Exploring single channel speech separation for short-time text-dependent speaker verification

The automatic speaker verification (ASV) has recently achieved great progress. However, the performance of ASV degrades significantly when the test speech is corrupted by interference speakers, especially when multi-talkers speak at the same time. Although the target speech extraction (TSE) has also attracted increasing attention in recent years, its TSE ability is constrained by the required pre-saved anchor speech examples of the target speaker. It becomes impossible to directly use existing TSE methods to extract the desired test speech in an ASV test trial, because the speaker identity of each test speech is unknown. Therefore, based on the state-of-the-art single channel speech separation technique—Conv-TasNet, this paper aims to design a test speech extraction mechanism for building short-time text-dependent speaker verification systems. Instead of providing a pre-saved anchor speech for each training or test speaker, we extract the desired test speech from a mixture by computing the pairwise dynamic time warping between each output of Conv-TasNet and the enrollment utterance of speaker model in each test trial in the ASV task. The acoustic domain mismatch between ASV and TSE training data, the behaviors of speech separation in different stages of ASV system building, such as, the voiceprint enrollment, test and PLDA backend are all investigated in detail. Experimental results show that the proposed test speech extraction mechanism in ASV brings significant relative improvements (36.3%) in overlapped multi-talker speaker verification, benefits can be found not only in ASV test stage, but also in target speaker modeling.