The ESAT speech recognition system

The ESAT speech group is one of the few remaining labs that has a fully in-house developed state-of-the-art continuous speech, large vocabulary, speaker independent recognition system. The current system is the result of 20 years of research of development in speech recognition at K.U.Leuven. The first developments in the late 80's targeted a small footprint isolated word recognizer. Later on the focus was on noise robust recognition for in-car applications. Since the mid 90's the development has focused on large vocabulary continuous recognition. Apart from being our major tool for research, several versions of the system have found their way into commercial applications including the Microsoft AutoPC (via Lernout & Hauspie).

A full description of the 2006 system can be found here.

In order to stay in tune both with research and software developments there have regular major rewrites of the software core. In February 2006 the latest rewrite, and largest effort ever, was started in framework of the STEVIN project SPRAAK.

	Expertise

Much of the expertise of the ESAT Speech Group is linked directly or indirectly to the large vocabulary speech recognizer. Over the years a wide range of expertise has been gathered in the wider field of speech processing, including:

• Speech Recognition:
  • ASR Systems
  • ASR Architectures: Example Based Recognition, Mixed bottom-up top-down search
  • Noise Robust Recognition: Missing Data Techniques, Spectral subtraction
• Speech Corpora:
  • automatic segmentation and alignment from rough annotations
• Speech Enhancement & Coding:
  • Perceptual Audio Coding with TLS
  • Subspace based speech enhancement
  • Broadband Beamforming, Blind Signal Separation
• Applications:
  • Computer Aided Learning
  • Hearing Aids