|
Robustness
to noise in automatic speech recognition is essential for the
development of successful applications. Noise reduction techniques have
been applied with some success in the past, but there remains a large
performance gap between the best ASR implementations and human
recognition, especially when the noise is non-stationary. This project
tackles the noise robustness problem in ASR through missing data
techniques (MDT) by addressing important open R&D issues for
accuracy improvement and computational efficiency. Detectors of missing
data will make minimal assumptions on the noise, while incorporating
more knowledge about speech. The acoustic model in the recognizer's
back-end will be refined and its evaluation will be made faster through
algorithmic research. The developed algorithms will be integrated in
the result of the SPRAAK software (also a project of the NTU) and made
available through its distribution channels. This project addresses
three STEVIN priorities: 1) robustness of speech recognition, 2) tools
and data for the development of robust speech recognition, and 3)
confidence measures. In this project we will base our research on a
'real-life' test suite that contains test material from the Dutch
SpeechDat Car and Speecon databases.
|
