Welcome to the ASR group homepage

[ FPMs > TCTS > ASR group ]

[Introduction]

[People]

[Publications]

[Contact]

Current research on speech recognition covers speaker-independent isolated word recognition, large vocabulary continuous speech recognition based on keyword spotting, speaker adaptation, and robust speech recognition using multi-band models (patent pending), in the framework of the THISL, RESPITE, SPRACH, and LKIT research projects.

Our twenty years of experience in this area is materialized in the STRUT Speech Recognition software toolkit, which includes hybrid HMM/ANN (Hidden Markov Models/Artificial Neural Networks) speech modeling, for which we are recognized as one of the major research units worldwide.

Applications: speech recognition systems for embedded applications (car navigation, consumer products), computer-assisted language learning, multimodal interfaces.

See "A short introduction to speech recognition" for further details about the ASR technology.

Check the speech recognition part of comp.speech for interesting links to ASR sites.

Research areas

Through several European and national projects, but also in collaboration with the MULTITEL ASBL research center, the ASR team addresses many aspects of the speech recognition problem:

• Hybrid HMM/ANN: combining hidden Markov models with artificial neural networks is a powerful alternative to classical stochastic models. This technique has been extensively studied and is now used as the baseline approach of our ASR systems. 

• Software engineering: over the last years we have built the STRUT toolkit for the fast development of ASR applications. Based on the plug and play programming philosophy, this software library implements many ASR related algorithms: signal processing, feature extraction, GMM, ANN, Viterbi decoding, state alignment, ...

• Robust speech recognition: through several projects (RESPITE, MODIVOC, AURORA) we have deeply investigated the problem of robust speech recognition: spectral subtraction, wiener filtering, noise estimation, multi-band, mixture of experts, missing data, microphone arrays, ...

• Keyword spotting: the problem of keyword spotting is crucial in real life application in order to partly handle spontaneous speech in man-machine dialogues. Next to keyword spotting, the problem of rejection of out of vocabulary words or poorly recognized word, has also been studied through the estimation of relevant confidence levels.

• Model adaptation: fast model adaptation allows to update the statistical models with very few data. Noise adaptation and speaker adaptation have been studied, more particularly in the framework of the hybrid HMM/ANN systems.

• Automatic phonetization

• Microphone arrays

• Dialogue management

• Distributed speech recognition: adapting voice interface to mobile environments imposes many constraints on the hardware capacities. Distributing the speech recognition processing consists in achieving a part (as light as possible) of this processing on the mobile device, transmitting some internal data representation and performing the rest of the recognition on fixed servers. This approach allows to control the CPU load on the mobiles and the useful bandwidth while guaranteeing the ASR performance. This problem has been adressed through the MODIVOC and AURORA projects.

• Embedded speech recognition.

^ Top ^