Demo · live

Earlybird, live

Press the microphone and talk, or play one of the recordings below. Earlybird writes each word before you finish saying it, guesses the next ones in grey, and shows how sure it is of every word.

Ready when you areEnglish · runs on our server, audio is not stored
Or play a recording through the model
Your words appear here as you speak.
Being spoken—
 
Next word, guessed—
 
Word after—
 

The 16.8 MB 4-bit Earlybird model runs on a small CPU server; your audio (or the recording you pick) is streamed to it in 80 ms steps and discarded. The recordings are held-out test clips the model never trained on. Confidences are the model's own probabilities, with no calibration step. Sessions are limited to two minutes.