Earlybird: a 29M streaming recognizer that outputs words before the speaker finishes them
A 16.8 MB, MIT-licensed model flips streaming ASR's +90–400 ms word latency to −90 ms, and exposes calibrated confidence on words still being spoken.

A 29M-parameter streaming recognizer called Earlybird outputs words before the speaker finishes them — a median 90 ms early on audiobooks, where released streaming models land 90–400 ms after the word ends. It is a single-author arXiv preprint with MIT-licensed weights and code, and its real contribution is treating "waiting for the word to end" as a design choice rather than a law of streaming ASR.
What shipped
Earlybird is a 29M-parameter encoder-decoder fine-tuned from Moonshine Tiny (34M params, MIT), with the vocabulary cut to 19,007 tokens to reach 29M parameters and 16.8 MB at 4-bit weights (paper, model card). The author, Sahil Mahendrakar, released the weights, full training and eval code, a timing harness, and result tables under MIT (GitHub).
The training trick is the whole idea. Utterances are cut at random 80 ms steps — including partway through a word — and the model outputs every word "due" at that point in full, plus a guess at the next two words. A word counts as due at max(s_i + min(0.5·(e_i−s_i), 50 ms), e_i − 300 ms): 300 ms before it ends, or once 50 ms of it has been heard (paper HTML). Knowledge distillation from Moonshine Tiny is applied only where the teacher's top choice matches the label, and three silence padding frames after the heard audio shift the median from −50 to −90 ms.
The headline numbers are the author's own timing-harness results, not independently replicated:
- Median word predicted 90 ms before the speaker finishes it on LibriSpeech test-clean, vs +90 to +400 ms for released streaming models (abstract).
- 69% of words predicted correctly before they end on audiobooks, vs 3–25% for the other models.
- Earliest on all six test sets the author timed — audiobooks, harder audiobooks, TED talks, public speech, earnings calls, and synthetic speech. Median word latency ranges from −46 ms (earnings calls) to −90 ms (audiobooks), against +90 to +535 ms for Moonshine Tiny, Whistle, and Whisper base (author write-up).
- On the eight English Open ASR Leaderboard sets, the 4-bit model averages 12.63% WER vs Moonshine Tiny's 14.35% and roughly ties same-size Whistle at 12.17% (result tables).
Why it matters
For a voice agent, word latency is dead time: the agent cannot start a search or plan a reply until the words are out. Earlybird's −90 ms median means roughly 400 ms of extra "thinking time" per word compared with a model that waits 300 ms after the word ends. That is the difference between an agent that starts acting while the user is still talking and one that waits for silence.
The more useful piece is the calibrated per-word confidence on words still being spoken. Earlybird's raw token probabilities give a confidence for every word, including the one mid-utterance, with an expected calibration error (ECE) of 3.7–5.5 on unfinished words and 0.8–4.4 on next-word guesses. At a stated confidence of ≥90% on an unfinished word, it is right 91–96% of the time (author write-up). That is what makes early output actionable: a downstream agent can decide to trust a high-confidence early word and act on it, while treating low-confidence ones as provisional. This is the pattern to steal — a recognizer that exposes when it is confident enough to act early, not just a faster transcript.
It is also a rare fully open systems paper: weights, training scripts, quantization, and the streaming timing harness are all public under MIT, so the numbers are reproducible rather than a press release.
Caveats
This is a preprint, single-author, and not peer-reviewed. Everything above is the author's own reporting; there is no independent replication yet.
- Latency excludes compute time by design. The harness measures what each model decides to output, not wall-clock end-to-end latency, so real on-device latency will be larger than the headline −90 ms.
- The baselines are set up unusually. Whisper base is not a streaming model and is re-decoded every 0.48 s; Moonshine Tiny is compared via its quantized runtime (14.35% mean) while its own published full-precision leaderboard entry scores 12.0%.
- Earliness costs stability. The model revises words ~1.3 times per word on audiobooks and 2–6 times on spontaneous or noisy speech, versus 0.2 for Whistle and 0.4 for Whisper base. Next-word guesses are weak — at ≥50% confidence, only about 1 in 7 is guessed correctly before it ends.
- Noise robustness is poor. At 10 dB SNR the author reports 32% WER vs 8.9% for Whisper base. It is English-only, and out-of-domain calibration degrades (SLURP ECE 8.8).
- The speed claims are internally inconsistent. The timing harness reports ~39 ms per 80 ms step (about 2× real time on one CPU thread), while the Open ASR tables separately claim 10–11× audio-seconds-per-compute-second on a different machine. The author does not reconcile the two.