Castro Martínez, Angel Mario and Meyer, Bernd T. (2015) Mutual benefits of auditory spectro-temporal Gabor features and deep learning for the 3rd CHiME Challenge. In: CHiME Challenge. (Unpublished)

Restricted to Repository staff only bis 30 September 2016. Volltext (1091Kb)

Abstract

In this study, we combine our feature extraction scheme inspired by the human auditory system with a deep learning approach in the context of the 3rd CHiME Speech Separation and Recognition Challenge. The use of spectro-temporal Gabor features is motivated by physiological measurements in the primary auditory cortex that indicated neurons to be sensitive to localized, spectro-temporal patterns of an acoustic stimulus. We exploit such information by replacing the baseline filter bank features with features from a two-dimensional convolution of the log-Mel-spectrogram and the proposed Gabor filter bank. Together with some other minor parameter adjustments, we obtained relative improvements over the real data of 26% for the development set, and 34.24% for the evaluation set compared to the CHiME 3 baseline.

Item Type:	Conference or Workshop Item (Paper)
Uncontrolled Keywords:	speech recognition, auditory features, deep learning, discriminative training
Subjects:	Generalities, computers, information > Computer science, internet Language > English Science and mathematics > Physics
Divisions:	Faculty of Medicine and Health Sciences > Department of Medical Physics and Acoustics
Date Deposited:	24 Sep 2015 12:39
Last Modified:	24 Sep 2015 12:39
URI:	https://oops.uni-oldenburg.de/id/eprint/2509
URN:	urn:nbn:de:gbv:715-oops-25908
DOI:
Nutzungslizenz:

Actions (login required)

View Item