Variable frame rate analysis for automatic speech recognition
Zheng‐Hua Tan · Proceedings of SPIE, the International Society for Optical Engineering/Proceedings of SPIE · 2007
In this paper we investigate the use of variable frame rate (VFR) analysis in automatic speech recognition (ASR). First, we review VFR technique and analyze its behavior. It is experimentally shown that VFR improves ASR performance for signals with low signal-to-noise ratios since it generates improved acoustic models and substantially reduces insertion and substitution errors although it may increase deletion errors. It is also underlined that the match between the average frame rate and the number of hidden Markov model states is critical in implementing VFR. Secondly, we analyze an effective VFR method that uses a cumulative, weighted cepstral-distance criterion for frame selection and present a revision for it. Lastly, the revised VFR method is combined with spectral- and cepstral-domain enhancement methods including the minimum statistics noise estimation (MSNE) based spectral subtraction and the cepstral mean subtraction, variance normalization and ARMA filtering (MVA) process. Experiments on the Aurora 2 database justify that VFR is highly complementary to the enhancement methods. Enhancement of speech both facilitates the frame selection in VFR and provides de-noised speech for recognition.