Robust i-vector extraction tightly coupled with voice activity detection using deep neural networks

This paper describes an extended framework of i-vector feature extraction for improving speaker recognition under noisy conditions. In such speaker recognition applications, voice activity detection (VAD) has been a pre-processing step independent from the succeeding i-vector extraction, just for excluding noisy sounds other than voice of a target speaker. In the proposed framework, i-vector extraction is tightly coupled with VAD. It first estimates voice quality in an utterance as frame- wise voice posteriors using deep neural network-based sound classification, and then it extracts an i-vector of the utterance on the basis of the voice posteriors. The proposed method is able to produce a reliable speaker feature under noisy conditions with help of LSTM-based soft decision VAD. Speaker verification experiments on NIST 2016 SRE evaluation set demonstrate a 14.7% reduction in equal error rates.