Import
Local video file, audio file (MP3, WAV…) or YouTube link with a resolution picker (up to 4K).
- Beta 1.21
- Sep 20, 2026
- MIT license
- Windows + Pinokio
- Free, 100% local
ZastTranslate
1-click video translation and dubbing, in your own voice. 100% local, no API keys, no subscription.
ZastTranslate is an open-source app for Pinokio. It transcribes a video with WhisperX, separates voice from music with Demucs, translates with a local language model (Qwen, EuroLLM), then re-synthesizes the speaker's voice in 33 languages with VoxCPM 2. Everything runs on your PC: your videos never leave your machine.
Video tutorial
How to dub any video with your own voice, for free, with local AIThe complete step-by-step walkthrough, from installation to multi-language dubbing with voice cloning and subtitle stabilization. The video is in French.Watch on YouTube →
What ZastTranslate does
Accurate transcription
WhisperX with word-level timestamps and an 8-step subtitle stabilization (no one-word cues, no overlapping timecodes).
Audio separation
Demucs isolates the voice from music and effects, then remixes them with the dubbed voice.
Local translation
Qwen2.5, Qwen3.5 or EuroLLM. Two versions: a faithful translation and a shortened "Fitted" version that fits each segment's duration.
Voice cloning
VoxCPM 2 clones the speaker's voice (or a 10 to 30 s sample) in 30 languages, with no API key.
Segment editor
Interactive timeline: click to seek, adjust start and end, fix the text, re-synthesize a single line.
Bulk mode
Several languages at once: dubbed videos, audio tracks, subtitles and localized titles/descriptions. The "subtitles and metadata only" mode takes under 15 seconds.
YouTube SEO Studio
Titles, 200 to 300 word descriptions, full-length chapters and 4 hashtag packs.
WordPress Blog Studio
Turns a video into an article (meta description, slug, 1080p/2K frames pulled from the video) with copy-ready Gutenberg HTML.
Vertical 9:16 Shorts
Detects 1 to 5 viral moments, reframes them vertically, adds word-by-word karaoke subtitles.
AI thumbnails (optional)
Qwen-Image-2.1: YouTube thumbnails, RGBA transparency and 3-variant A/B packs. Downloaded only on request (~33 GB).
Built-in guides
Every tab has a collapsible guide: goal, options explained and click order.
How it works
Import
Drop a video or audio file, or paste a YouTube link.
Transcribe
Pick the source language and Whisper model, review the text, then click "Validate Transcription".
Translate
Pick the target language. Edit the translation and the Fitted version, then validate.
Dub
Choose a voice mode (default voice, clone from the original, clone from a file) and generate the dubbed MP4.
Export in bulk
Bulk mode repeats the process for every language you pick, with SRT, VTT, SBV subtitles and a ZIP archive.
Installation
- Install and open Pinokio.
- Add the zast57/ZastTranslate repository from the discover page, or paste its address.
- Click Install: Python, PyTorch CUDA, VoxCPM 2 and all dependencies are set up automatically.
- Click Start: the Gradio interface opens in your browser. Models download on first use.
System requirements
Core app
| GPU | NVIDIA with 6 GB VRAM or more (CUDA). RTX 50-series supported. |
|---|---|
| Memory | 16 GB RAM recommended |
| Disk | about 8 GB for models (downloaded on first use) |
| OS | Tested on Windows only (with Pinokio) |
Qwen-Image-2.1 thumbnail studio (optional)
| GPU | NVIDIA with 8 to 12+ GB VRAM (with CPU offload), or 24 GB for native BF16 |
|---|---|
| Memory | 32 GB RAM recommended |
| Disk | about 33 GB for model weights |
The thumbnail studio is optional: the app works fully without downloading it.
Current limitations
- No lip-sync: the dubbed audio replaces the original voice, but the picture is not modified.
- Single voice: all segments are dubbed with the same voice, multi-speaker support is not available yet.
- Tested on Windows only.
Frequently asked questions
Is ZastTranslate free?
Yes. It is open source under the MIT license and runs locally: no API key, no subscription, and your videos are never sent to an online service.
What hardware do I need?
An NVIDIA GPU with 6 GB of VRAM or more, 16 GB of RAM and about 8 GB of disk space for the models. It is tested on Windows with Pinokio.
How many languages can it dub into?
The VoxCPM 2 voice engine supports 30 languages. The list you see depends on both the voice engine and the translation model: EuroLLM, for example, only covers European languages, while Qwen covers them all.
Can it clone my voice?
Yes. The "Clone from original" mode reproduces the speaker's voice from the extracted vocal track. You can also provide a clean 10 to 30 second sample.
Are my videos uploaded to the internet?
No. Transcription, translation and voice are all computed on your machine. The internet is only used to download models and, if you ask for it, a YouTube video.
Do I need an API key or a subscription?
No. Everything relies on open models: WhisperX, Demucs, Qwen, EuroLLM and VoxCPM 2.
How do I update it?
Use the project update in Pinokio. The version list below summarizes the changes; the full detail is in the GitHub repository.
Version history
- Beta 1.21Sep 20, 2026
Qwen-Image-2.1 thumbnail studio (optional): RGBA transparency, 12 styles, 3-variant A/B packs, AI metadata stripped on export. Cleanup scripts and buttons to free VRAM and disk.
- Beta 1.20Sep 8, 2026
Quadruple acceleration with zero quality loss (voice caching, NVENC encoding, C++ GGUF engine, KV cache). No more dangling sentence fragments in subtitles. Complete SEO titles.
- Beta 1.19Sep 7, 2026
Much faster translation and Fitted pipeline, VRAM protection on 8 GB cards.
- Beta 1.18Sep 6, 2026
Subtitle pipeline overhaul (4 critical bugs), 222-rule correction dictionary, Blog Studio, viral Shorts, YouTube SEO kit.
- Beta 1.10Aug 30, 2026
YouTube SEO Studio, full-timeline chapters, landmark detection, 4 hashtag packs, subtitles & metadata only mode.
- Beta 1.08Aug 25, 2026
NVIDIA RTX 50-series (Blackwell, CUDA 12.8), WhisperX priming with tech vocabulary, brand name normalizer.
- Beta 1.07Aug 22, 2026
Filler word cleanup, SRT, VTT and SBV export, YouTube title and description localization.
- Beta 1.06
Interactive segment editor with timeline, click to seek, single-line re-synthesis.
Technologies used
ZastTranslate builds on excellent open projects:
Who is behind ZastTranslate?
I am Zast, a French web creator since 1996: CineAlliance (1998), ParadoxeTemporel (2011), MusiqueAlliance (2014), VisiteEnFamille (2016). I built ZastTranslate to translate my own YouTube videos without relying on a paid service, then opened it to everyone.