Skip to content

v0.2.0-beta.1 — Synthetic voice and bidirectional translation

Pre-release
Pre-release

Choose a tag to compare

View all tags
@alexwing alexwing released this 02 Aug 00:08
· 16 commits to main since this release
v0.2.0-beta.1
08493b0

Beta. Works end to end, tested on one machine (RTX 3060 12 GB, Windows 11), not yet in a real meeting with other people.

You are in a Teams call. The others speak English. You speak Spanish into your microphone, and they hear it in English, in your own voice, through a virtual microphone. At the same time, what they say you read in Spanish on screen.

Everything runs locally on your GPU. No API key, no audio leaves the machine.

What it does

  • Transcribes what plays on your PC and your microphone at once, as two independent sessions.
  • Translates both, each in its own direction — four language slots, not one.
  • Speaks your side in the other language, in a clone of your voice (Chatterbox, 23 languages) or a neutral one (Kokoro, 8), out through VB-CABLE so the meeting hears it and your speakers don't.
  • Meeting view, subtitle overlay, configuration profiles, export to .txt/.srt, interface in English and Spanish.

Full walkthrough — install, wiring, first meeting — in the README.

Install

Download the .exe, run it, then provision Python and the models from the app's own folder — the installer carries the scripts, so there is nothing to clone:

& "$env:LOCALAPPDATA\LiveTranscriber\install.cmd" -InstallPython -WithVoice

Use install.cmd, not the .ps1: a fresh Windows refuses to run downloaded scripts.

Needs an NVIDIA GPU, Turing or newer, and 15 GB of disk — 27 GB with the voice. For the virtual microphone, VB-CABLE, installed separately. Detail in INSTALL.md.

Configuration and logs live in %APPDATA%\LiveTranscriber\.

Measured on an RTX 3060 12 GB
Recognition (lookahead 3) 320 ms, 2.40 GB VRAM per session
Translation ~160 ms per sentence, 1.27 GB
Voice, cloned (Chatterbox) 0.84–1.03x, 3.40 GB
Voice, neutral (Kokoro) ~47x, 0.56 GB

The voice figures come from a separate bench, and were taken with the voice alone — never with the recognizers and translator on the same card.

Limits

  • Chatterbox sits at the edge of real time. The margin over 1.0x is 3%, and grouping plus a lag indicator exist because of it. Kokoro has no such problem but gives up voice cloning.
  • Interpreting one sentence behind, not simultaneous. Nothing is spoken until its sentence is complete and translated.
  • A cascade: a misheard word propagates to the translation and gets spoken in your voice. The screen always shows what was said in your name.
  • NLLB-200 is CC-BY-NC — non-commercial. The voice engines are MIT and Apache 2.0; the app's code is MIT.
  • Chatterbox audio carries Resemble AI's Perth watermark.
  • Windows and NVIDIA only. Unsigned, so SmartScreen will warn.
Assets 4
Loading