Under Construction

Meeting Recorder: Portable Speech Recognition

Speech recognition for Meeting Recorder is challenging for a number of reasons:

It must work in a wide variety of acoustic environments (background noise, reverberation, multiple speakers, and so on).
Requires real-time feedback.
Limited computational and data storage resources.
Power use (i.e. battery life) is a critical factor.
Requires large vocabulary.

Speech recognition on existing PDAs is extremely limited in vocabulary size, speaker independence, performance, and noise robustness.

Stuff here...

[Hybrid Speech Recognition Block Diagram]

Using multiple signal processing and phone probability estimation components improves acoustic robustness.
Battery life can be prolonged by using the robustness components only when necessary.
The signal processing and phone probability estimation components are easily vectorized.
The decoder component consists of a large search that requires pruning to be tractable. Some progress has been made in vectorizing this component, but more research is required.