Murmur

I switch between Chinese and English mid-sentence constantly, and nearly every dictation tool falls apart exactly there. So I built one for myself first: hold a hotkey, talk, let go, and the cleaned-up text lands at your cursor. Three hotkeys, three uses — dictate, translate, ask — with a floating capsule on screen telling you it's listening.

To find out which speech engine is genuinely accurate on mixed Chinese-English, I built a benchmark: run my own real recordings through several engines and rank them separately for pure Chinese, mixed speech and pure English — never averaged, because no single model wins all three. That stubbornness gave me an answer I couldn't get anywhere else, and it decided which engine ships in the product.

In September 2026 it became publicly downloadable for the first time: signed, notarised, free to download. On-device recognition works out of the box — no sign-up, nothing to configure, and it still types when you're offline. The cloud tier is invite-only for now.

What I'm building now is its other hand: a Mac input method. Switch to the Murmur input source and you type pinyin for Chinese or plain English exactly as usual — then hold that one key, speak, and the words land at your cursor. It also learns from what you type, so what you say comes out more accurate over time. Typing and speaking should have been two hands of one product all along. The dictionary I built for it now runs past half a million entries.

It also learned one small thing I need daily: say “remind me about X in half an hour” and it just remembers, then glows softly at the edge of the screen when the time comes. No network, no cost, and it survives a restart.

The name and the domain aren't settled yet — it's borrowing another project's domain for now. But the order is clear to me: make it genuinely good first, worry about the rest after.