Large margin techniques for parameter estimation in speech recognition

Continuous-density Hidden Markov models (CD HMMs) are widely used to model real-valued random variables, for example, in automatic speech recognition (for acoustic modeling of speech signals), optical character recognition and video sequences. In this project, we have proposed a large margin learning framework to estimate parameters of CD HMMs. Our approach demonstrated superior performance in standard speech recognition benchmarks to traditional approaches such as maximum likelihood estimation, conditional likelihood maximization, minimum classification errors.

Left. Use CD HMMs to model acoustic signals. Right. Use CD HMMs as classifiers for sequences.
Left. Contrast in training methods for CD HMMs. Large margin training criteria is more stringent than traditional criteria. Right. Superior performance of large margin CD HMMs over other approaches.

Related publications