Earlybird: Predictive Streaming Speech Recognition in 29M Parameters
arXiv cs.CL · 11 Nov 2026, 03:59 AEST
Earlybird is a new streaming speech recognizer that predicts words before the speaker finishes them. It achieves this with 29 million parameters, outperforming existing models in speed and accuracy.
Key facts
- Streaming speech recognizers output a word only after the speaker has finished it, typically 90-400 ms later.
- Earlybird learns from audio cut off at random points, including partway through a word, and is asked to predict every word that is due by then in full, plus a guess of the next two words.
- On clean audiobook speech (LibriSpeech test-clean), 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.
- On the same audiobooks, 69% of words are predicted correctly before they end, against 3-25% for the other models.
- 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.
Read aloud
Earlybird is a new streaming speech recognizer that predicts words before the speaker finishes them. It achieves this with twenty nine million parameters, outperforming existing models in speed and accuracy. Streaming speech recognizers output a word only after the speaker has finished it, typically ninety-four hundred ms later. Earlybird learns from audio cut off at random points, including partway through a word, and is asked to predict every word that is due by then in full, plus a guess of the next two words. On clean audiobook speech LibriSpeech test-clean, the median word is predicted correctly ninety ms before the speaker finishes it, while released streaming models get it ninety to four hundred ms after it ends. On the same audiobooks, sixty nine percent of words are predicted correctly before they end, against three-twenty five percent for the other models. 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.
Sources in this cluster
- Earlybird: Predictive Streaming Speech Recognition in 29M Parameters
arXiv cs.CL · 11 Nov 2026, 03:59 AEST