Earlybird

Predictive speech-to-text in 17 MB. Earlybird writes each word before the speaker finishes saying it, and says how sure it is about every word, including the one still being spoken and its guesses of the next two.

Pick a recording and press play to see what each model had on screen at every moment. Blue italic words are Earlybird's guesses of what comes next; small numbers are its confidence. Try 0.25× to watch words appear before they're finished.

Try it live with your microphone →Read the write-up
How the comparison runs. The same harness as the write-up: each clip fed in 80 ms chunks followed by 1 s of silence, every update logged on the audio clock (compute time not added). Earlybird (fine-tuned from Moonshine Tiny) is the released 4-bit model with 2 guessed words. Moonshine Tiny is the official moonshine-voice runtime, updated every 80 ms. Whisper is re-run on all audio so far every 0.48 s, raw output. Whistle is the Cactus engine in live mode. Word ends come from forced alignment of the reference transcript. Clips are held out from training: SLURP (CC BY 4.0), The People's Speech (CC BY / CC BY-SA), LibriSpeech test-clean (CC BY 4.0).

Model: sahilmahendrakar/Earlybird-29M