Research on Adaptive Step Decoding in Segment-based LVCSR

In this paper, a novel adaptive step decoding method using steady-energy pieces (SEPs) is explored in segment model (SM) based LVCSR system. Using speech analysis methods and statistical classification tools, the start and end points of SEPs are detected firstly. In SM decoding stage, frame-by-frame decoding for segments which start or end in SEPs are overleaped, replaced by SEP-based decoding. In our primary study, the SM-based LVCSR system reduced 20.55% of its runtime using our adaptive step decoding, with no increase on WER. The adaptive step decoding using pre-analysis of speech is proved to be a potential way to improve LVCSR system's performance.