Fully offline
Models are downloaded once and verified; after that it works without a network. Nothing you say is uploaded.
Offline speech translation · Go
Speak in any language,hear it in another
PTranslator listens to your microphone, recognises the language, translates the sentence and reads it aloud. Everything runs on your own computer: no cloud service, and no audio ever leaves the machine.
- Offline
- Whisper
- Silero VAD
- Bergamot
- Supertonic
- Go
$ ptranslator live --to ja [1] en (2.82s): Hello, how are you today? → こんにちは、今日はお気持ちですか? VAD 504ms · STT 223ms · translate 226ms · TTS 291ms # every stage is timed, every sentence
0bytes of audio sent to any server
6languages: English, Vietnamese, Japanese, Korean, French, Chinese
~1 sfrom the end of a sentence to speech, on the light model
3recognition sizes, from light to accurate
What it does
A whole interpreter on your computer
Four steps in one pipeline — and nothing to sign up for.
Knows who is speaking
Speak in any supported language; PTranslator detects it for each sentence, so a conversation can switch languages.
Translates through English
Any pair of Vietnamese, Japanese, Korean, French and Chinese goes through English, using Mozilla's open Firefox Translations models.
Reads it aloud
A natural synthetic voice in the target language, with ten voices and adjustable speed.
Models on demand
Missing models are fetched with their size and licence shown first, resumable, checked by checksum, and kept under a memory limit.
Measured, not promised
Every sentence reports the time of each stage, and the project documents its known limits and test results openly.
How it works
From voice to voice in one second
Five stages run in parallel, so the next sentence can already be heard while this one is spoken.
- 1
Listen
The microphone is watched by a voice-activity detector that cuts the audio into sentences.
- 2
Recognise
Whisper turns speech into text and detects the language.
- 3
Translate
The sentence is translated, through English when neither side is English.
- 4
Speak
A synthetic voice reads the translation in the target language.
- 5
Show
The original, the translation and the time of each stage are printed.
FAQ
Common questions
Does it need the internet?
Only once, to download the models (roughly 800 MB for the core set and two languages). After that it works fully offline.
Which languages does it translate?
English, Vietnamese, Japanese, Korean, French and Chinese, in any direction through English. Speech can be read aloud in many more languages.
How accurate is it?
On the project's own test of synthetic voices, the largest recognition model made errors on about 5% of words in clean audio and 13% with noise. Real voices, accents and small translation models will do worse.
Is there an app?
Not yet: PTranslator is a command-line program for now, with live translation, file transcription, translation and text-to-speech commands.
Can I use it?
PTranslator is in development by the Calix team. Get in touch if you would like to try it.
Want a translator that keeps your voice private?
Tell us the languages and the situations you need it for.
Get in touch