Registration for ICASSP is free of charge, but registration is required to view the videos. If you have not yet registered, please visit: https://cmsworkshops.com/ICASSP2020/Registration.asp.Access the full virtual conference by visiting: https://2020.ieeeicassp-virtual.org/attendee/login. Your username is your email address and your password is your confirmation number/registration ID.

You need an account to view media

Sign in to view media

Don't have an account? Please contact us to request an account.

Speech Processing
SPE-L11.5
Lecture
Speech Separation and Extraction I: Single Channel

MIXUP-BREAKDOWN: A CONSISTENCY TRAINING METHOD FOR IMPROVING GENERALIZATION OF SPEECH SEPARATION MODELS

Max W. Y. Lam

Date & Time

Thu, May 7, 2020

10:00 am – 12:00 pm

Location

On-Demand

Abstract

Deep-learning based speech separation models confront poor generalization problem that even the state-of-the-art models could abruptly fail when evaluating them in mismatch conditions. To address this problem, we propose an easy-to-implement yet effective consistency based semi-supervised learning (SSL) approach, namely Mixup-Breakdown training (MBT). It learns a teacher model to "breakdown" unlabeled inputs, and the estimated separations are interpolated to produce more useful pseudo "mixup" input-output pairs, on which the consistency regularization could apply for learning a student model. In our experiment, we evaluate MBT under various conditions with ascending degrees of mismatch, including unseen interfering speech, noise, and music, and compare MBT's generalization capability against state-of-the-art supervised learning and SSL approaches. The result indicates that MBT significantly outperforms several strong baselines with up to 13.77% relative SI-SNRi improvement. Moreover, MBT only adds negligible computational overhead to standard training schemes.


Presenter

Max W. Y. Lam

Tencent AI Lab
Sign in to join the conversationDon't have an account? Please contact us to request an account.
Sign in to view documentsDon't have an account? Please contact us to request an account.

Session Chair

Tomohiro Nakatani

NTT Corporation