Online Learning and Acoustic Feature Adaptation in Large-Margin Hidden Markov Models

Abstract

Abstract-We explore the use of sequential, mistake-driven updates for online learning and acoustic feature adaptation in large-margin hidden Markov models (HMMs). The updates are applied to the parameters of acoustic models after the decoding of individual training utterances. For large-margin training, the updates attempt to separate the log-likelihoods of correct and incorrect transcriptions by an amount proportional to their Hamming distance. For acoustic feature adaptation, the updates attempt to improve recognition by linearly transforming the features computed by the front end. We evaluate acoustic models trained in this way on the TIMIT speech database. We find that online updates for large-margin training not only converge faster than analogous batch optimizations, but also yield lower phone error rates than approaches that do not attempt to enforce a large margin. Finally, experimenting with different schemes for initialization and parameter-tying, we find that acoustic feature adaptation leads to further improvements beyond the already significant gains achieved by large-margin training.Index Terms-Acoustic feature adaptation, automatic speech recognition (ASR), discriminative training, hidden Markov models (HMMs), large-margin classification, online learning.

Robust hidden semi-Markov modeling of array CGH data

“…However, discriminative training is time consuming and we need to choose the learning rates [1]. Another way to consider the inaccuracy of models is to use robust methods [2].…”

“…Hidden Markov models (HMM) are generative classifiers. Discriminative training of HMMs has a long history, especially in the automatic speech recognition area [1]. Of these discriminative trained models, recently, the large-margin HMMs show better performance.…”

“…Of these discriminative trained models, recently, the large-margin HMMs show better performance. However, discriminative training of HMMs is more complicated than maximum likelihood estimation because discriminative training methods should separate correct hypothesis from incorrect ones [1]. EM algorithm can not be used for discriminative training and genetic optimization algorithms should be used.…”