Histogram equalization for noise robust speech recognition

No Thumbnail Available

URL

Journal Title

Journal ISSN

Volume Title

Helsinki University of Technology | Diplomityö
Checking the digitized thesis and permission for publishing
Instructions for the author

Date

2009

Major/Subject

Informaatiotekniikka

Mcode

T-61

Degree programme

Tietotekniikan tutkinto-ohjelma

Language

en

Pages

(5+) 50

Series

Abstract

Automatic speech recognition (ASR) is a fascinating field of science where the machine almost becomes human. Being able to communicate naturally with a machine has been a dream for a long time. Today, the technology makes it possible for the machine to understand human speech. However the quality of recognition suffers a lot from surrounding noise, and noisy environments are our everyday life conditions. This work presents a technique to improve noise robustness of ASR systems based on histogram equalization. This method has been proven efficient in the field of image processing and here we show that it can he successfully applied to audio data too. The idea behind it is to equalize noisy data and make it "sound like" clean data so that ASR systems trained on clean speech can recognize noisy speech more accurately. Experiments are conducted on Helsinki University of Technology's ASR system, and show a significant improvement in large vocabulary continuous speech recognition of noisy data on Aurora 4 database.

Description

Supervisor

Oja, Erkki

Thesis advisor

Kurimo, Mikko

Keywords

speech recognition, histogram equalization, noise robustness

Other note

Citation