Graduate School and Research Center in Digital Sciences

Online non-negative convolutive pattern learning for speech signals

Wang, Dong; Vipperla, Ravichander; Evans, Nicholas; Zheng, Thomas Fang

Research Report RR-11-261

The unsupervised learning of spectro-temporal patterns within speech signals is of interest in a broad range of applications. Where patterns are non-negative and convolutive in nature, relevant learning algorithms include convolutive non-negative matrix factorization (CNMF) and its sparse alternative, convolutive non-negative sparse coding (CNSC). Both algorithms, however, place unrealistic demands on computing power and memory which prohibit their application in large scale tasks. This paper proposes a new online implementation of CNMF and CNSC which processes input data piece-by-piece and updates learned patterns gradually with accumulated statistics. The proposed approach facilitates pattern learning with huge volumes of training data that are beyond the capability of existing alternatives. We show that the new online learning algorithm almost surely converges to the same cost value as the standard batch learning approach when both computing resources and data are unlimited and that it outperform batch learning in two experiments with practical computing resources and data quantities.

Document Bibtex

Title:Online non-negative convolutive pattern learning for speech signals
Keywords:non-negative matrix factorization, convolutive NMF, online pattern learning, sparse coding, speech processing, speech recognition
Type:Report
Language:English
City:
Date:
Department:Digital Security
Eurecom ref:3576
Copyright: © EURECOM. Personal use of this material is permitted. The definitive version of this paper was published in Research Report RR-11-261 and is available at :
Bibtex: @techreport{EURECOM+3576, year = {2011}, title = {{O}nline non-negative convolutive pattern learning for speech signals}, author = {{W}ang, {D}ong and {V}ipperla, {R}avichander and {E}vans, {N}icholas and {Z}heng, {T}homas {F}ang}, number = {EURECOM+3576}, month = {12}, institution = {Eurecom}, url = {http://www.eurecom.fr/publication/3576},, }
See also: