A comparative study of fMPE and RDLT approaches to LVCSR

Abstract

This paper presents a comparative study of two discriminatively trained feature transform approaches, namely feature-space minimum phone error (fMPE) and region-dependent linear transform (RDLT), to large vocabulary continuous speech recognition (LVCSR). Experiments are performed on an LVCSR task of conversational telephone speech transcription using about 2,000 hours training data. Starting from a maximum likelihood (ML) trained GMM-HMM based baseline system, recognition accuracy and run-time efficiency of different variants of the above two methods are evaluated, and a specific RDLT approach is identified and recommended for deployment in LVCSR applications.