Earlybird: Predictive Streaming Speech Recognition in 29M Parameters
On clean audiobook speech, Earlybird correctly predicts 69% of words before the speaker finishes them.
The paper reports a median correct prediction 90 milliseconds before a word ends, compared with 90–400 milliseconds afterward for the released streaming models tested. Earlybird trains on audio cut off at random points, learning to complete words and guess the next two. It produced the earliest predictions across all six timed test sets, spanning audiobooks, talks, public speech, earnings calls and synthetic speech. Across eight English recognition benchmarks, its accuracy beat Moonshine Tiny on every set and roughly matched the similarly sized Whistle.
ArXiv · AI/CL/LG's note
The paper reports a median correct prediction 90 milliseconds before a word ends, compared with 90–400 milliseconds afterward for the released streaming models tested. Earlybird trains on audio cut off at random points, learning to complete words and guess the next two. It produced the earliest predictions across all six timed test sets, spanning audiobooks, talks, public speech, earnings calls and synthetic speech. Across eight English recognition benchmarks, its accuracy beat Moonshine Tiny on every set and roughly matched the similarly sized Whistle.
ArXiv · AI/CL/LG's note
score 5