Abstract:
Fiber Bragg grating (FBG)-based throat microphone's superior background noise suppression makes them ideal for wearable automatic speech recognition (ASR) devices. However, achieving naturalness and intelligibility remains challenging due to the low-pass filtering effects of tissue and bones. This study presents a deep learning framework using a long short-term memory (LSTM) recurrent neural network for speech enhancement in FBG microphones. Also, it explores the impact of microphone placement and sex on ASR performance. The microphone, designed with a 1530.12 nm prestrained FBG, captured vocal vibrations from six participants reciting Harvard sentences. An LSTM model trained with spectral mapping restored high-frequency components, improving the non-intrusive short-time objective intelligibility (NI-STOI) score by 2%. Character error rate (CER) and NI-STOI score showed significantly better performance at the lower throat position, emphasizing the importance of optimal microphone placement. Speaker sex, however, had no significant effect on CER or intelligibility.