Partenaires et organisations internationales
(Anglais)
|
Sheffield University (UK), Ruhr-Universiät Bochum (D), Daimler-Benz (D), Institut National Polytechnique de Grenoble (F), University of Keele (UK), University of Patras (UK), IDIAP
|
Résumé des résultats (Abstract)
(Anglais)
|
The goals of this research network were to achieve better understanding of auditory processing and to deploy this understanding in automatic speech recognition in noisy environments, typically for car applications (with Daimler-Chrysler as the main industrial partner). This project covered several themes, including computational auditory scene analysis, sound-source segregation and new recognition techniques based on multi-band and multi-stream processing (the latter theme being mainly covered by IDIAP).
In the framework of the SPHEAR project, IDIAP mainly contributed to the investigation, development and testing of advanced speech recognition techniques based on different multi-band and multi-stream speech recognition paradigms, also involving different stream reliability measures, as well as supervised/unsupervised, and offline/online, adaptation schemes.
As opposed to other noise robust approaches, the proposed multi-channel approaches do not require any estimation of the noise spectrum or any noise model, which is particularly difficult in the case of unknown noisy conditions or non-stationary noise (often appearing in real life conditions). As a matter of fact, numerous experiments (including international evaluations) largely demonstrated the superiority of the multi-channel approach, especially in the case of non-stationary noise. This multi-channel approach is now recognized as part of the stat-of-the art in robust speech recognition.
Finally, in the framework of this project, and as part of the TMR constraints, two foreign PhD students (one French and on German) graduated at IDIAP with a PhD.
|
