- Realtime speech recognition on a 200 MHz fixed point linux core processor for vocabularies running into thousands to tens of thousands of words, depending on grammar constraints.
- Zero performance degradation over corresponding floating point version (demonstrable).
- Transcriptions returned in romanized script, with additional wrapper to convert to native script.
- Language models and grammars can be externally specified (defaults are provided). Natural language accepted.
- Speaker independent or speaker configurable recognition.
- Intrinsic runtime environmental noise compensation (not user activated).
- Optimal performance for speech captured over closetalking and farfield microphones (can be integrated into mobile telephony applications).
- Versions available for several languages.
|
Released Sept 06
.
|