An improved uncertainty decoding scheme with weighted samples for multi-channel DNN-HMM hybrid systems

Kellermann W, Hümmer C, Astudillo RF (2017)


Publication Language: English

Publication Type: Conference contribution, Conference Contribution

Publication year: 2017

Pages Range: 31-35

Event location: San Francisco US

ISBN: 978-1-5090-5925-6

DOI: 10.1109/HSCMA.2017.7895556

Abstract

In this paper, we improve a recently-proposed uncertainty decoding scheme for DNN-HMM (deep neural network - hidden Markov model) hybrid systems, which models acoustic features as random variables. This uncertainty decoding strategy averages DNN outputs produced by a finite set of feature samples to approximate the posterior likelihoods of the context-dependent HMM states. As main innovation, we propose a weighted (instead of arithmetic) DNN-output averaging based on a minimum classification error criterion and apply it to a new probabilistic distortion model for multi-channel front-end signal enhancement schemes. The experimental evaluation is performed on the 8-channel REVERB Challenge task using a DNN-HMM hybrid system with spatial filtering of the microphone signals. It is shown that the recognition accuracy of the DNN-HMM hybrid system improves by incorporating uncertainty decoding and that the proposed weighted DNN-output averaging further reduces the word error rate scores

Authors with CRIS profile

Involved external institutions

How to cite

APA:

Kellermann, W., Hümmer, C., & Astudillo, R.F. (2017). An improved uncertainty decoding scheme with weighted samples for multi-channel DNN-HMM hybrid systems. In Proceedings of the Joint Workshop on Hands-free Speech Communication and Microphone Arrays (HSCMA) (pp. 31-35). San Francisco, US.

MLA:

Kellermann, Walter, Christian Hümmer, and Ramón Fernández Astudillo. "An improved uncertainty decoding scheme with weighted samples for multi-channel DNN-HMM hybrid systems." Proceedings of the Joint Workshop on Hands-free Speech Communication and Microphone Arrays (HSCMA), San Francisco 2017. 31-35.

BibTeX: Download