LiveTranscriber: understand and speak another language, in your own voice

LiveTranscriber does two things at once: it shows you in writing, in your language, whatever is playing on your computer; and it lets you speak another language in your own voice. Everything runs locally — no cloud, no API bill, and the audio never leaves the machine.

Reading what everyone else says

It captures any system audio — a Teams call, a film, a video in the browser — and transcribes it live. It can also listen to the microphone at the same time, as separate sessions, so you get both sides of a conversation.

On top of the transcript you can switch on translation into 40+ languages. Each sentence is translated the moment its punctuation closes it, not when the paragraph ends, so the translated text runs one sentence behind instead of appearing in blocks whenever someone stops talking. On screen it is grouped into paragraphs so it reads like normal text.

Speaking in your own voice

This is the part that sets it apart. You speak into the microphone in Spanish, and what comes out the other side is that same sentence in the target language and in your voice — not a generic synthetic one: the model clones your timbre from a sample of you.

And it doesn't stay inside the app. The voice plays through a virtual microphone, so any program that takes a mic input — Teams, Meet, Discord — hears it as if you were speaking. It is verified end to end: the cloned voice went in through the virtual cable and the recogniser, listening on the other side, transcribed it word for word, punctuation included. What the other person would hear is exactly what you said.

There are two modes. With voice cloning, 23 languages and your own timbre. Without it, a much faster neutral voice in 8 languages, useful when fluency matters more. The feature is entirely optional: switched off it costs nothing — no process, no VRAM.

The meeting in English

The scenario that joins both halves: you are in a meeting held in English. You read what the others say in Spanish, as they say it. And when it is your turn, you speak Spanish and they hear you in English, in your voice. No switching languages in your head in either direction.

Only the translation of your microphone is ever spoken, never other people's sentences — playing their own voice back at them would make no sense — and it handles echo: if your voice comes out of the speakers and the mic picks it up, it does not re-enter the loop.

What you take away

A session can be exported to text or .srt subtitles, in the original language, translated, or bilingual with both versions together. Files are named by date and never overwritten: two exports on the same day give you two files, not one lost.

A note on how it is built

Audio capture and all the logic are in Rust, the app is Tauri with a React frontend, and the models run in separate processes: recognition with NVIDIA's Nemotron-3.5-ASR, translation with NLLB-200 and voice with Chatterbox Multilingual. Each engine sits behind an interface, so swapping any of the three touches neither the capture layer nor the UI.

It needs Windows and an NVIDIA GPU; an RTX 3060 handles it comfortably. And it is worth keeping in mind that it is a chain: if recognition mishears a word, the translation carries the error along. This is translated subtitling one sentence behind, not professional simultaneous interpretation.

Source on GitHub.