Self-supervised speaker embeddings
This addresses the limitation of speaker embeddings for speech processing by enabling use of unlabeled data, though it is incremental as it builds on existing embedding methods.
The paper tackles the problem of speaker embeddings being unable to use unlabeled data by proposing a self-supervised training method that reconstructs speech frames using embeddings and phone sequences, resulting in notable improvements on VoxCeleb and Speakers in the Wild datasets.
Contrary to i-vectors, speaker embeddings such as x-vectors are incapable of leveraging unlabelled utterances, due to the classification loss over training speakers. In this paper, we explore an alternative training strategy to enable the use of unlabelled utterances in training. We propose to train speaker embedding extractors via reconstructing the frames of a target speech segment, given the inferred embedding of another speech segment of the same utterance. We do this by attaching to the standard speaker embedding extractor a decoder network, which we feed not merely with the speaker embedding, but also with the estimated phone sequence of the target frame sequence. The reconstruction loss can be used either as a single objective, or be combined with the standard speaker classification loss. In the latter case, it acts as a regularizer, encouraging generalizability to speakers unseen during training. In all cases, the proposed architectures are trained from scratch and in an end-to-end fashion. We demonstrate the benefits from the proposed approach on VoxCeleb and Speakers in the wild, and we report notable improvements over the baseline.