Investigation of an Encoder-Decoder LSTM model on the enhancement of speech intelligibility in noise for hearing impaired listeners

摘要

Hearing impaired (HI) listeners often struggle to follow conversations when exposed in a complex acoustic environment. This is partly due to the reduced ability in recovering the target speech Temporal Envelope (ENV) cues from Temporal Fine Structure (TFS). This study investigates the enhancement of speech intelligibility in HI listeners, by processing the ENV of speech signals corrupted by real-world environmental noise. An Encoder-Decoder Long Short Term Memory (LSTM) model is exploited after perceptually motivated processing stages to compensate for the important ENV characteristics of comprehensible speech for hearing impairment. The computational model is evaluated using the Short-Time Objective Intelligibility (STOI) measure for speech intelligibility. Finally, results indicate a 6% improvement in the mean STOI measure across different SNR values.