Improving generalization of Monte Carlo dropout based DNN ensemble model for speech enhancement and results on real world, traffic noise

Nazreen, PM and Ramakrishnan, AG (2019) Improving generalization of Monte Carlo dropout based DNN ensemble model for speech enhancement and results on real world, traffic noise. In: 2019 IEEE 16th India Council International Conference (INDICON), 13-15 Dec. 2019, Rajkot, India, India.

PDF IND_COU_INT_CON_2019.pdf - Published Version Restricted to Registered users only Download (772kB) | Request a copy

Abstract

We propose a threshold-based algorithm to choose between model uncertainty-based and DNN-classifier-based selection of noise-specific DNN models for speech enhancement, using Monte Carlo dropout. This method tries to compensate for the poor performance of the former scheme on speech with seen noises compared to classifier-based scheme. We show some promising results on speech corrupted with a mixture of unseen noises and on time varying, non-stationary noises, affecting random segments of speech. We use TIMIT speech, NOISEX-92 noises, and real world, traffic noise recorded by us. Our algorithm performs well on real world, traffic noise from 10 down to -10 dB. © 2019 IEEE.

Item Type:	Conference Paper
Publication:	2019 IEEE 16th India Council International Conference, INDICON 2019 - Symposium Proceedings
Publisher:	Institute of Electrical and Electronics Engineers Inc.
Additional Information:	cited By 0; Conference of 16th IEEE India Council International Conference, INDICON 2019 ; Conference Date: 13 December 2019 Through 15 December 2019; Conference Code:158465
Keywords:	Deep neural networks; Monte Carlo methods; Noise pollution; Speech enhancement; Uncertainty analysis, Classifier based scheme; Ensemble modeling; Model uncertainties; Nonstationary noise; Poor performance; Random segments; Time varying; Traffic noise, Acoustic noise
Department/Centre:	Division of Electrical Sciences > Electrical Engineering
Date Deposited:	07 Sep 2020 07:10
Last Modified:	07 Sep 2020 07:10
URI:	http://eprints.iisc.ac.in/id/eprint/65254

Actions (login required)

View Item