Speech Processing Group

Speech@FIT (Speech Processing Group at Faculty of Information Technology, Brno University of Technology - FIT BUT, Czech republic) was formed in 1997 at the Faculty of Electrical Engineering and Computer Science at BUT, and joined the Department of Computer Graphics and Multimedia of FIT at the creation of FIT in January 2002.

The group is advised by Prof Hermansky, who has a significant research and educational track and a broad industrial and academic experience (US-West, Panasonic, ICSI Berkeley, OGI Portland, IDIAP Switzerland). The executive director of Speech@FIT - Jan "Honza" Cernocky - has significant experience in speech and signal processing education, research and project management.

Speech@FIT has participated in EC-sponsored projects aiming at speech corpora collection – it was at the collection of the first large-scale Czech speech database (SpeechDat-E, FP4, No. 977017) and participated also in SpeeCon (IST-1999-10003), as well as other industry-funded speech-collection. As full research partners, we have participated in Multimodal Meeting Manager (M4, FP5, IST-2001-34485), and Augmented Multimodal Interfaces (AMI, FP6, 506811). Currently, we are partners of Content Analysis and REtrieval Technologies to Apply Knowledge Extraction to massive Recording (CareTaker, FP6, 027231) and Augmented Multimodal Interfaces with Distant Access (AMIDA, FP6, IST-033812). Speech@FIT is also participating in several locally funded research projects.

Main expertise of the group is in perceptually-based robust acoustic processing, speech recognition, keyword spotting and speaker and language identification. The best phoneme recognition results on TIMIT, and excellent results in NIST LRE 2005 and NIST SRE 2006 evaluations can be considered as its main achievement. Speech@FIT is also key member of AMI[DA] team developing large vocabulary recognizer for meetings and its STK, phnrec and SNet software are used in several labs worldwide. The group is also experienced in very low bit rate coding, automatic determination of speech units and large scale speech database collection and processing.

The group disposes of equipment to perform serious experiments in speech recognition: 6 IBM-Blade centers, with 66 servers, all running Linux, file servers with total capacity of 50 TeraBytes and speech and language databases.