Feature Space Generalized Variable Parameter HMMs for Noise Robust Speech Recognition

Handling variable ambient noise is a challenging task for au-tomatic speech recognition (ASR) systems. To address this is-sue, multi-style training using speech data collected in diverse noise environments, noise adaptive training or uncertainty de-coding techniques can be used. An alternative approach is to explicitly approximate the continuous trajectory of Gaussian component or model space linear transform parameters against the varying noise, for example, using generalized variable pa-rameter HMMs (GVP-HMM). In order to reduce the computa-tional cost of conventional GVP-HMMs when model parameter update against the varying noise condition is required, this pa-per investigates a novel and more efficient extension of GVP-HMMs that can also model the trajectories of feature space lin-ear transforms. Significant error rate reductions of 9.3 % and 18.5 % relative were obtained over the multi-style training base-line system on Aurora 2 and a medium vocabulary Mandarin Chinese speech recognition task respectively. Index Terms: feature transform trajectory, generalized variable parameter HMM, variable noise, robust speech recognitio