Effect of Reverberation on Phonemes in Automatic Speech Recognition under Reverberant Environment

Automatic speech recognition (ASR) technology needs further performance enhancement for its applications in various reverberant environments. To provide more tools to enhance ASR performance in reverberant environments, effect of reverberation on phonemes in ASR is quantitatively analyzed in this work. Mean phoneme coherence (MPC) is presented to quantify the amount of spectral contamination on phonemes in reverberated speeches. MPC effectively represents the amount of reverberation's effect on phonemes which depends on not only reverberant condition but also phoneme arrangement in reverberated speeches. By assessing reverberated speeches with a speech recognizer, words in reverberated speeches are classified into correctly recognized words and wrongly recognized words. Then MCP distributions are compiled separately for each phoneme in correctly recognized words and wrongly recognized words. By comparing difference in MPC distributions in for each phoneme in correctly and wrongly recognized words, we identified which phonemes are more and less affected by reverberation on ASR.