Joint Uncertainty Decoding for Robust Large Vocabulary Speech Recognition
H. Liao and M.J.F. Gales
CUED/F-INFENG/TR-552. Nov 2006.

Standard techniques to increase automatic speech recognition noise
robustness typically assume recognition models are clean trained. This
"clean" training data may in fact not be clean at all, but may contain
channel variations, varying noise conditions, as well as different
speakers. Hence rather than considering noise robustness techniques as
compensating clean acoustic models for environmental noise, they may
be thought of as reducing the acoustic mismatch between training and
test conditions. This report examines the application of VTS model
compensation or model-based Joint uncertainty decoding to clean and
multistyle trained systems. An EM-based noise estimation procedure is
also presented to produce ML VTS or Joint noise models depending on
the form of compensation used. Alternatively, compared to multistyle
training, adaptive training with Joint uncertainty transforms, also
referred to as JAT in this work, provides a better method for handling
heterogeneous data.  With JAT, the uncertainty bias added to the model
variances de-weights observations proportional to the noise level. In
this way, Joint transforms normalise the noise from the data allowing
the canonical model to solely represent the underlying "clean"
acoustic signal. This report presents a novel Joint adaptive training
framework including formula for estimating the transforms and
canonical model parameters. Lastly, large vocabulary systems are often
trained on multistyle data sets such as broadcast news or
conversational telephone speech that have a variety of noise
conditions. However, to date not much research has been done on
compensating such systems built with non-artificially corrupted
data. In this report, experiments are conducted on an artificially
corrupted Resource Management database and the large vocabulary
Broadcast News corpus of collected broadcast recordings.