Abstract
It is challenging to perform automatic speech recognition when multiple people talk simultaneously. To solve this problem, speaker-aware selective methods have been proposed to extract the speech of the target speaker, relying on the auxiliary speaker characteristics provided by an anchor (a clean audio sample of the target speaker). However, the extraction performance depends on the duration and quality of the anchors, which is unstable. To address this limitation, the authors propose a target speaker extraction network (TEnet) which applies the robust speaker embedding to extract the target speech from the speech mixture. To get more stable speaker characteristics during training, the robust speaker embeddings are accumulated over all the speech of each target speaker, rather than utilising the embedding produced by a single anchor. As for testing, very few anchors are enough to get decent extraction performance. Results show the TEnet trained with accumulated embedding achieves better performance and robustness compared with the single-anchored TEnet. Moreover, to exploit the potential of the speaker embedding, the authors propose to feed the extracted target speech as anchor and train a feedback TEnet, whose results are superior to the short-anchored baseline for 22.5% on word error rate and 15.5% on signal-to-distortion rate.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Disclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.