New AI Predicts Words Before You Finish Speaking
This could make voice assistants feel instant and save you time.
Earlybird is a 29M-parameter streaming speech recognizer—16.8 MB at 4 bits—that can predict words before a speaker finishes saying them, so a voice agent can start a search or plan its reply sooner. On clean audiobook speech, the median word is predicted correctly 90 ms before the speaker finishes it, while released streaming models get it 90 to 400 ms after it ends.
Earlybird is the earliest on all six test sets timed, covering audiobooks, TED talks, public speech, earnings calls, and synthetic speech; on the same audiobooks, 69% of words are predicted correctly before they end, versus 3–25% for the other models. For words still being spoken, confidence is calibrated with an ECE of 3.7 to 5.5, and on the eight English Open ASR Leaderboard sets it is more accurate than Moonshine Tiny on every set and roughly ties Whistle, a model of the same size.
- Earlybird predicts words before you finish speaking, making voice assistants respond faster.
- It's tiny (16.8 MB) and can run on your phone without the cloud, saving data and improving privacy.
- On audiobooks, it correctly guessed 69% of words before they ended, versus 3-25% for other models.
Why It Matters
Faster voice assistants mean quicker answers, less waiting, and smoother conversations with AI.