Earlybird-29M: 16.8MB speech model that types ahead of speech

Earlybird's design choice is unusual: most streaming speech models wait for a word to finish before displaying it, while this one predicts the word as it is being spoken and revises if needed. The developer's published figures put its median word display at 90ms before completion, compared with 90–400ms after completion for other streaming models, and the model attaches a 0–1 confidence score to every displayed word, including unfinished ones. According to the developer, unfinished words with confidence above 90% are correct 91–96% of the time.

The trade-offs are explicit. Accuracy on clean English read speech is strong for its size — 4.37% error on LibriSpeech clean — but jumps to 21.91% on earnings call audio and 21.23% on meeting audio, suggesting it is best suited to quiet, single-speaker content. The 4-bit weight format also means loading goes through a bundled earlybird.py file rather than standard Transformers auto-loading, which adds a file to manage in embedded deployments. And while the model runs on CPU, whether it keeps up with real-time audio depends heavily on the machine, so the developer advises measuring streaming speed in your own environment first.