A Study on Combined Effects of Reverberation and Increased Vocal Effort on ASR

This study analyzes the individual and combined effect of room reverberation and increased vocal effort on automatic speech recognition. Robustness of several state-of-the-art front-end feature extraction strategies and normalizations to these sources of speech signal variability is evaluated in the context of large and small vocabulary recognition tasks on American English and Czech speech corpora. For the large vocabulary task, speech material from the UT-Scope database comprising American English utterances is used. The Czech speech samples are drawn from the CLSD‘05 data corpus and used for the small vocabulary tasks. Both databases contain neutral as well as increased vocal effort recordings. Simulated reverberant test conditions are generated using measured room impulse responses from the AIR database and utilized in the evaluations. It is shown that the robustness of a common MFCC front-end to reverberation and increased vocal effort can be considerably improved when paired with cepstral gain normalization and modified RASTA filtering. A combination of recently proposed mean Hilbert envelope coefficients and modified RASTA is found to provide balanced performance across all reverberation and vocal effort conditions.