Recognition of Mobile and Rich Speech (MARS)
Motivation & Project Description
- New model training algorithms for distant speech
- Training using noise reduction algorithms and normalizing transforms
- Context clustering for room acoustics
- Non-native, multi-lingual and cross-lingual speech processing
- Meta data extraction on distant, telephone, and wideband speech
- ID, age, gender, emotion, channel, language, socio-economic status
- Online acoustic change detection
- Speaker detection, clustering and adaptation
- In-house ASR system as benchmark for external suppliers
Expected Outcome:
- Janus-based ASR modules using 16kHz English distant-speech AMs
- Janus-based Inspire recognizer
- Janus-based ASR for Ivistar info displays (with VCE)
- Meta-data extraction modules and integration with Janus Recognition Toolkit
Time Frame: 07/2007-12/2008
T-labs Team Members: Florian Metze
Students: Peter Bourgonje, Stefan Schaffer
Partners: Jitendra Ajmera
Funding by: Deutsche Telekom Laboratories