Coupling acoustic and language models in speech recognition in noisy conditions

No Thumbnail Available

URL

Journal Title

Journal ISSN

Volume Title

School of Science | Master's thesis
Checking the digitized thesis and permission for publishing
Instructions for the author

Date

2012

Major/Subject

Informaatiotekniikka

Mcode

T-61

Degree programme

Language

en

Pages

[7] + 45

Series

Abstract

Automatic speech recognition systems have difficulties with adapting to different speakers and acoustic environments, unlike humans that can overcome them without great difficulties. This represents an obstacle in developing real world applications where the speech recognition system is used in varying noisy environments. This work addresses the problem of automatic speech recognition in noisy conditions, focusing on coupling acoustic and language models to achieve better speech recognition results. Specifically, the letter error rates and word error rates are compared for different coupling values of the acoustic and language models in the same data set with different values of SNR artificially added noise. Results of these experiments show a near linear relationship between SNR in dB and the factor that adjusts language model logarithmic likelihood. An approach based on spectral subtraction is presented to estimate the optimal scaling factor based on mapping an estimation of the SNR to the language model scale. The results obtained show better recognition rates than the baseline fixed o to a constant LM scale value. Performance near to optimal LM scale on average for each SNR is achieved by the given approach. However, the results are still far from the ones obtained from the best utterance wise LM scale.

Description

Supervisor

Kurimo, Mikko

Thesis advisor

Palomäki, Kalle
Kallasjoki, Heikki

Keywords

automatic speech recognition, acoustic modelling, language modelling, SNR estimation, spectral substraction

Other note

Citation