Skip to content
Voxa
Open source · MIT

Dub any video into another language — and keep it in sync.

Whisper transcription, context-aware LLM translation, four TTS engines, and placement locked to the speaker. MIT-licensed. Runs on your machine.

Get startedView on GitHub
bash
$ voxa talk.mp4 --target_lang ru

Demo

See it in action

The same clip, dubbed into three languages — each locked to the speaker.

One source video, dubbed by Voxa. Pick a language.

Why Voxa

The dub never falls behind the speaker.

Naive tools concatenate clips and drift. Voxa anchors every clip to the source timeline, so the dub stays locked to the speaker — no growing lag, no re-syncing by hand.

Features

Everything the pipeline needs

A complete dubbing toolchain — transcription to final mux — in a single command.

Anchored placement
Every clip is pinned to the source timeline. Drift stays at zero.
Four TTS engines
Edge, OpenAI, Piper (offline), and XTTS voice cloning.
Context-aware translation
Google, Ollama, OpenAI, or Anthropic — translate with full-sentence context.
Quality gate
Optional WER scoring re-synthesizes lines that miss the mark.
Self-hostable
Point any OpenAI-compatible TTS endpoint with --openai-tts-base-url.
Resumable
Checkpoints let you stop and pick up exactly where you left off.

How it works

Seven steps, one command

Voxa runs a deterministic pipeline from raw audio to a synced dub.

  1. 01

    Extract audio

    Pull the audio track from the source video with ffmpeg.

  2. 02

    Denoise

    Clean background noise so transcription stays accurate.

  3. 03

    Transcribe

    Whisper turns speech into timed text.

  4. 04

    Merge sentences

    Group fragments into natural, translatable sentences.

  5. 05

    Translate

    An LLM translates with full-sentence context.

  6. 06

    Synthesize

    The chosen TTS engine speaks each line.

  7. 07

    Mux

    Anchor every clip to the timeline and remux the video.

Supported models

Pick your engine

Mix and match transcription, translation, and speech to fit your setup.

StageOptionsNotes
TranscriptionWhisper (faster-whisper / openai-whisper)Runs locally, multiple model sizes
TranslationGoogle · Ollama · OpenAI · AnthropicFree Google or context-aware LLMs
Text-to-speechEdge · OpenAI · Piper · XTTSCloud, offline, or voice cloning

Voices

Hear the OpenAI voices

Six OpenAI text-to-speech voices (tts-1-hd). Pick one and press play.

OpenAI · tts-1-hd

“Voxa dubs your videos into another language and keeps them in sync with the speaker.”

Installation

Up and running in a minute

Requires Python 3.10+ and ffmpeg on your PATH.

clone
$ git clone https://github.com/akshinmrv/Voxa.git && cd Voxa
install
$ pip install .
run
$ voxa talk.mp4 --target_lang tr

The voxa command comes from pip install .; you can also run python voxa.py directly.

Quick start

Common commands

Copy, paste, dub.

bash
$ voxa talk.mp4 --target_lang ru

FAQ

Questions, answered

Does the dub drift out of sync?

No. Voxa anchors every clip to the source timeline, so placement stays locked to the speaker.

Can it run fully offline?

Yes — Whisper locally, Ollama for translation, and Piper for speech. No cloud calls required.

Can I use my own TTS server?

Yes. Any OpenAI-compatible endpoint works via --openai-tts-base-url.

Which languages are supported?

Whisper covers around 100 source languages; targets depend on the TTS engine you choose.

Is it really free?

Yes, Voxa is MIT-licensed. Some optional engines carry their own licenses.

Author

Built by Servoogle

Voxa is an open-source project by Servoogle, created and maintained by Akshin Miranov — practical, self-hostable tools with no lock-in and no telemetry.

Servoogle
Open-source developer tools
View the repository
Akshin Miranov
Creator & maintainer
GitHub profile