In development · macOS and iPhone
Dictation that keeps up when you switch 中文 and English mid-sentence.
Most dictation tools are built English-first. They handle a clean English sentence well, then fall apart the moment you say a Chinese clause with three English product names in it. Voicelna is built for that sentence.
- ~1.6 s
- Median time from the moment you stop speaking to cleaned text at your cursor (90th percentile 2.5 s). Raw text lands at ~0.5 s if you prefer speed over polish.
- ≤ 40 terms
- Your own vocabulary — names, models, jargon — sent with each dictation so the recogniser stops guessing. Kept on your device.
Why bilingual dictation is the hard case.
The recogniser picks a lane
Most engines detect a language and commit to it. Say a Chinese sentence with “Kubernetes” and “staging” in it, and the English words come back as phonetic Chinese — or the whole line flips to English.
Proper nouns are guesswork
Product names, client names, internal jargon: the recogniser has never seen them. It substitutes the nearest common word, and you fix it by hand.
How Voicelna works.
Hold a key and speak
On a Mac, hold right Option in any app and talk. It connects while you are still speaking, so there is no wait at the end.
Raw text comes back first
Streaming recognition, tuned for sentences that switch language mid-clause. Median 0.47 s from the moment you stop speaking.
A language model cleans it up
Punctuation, sentence breaks, filler words removed, lists formatted as lists. This is the step that turns speech into text you'd send.
It lands at your cursor
Not a transcript window you copy from. The finished text goes where you were already typing — email, editor, terminal, anywhere.
Where Voicelna sits.
| Built-in macOS dictation | English-first AI dictation | Voicelna | |
|---|---|---|---|
| Chinese-English in one sentence | Partial | Weak | Yes, mid-sentence |
| Cleans up filler and punctuation | No | Yes | Yes |
| Learns your proper nouns | No | Rarely | Yes, up to 40 per dictation |
Where your words go.
Voicelna is not released yet, and the service that recognises the audio may still change before launch; the privacy notice it ships with will name it. What is settled: the names and terms Voicelna learns are kept in a file on your own device, and each dictation sends along only the few that matter for it.
Questions.
Can it learn names it keeps getting wrong?
Yes. Voicelna keeps a personal vocabulary of the terms you use, together with the wrong spellings the recogniser has produced for them. Both are sent with the next dictation, so a name you corrected once tends to stay correct. The list is capped on purpose — an oversized glossary makes accuracy worse, not better.
Which platforms does it run on?
macOS, with a global hotkey that works in any application, and iPhone, as a third-party keyboard so dictation works inside other apps.
Is it available yet?
Not yet. Voicelna is in development, and we use it every day ourselves. Email hello@aivmelna.com to be told when it ships.