A comparative study for throat microphone speech enhancement with different approaches

Throat microphones (TM) offer significant advantages in noisy environments by capturing speech signals directly from the throat, thus minimizing external noise. However, TM signals often lack clarity and intelligibility compared to conventional microphones. This paper presents a comparative study of three prominent feature extraction techniques—Mel-frequency cepstral coefficients (MFCC), Linear Predictive Cepstral coefficients (LPCC), Perceptual Linear Prediction (PLP) for enhancing speech captured by throat microphones. Each technique is evaluated based on its ability to enhance speech clarity and reduce noise interference. Experimental results on the ATR503 dataset, consisting of throat and close-talk microphone recordings, reveal that LPCC achieved an average Signal-to-Noise Ratio (SNR) improvement of 3dB and a Perceptual Evaluation of Speech Quality (PESQ) score increase of 1.3133 and 0.9553 compared to MFCC and PLP. In subjective evaluations the highest mean rating of 8.46 for LPCC indicates it was perceived as the most intelligible and clear. LPC spectra analysis demonstrates that Linear Predictive Cepstral Coefficients (LPCC) in retrieving missing frequencies in speech captured by throat microphones. These findings suggest that LPCC is a robust method for throat microphone speech enhancement, offering significant improvements in speech intelligibility and quality in noisy environments.