Sorry, you need to enable JavaScript to visit this website.

facebooktwittermailshare

LOW-RANK AND SPARSE SOFT TARGETS TO LEARN BETTER ! DNN ACOUSTIC MODELS

Abstract: 

Conventional deep neural networks (DNN) for speech acoustic modeling rely on Gaussian mixture models (GMM) and hidden Markov model (HMM) to obtain binary class labels as the targets for DNN training. Subword classes in speech recognition systems correspond to context-dependent tied states or senones. The present work addresses some limitations of GMM-HMM senone alignments for DNN training. We hypothesize that the senone probabilities obtained from a DNN trained with binary labels can provide more accurate targets to learn better acoustic models. However, DNN outputs bear inaccuracies which are exhibited as high dimensional unstructured noise, whereas the informative components are structured and low dimensional. We exploit principal component analysis (PCA) and sparse coding to characterize the senone subspaces. Enhanced probabilities obtained from low-rank and sparse reconstructions are used as soft-targets for DNN acoustic modeling, that also enables training with untranscribed data. Experiments conducted on AMI corpus shows 4.6% relative reduction in word error rate.

up
0 users have voted:

Paper Details

Authors:
Pranay Dighe, Afsaneh Asaei, Hervé Bourlard
Submitted On:
7 March 2017 - 12:15pm
Short Link:
Type:
Poster
Event:
Presenter's Name:
Pranay Dighe
Paper Code:
3483
Document Year:
2017
Cite

Document Files

icassp_poster.pdf

(603 downloads)

Subscribe

[1] Pranay Dighe, Afsaneh Asaei, Hervé Bourlard, "LOW-RANK AND SPARSE SOFT TARGETS TO LEARN BETTER ! DNN ACOUSTIC MODELS", IEEE SigPort, 2017. [Online]. Available: http://sigport.org/1687. Accessed: Oct. 18, 2017.
@article{1687-17,
url = {http://sigport.org/1687},
author = {Pranay Dighe; Afsaneh Asaei; Hervé Bourlard },
publisher = {IEEE SigPort},
title = {LOW-RANK AND SPARSE SOFT TARGETS TO LEARN BETTER ! DNN ACOUSTIC MODELS},
year = {2017} }
TY - EJOUR
T1 - LOW-RANK AND SPARSE SOFT TARGETS TO LEARN BETTER ! DNN ACOUSTIC MODELS
AU - Pranay Dighe; Afsaneh Asaei; Hervé Bourlard
PY - 2017
PB - IEEE SigPort
UR - http://sigport.org/1687
ER -
Pranay Dighe, Afsaneh Asaei, Hervé Bourlard. (2017). LOW-RANK AND SPARSE SOFT TARGETS TO LEARN BETTER ! DNN ACOUSTIC MODELS. IEEE SigPort. http://sigport.org/1687
Pranay Dighe, Afsaneh Asaei, Hervé Bourlard, 2017. LOW-RANK AND SPARSE SOFT TARGETS TO LEARN BETTER ! DNN ACOUSTIC MODELS. Available at: http://sigport.org/1687.
Pranay Dighe, Afsaneh Asaei, Hervé Bourlard. (2017). "LOW-RANK AND SPARSE SOFT TARGETS TO LEARN BETTER ! DNN ACOUSTIC MODELS." Web.
1. Pranay Dighe, Afsaneh Asaei, Hervé Bourlard. LOW-RANK AND SPARSE SOFT TARGETS TO LEARN BETTER ! DNN ACOUSTIC MODELS [Internet]. IEEE SigPort; 2017. Available from : http://sigport.org/1687