frontier models
The Recognizer Predicts the Word Before It Ends
Earlybird trains on audio cut off mid-word and emits calibrated guesses from a 29-million-parameter streaming model.
Summary
Earlybird trains on audio cut off mid-word and emits calibrated guesses from a 29-million-parameter streaming model.
Earlybird is a 29-million-parameter, 16.8-megabyte-at-four-bits speech recognizer trained to predict words from audio truncated at random moments, including mid-word. On LibriSpeech test-clean, the authors report that the median word was predicted correctly 90 milliseconds before the speaker finished it, while compared released streaming models emitted words 90 to 400 milliseconds afterward. Sixty-nine percent of words were predicted correctly before completion on that set. Results across six timing sets and eight Open ASR sets position the model as a low-latency research system, not a universal guarantee across speakers or environments.
Why it matters
Earlybird trains on audio cut off mid-word and emits calibrated guesses from a 29-million-parameter streaming model.
Limits and context
- Results across six timing sets and eight Open ASR sets position the model as a low-latency research system, not a universal guarantee across speakers or environments.
Key claims
Earlybird trains on audio cut off mid-word and emits calibrated guesses from a 29-million-parameter streaming model.
Qualification: Results across six timing sets and eight Open ASR sets position the model as a low-latency research system, not a universal guarantee across speakers or environments.
Evidence: source-2026-11-11-011
Sources
- arXiv:2611.14421v1arXiv / Cornell University · primary research
Corrections
No corrections have been recorded for this story.