This was the initial prototype version of Parakeet. For the latest information on Parakeet, go here.

In this project, I investigated using speech recognition on the Nokia N800 internet tablet. I used a bluetooth microphone for audio capture and the PocketSphinx recognizer for the core speech recognition. Using a 5K-word vocabulary, I found I could achieve good accuracy and near real-time recognition performance. I developed a stylus-based interface which allowed easy correction of any errors. The correction interface was based on visualizing the likely alternates for each word in the best hypothesis.

The sections below detail how I got things going. They give results of a variety of recognition experiments. They also include links to various goodies to download if you want to have a play.

Audio and acoustic modeling
Decoding
Language modeling and recognition results
Test sets and adaptation results
Improving recognition accuracy
WSJ Dictation task
Confusion network experiments
Stylus-based correction interface
Discussion and recommendations

Correcting an error by selecting an alternate word.

Correcting an error by copying a word from another area.

Correcting an error using the onscreen keyboard.