Earlybird is a small transformer model that outputs every rising word from its first sounds. On LibriSpeech test‑clean, 69 % of the words are correctly predicted before they finish; the median word appears 90 ms early. Across six speech domains— audiobooks, TED talks, public speeches, earnings calls, and synthetic speech—Earlybird keeps this early‑prediction advantage. The model’s confidence estimates for words still being spoken are well calibrated, with an expected calibration error between 3.7 % and 5.5 %.

The network was built by taking a pre‑trained Moonshine Tiny encoder‑decoder and trimming its vocabulary to 19 007 tokens, then training it to predict every word that has arrived in the audio window plus the next two words as guesses. No changes were made to the underlying architecture.

On LibriSpeech test‑clean, the median word latency of Earlybird is –90 ms, outpacing Moonshine Tiny (+270 ms), Whistle (+400 ms), and Whisper base (+90 ms). On the other five datasets the improvement ranges from 60 ms to 73 ms. Earlybird’s overall word error rate on the Open ASR Leaderboard is 11.7 % in full precision and 12.7 % after 4‑bit quantization, lower than Moonshine Tiny (12.0 %) and comparable to Whistle (12.2 %).