HCRF-based Model Compensation for Noisy Speech Recognition

Hidden conditional random fields (HCRFs) belong to a type of discriminative models for pattern classification. It is modified from conditional random fields framework and have been shown its advantages for acoustic modeling in speech recognition. This paper extends HCRF methodology to develop a robust technique for noisy speech recognition. We rearrange the linear chain structure of HCRF to its associated HMM and then take approximation of the Gaussian mixture models of the HMM with Taylor expansion. This makes it possible to obtain the proper relation in statistics between HCRF and HMM and then we propose a operative transformation for adapting the seed HCRFs to a set of noise matched HCRFs. This study addresses the following related issues: (1) how to implement the HCRFs-based compensation for noisy environment; (2) the integration of noise and channel bias compensation in HCRF frameworks; and (3) comparison of performance between HMM-based and HCRF-based noisy mixed-lingual (Mandarin and English) speech recognition. The experimental results indicate that proposed HCRF-based model compensation framework enjoys potential for development in robust speech recognition.