Live call translationHow to translate a Zoom, Meet or Teams call in real time
You speak your language and the other person hears a translated voice in theirs. They answer, and you hear it in yours through your headphones. Only you install chasqui: the other person joins the call as usual.
What you need
- Windows 10 or 11, and headphones. With open speakers, your mic picks up the translation and it loops.
- chasqui and an OpenAI API key with access to
gpt-realtime-translate. - VB-Audio Cable (free), which carries your translated voice into the call.
- A way to capture what the other person says: Stereo Mix turned on in Windows, or a second cable such as Hi-Fi Cable. Without one, only your side gets translated.
Step by step
- Open chasqui. The first time, the setup window checks your key, the cable and the device switcher.
- Choose your language in the left box (I SPEAK) and theirs in the right box (THEY SPEAK).
- Pick Headset mode and press START. Wait for the light to say ON ROUTE.
- Join the call. The app has to use the Windows default mic: if it has a specific mic selected, change it to the default one or to CABLE Output.
- When you finish, press STOP or close the window. chasqui gives your Windows mic and output back.
Prefer not to have your Windows defaults touched? Meet mode leaves them alone: inside the call app you set the mic to CABLE Output and the speaker to your second cable.
If the other person hears nothing
Call apps can treat a synthesized voice as noise and filter it out. The turn shows up in chasqui's log, but the other side hears silence. Set noise suppression to Low or Off:
| App | Setting |
|---|---|
| Google Meet | Settings > Audio > Noise cancellation |
| Microsoft Teams | Settings > Devices > Noise suppression |
| Zoom | Settings > Audio > Background noise suppression |
What it costs
A call runs two translation streams, one per direction, so it costs $0.102 a minute: $3.06 for 30 minutes, $6.12 for an hour. OpenAI bills it to your own key. There is no subscription, and the window shows the running total while you talk.
What to expect
- Delay: usually a second or two. We measured about 1.6 seconds from the original words to the translated voice, and the last words of a turn can come a little later.
- Mistakes: the model sometimes answers a short phrase instead of translating it, so "Did you get there okay?" can come back as a reply. Longer sentences hold up better. The window shows both sides, so you can catch it.
- Privacy: the audio of both people goes from your PC to OpenAI, with no chasqui server in between. Tell the other person you are using a translator.
- Languages: English, Spanish, Portuguese, French, German, Italian, Japanese, Korean, Mandarin, Arabic, Russian and Hindi.
Free and open source. You pay OpenAI by the minute, with your own key.