About

Nothing else was listening properly.

So this was built in the open, for the people every other dictation tool quietly asks to pick a language.

Half the world thinks in more than one language at once. The sentence starts in one and finishes in another, and the word in the middle only exists in the other one. Every dictation tool treats that as an error. It picks a language, then translates or mangles whatever doesn't fit.

WhisprStream doesn't pick. It listens to the whole utterance, again and again, several times a second, and never cuts it into pieces — because the cuts are where mixed sentences break.

It runs entirely on the machine in front of you. Not as a privacy feature bolted on afterwards, but because a model small enough to stay warm in memory is also the fastest way to get the words back.

It was built for vibe coding. Describing what you want is faster out loud than typed, and the description arrives in whatever mixture of languages you happened to think it in. The agent does the typing either way.

What's inside

No surprises in it.

Speech modelQwen3-ASR 0.6B, quantised to 8-bit
RuntimeA separately downloaded, verified MLX engine on Apple silicon
NetworkFirst-run engine and chosen model downloads, plus update checks; audio stays local.
RequiresmacOS 14 or later, Apple silicon
LicenceMIT

Anyone bilingual.

Not just Chinese and English, and not just two at a time.

Vibe coders.

Say the prompt out loud. It lands at the cursor, in whatever editor you're in.

Anyone who reads the source.

Every line of it is public, including the part that hears you.

MIT licensed

Read every line.

Software that listens to you should be software you can audit.

Get the app ZIP from Releases View on GitHub