Speaker recognition system of flexible throat microphone using contrastive learning

Recently, Flexible pressure sensor-based Throat Microphones (FTM) have attracted more attention in noise-robust speaker recognition and are promising for helping people with specific dysarthria to complete speaker recognition. FTM has outstanding flexibility compared with Hard Throat Microphones (HTM) and noise-robustness compared with Close-talk microphones (CM). However, speaker recognition for FTM is still an open task awaiting exploration since FTM has degradation problems and a lack of data sets. To tackle these two obstacles, referring to feature mapping methods for HTM, we introduce an FTM-oriented supervised contrastive learning (FTMSCL) method. An FTM speech data set is collected, then a contrastive loss function is designed to avoid the feature mapping methods' problems and effectively leverage label information from this data set. Furthermore, a critical parameter margin in this loss and several data augmentations for FTM are investigated. Experimental results show that, with no need for CM data, FTMSCL can achieve a False Acceptance Rate (FAR) of 2.97% and a False Rejection Rate (FRR) of 2.83%, which outperforms a conventional End-to-End one and an advanced feature mapping one significantly. Moreover, the best FAR and FRR of our FTMSCL method are only 0.86% and 0.83% higher than the best one using clean CM data.