v1.7.0: Modern LSTM, Blended Predictions, Bug Fixes

What's New

• Retrained LSTM on modern dialogue. The predictor now trains on 20 million tokens of movie and TV dialogue (OpenSubtitles) instead of 19th-century literature. Predictions are much more natural and conversational.
• Longer context window. Extended from 4 to 6 words for better use of sentence context.
• Blended predictions. The LSTM now works together with the n-gram model instead of replacing it. LSTM provides context-aware suggestions first, then n-gram fills remaining slots with words learned from your own writing.
• Fixed a context-corruption bug. Accepting a prediction was adding the word to the context buffer twice, so the next prediction ignored all preceding words. This was a long-standing bug that made predictions after accepting a word much worse.
• Fixed a padding bug. Short contexts were padded with the wrong token, degrading prediction quality.

Download

Download the .nvda-addon file below and open it to install.