratios, and thus significantly different reliabilities. The Uncertain Observation decoder derived for this research presents a strong probabilistic framework, integrating the likelihood distribution of each feature vector into the decoding process. This results in a continuously adaptive recognition system that integrates Bayes Predictive Classification into a HMM recognition system.; The technique compares favorably in stationary noise to model compensation methods such as Parallel Model Combination (PMC), and outperforms simpler feature transformation. In time-varying environments, for which the algorithm is designed, Uncertain Observation techniques outperform both model compensation and time-varying feature transformations. Uncertain Observation methods are further applied to packet lossy channels, compressed speech, and feature vector quantization, showing a promising method that can be integrated into distributed speech recognition systems for robust performance. Finally, Uncertain Observation methods are applied during training to aid speech model creation or adaptation when the available data contains a distortion. |
