If you're dubbing a 10-minute video weekly across 5 languages, this saves you about $1,800/year vs cloud tools — and the dub never leaves your machine.

1. Clone or unzip the repo
2. Double-click install.bat   ← installs everything (~5-10 min)
3. Double-click start.bat     ← browser opens at http://localhost:8910
4. Paste YouTube URL → pick language → Start

git clone https://github.com/TachikomaRed/tachidubb && cd tachidubb
chmod +x install.sh
./install.sh    # installs everything + creates start.sh
./start.sh

First dubbing run downloads the VoxCPM2 model (~5 GB) — one time.

This is what makes TachiDUBB different. You don't have to touch the UI to use it.

You:    Dub https://youtu.be/abc into French, Spanish and Japanese,
        then stitch them into one 60-second showcase reel.

Claude: [calls tachidubb_showcase(...)]
        [polls tachidubb_get_showcase(...)]
        Done — http://localhost:8910/outputs/showcase_sc_2f1a.../showcase.mp4

Add the MCP server in 10 seconds:

claude mcp add tachidubb python /path/to/tachidubb/tools/tachidubb_mcp.py

Or paste into ~/.claude.json:

{
  "mcpServers": {
    "tachidubb": {
      "command": "/path/to/tachidubb/venv/Scripts/python.exe",
      "args": ["/path/to/tachidubb/tools/tachidubb_mcp.py"],
      "env": { "TACHIDUBB_URL": "http://localhost:8910" }
    }
  }
}

The repo ships a Claude Code skill at .claude/skills/tachidubb/SKILL.md. Copy it to ~/.claude/skills/ and Claude knows when and how to drive the pipeline.

# Single language, blocking
python tools/tachidubb_cli.py dub https://youtu.be/abc --lang fr --wait

# Compare 5 languages side-by-side
python tools/tachidubb_cli.py compare ./clip.mp4 --langs es,fr,de,ja,pt --trim 60

# Stitched multilingual showcase reel
python tools/tachidubb_cli.py showcase https://youtu.be/abc \
  --langs es,fr,de,ja,pt --trim 60 --wait

# Re-dub an existing job into new languages — skips re-upload
python tools/tachidubb_cli.py redub 5038e404 --langs ja,it --mode showcase --wait

# Health, status, history
python tools/tachidubb_cli.py system
python tools/tachidubb_cli.py jobs --limit 20
python tools/tachidubb_cli.py status <job_id>

Drive a remote box: set TACHIDUBB_URL=http://192.168.0.10:8910

See examples/ for ready-to-run scripts.

📺 Watch the full demo (no audio, ~2 min) — submit a YouTube URL, pick 5 languages, get a stitched showcase reel.

YouTube URL or local file
        │
        ▼
   yt-dlp ───────────────────────► (downloads source)
        │
        ▼
   FFmpeg ───────────────────────► (extracts audio)
        │
        ▼
  faster-whisper ───────────────► (transcript + word timestamps)
        │
        ▼
   pyannote ─────────────────────► (speaker diarization, optional)
        │
        ▼
   Ollama (Qwen3 / Gemma3 / Aya) ► (translation, length-matched)
        │
        ▼
   VoxCPM2 ──────────────────────► (voice cloning per speaker, 48 kHz)
        │
        ▼
   FFmpeg ───────────────────────► (time-align, mix bg music, render)
        │
        ▼
   Dubbed MP4 + SRT subtitles

Every step is modular, swappable, and runs on your hardware.

28 target languages out of the box (via VoxCPM2 + edge-tts fallback):

Source detection is automatic (Whisper). Translation goes through whatever Ollama model you have — aya-expanse:8b is the default for best multilingual quality.

No GPU? It still runs — just expect long jobs. The pipeline auto-falls back to edge-tts (Microsoft cloud TTS) if VoxCPM2 won't load, which sacrifices voice cloning but produces intelligible output fast.

Total for full setup: ~18 GB. Skinny single-language setup without diarization or BGM preservation: ~12 GB.

Required tokens: NONE. The default install runs 100% offline once dependencies are downloaded. No OpenAI / ElevenLabs / Anthropic key needed — translation is local (Ollama), TTS is local (VoxCPM2), ASR is local (Whisper).

What "phones home" by default:

• yt-dlp reaches YouTube/Vimeo/etc. — only when you submit a URL
• huggingface.co for model downloads — first run only, then cached
• ollama.com for translation model pulls — first install only
• edge-tts for the cloud TTS fallback — only triggers if VoxCPM2 fails to load on your GPU

There's no telemetry, no analytics, no phone-home from TachiDUBB itself. Audit the network calls: search the repo for httpx. / requests. — only the integrations above.

Copy .env.example to .env and edit as needed:

# Speaker diarization (multi-speaker videos)
HF_TOKEN=hf_xxxxx                  # from huggingface.co/settings/tokens

# TTS model selection
VOXCPM_MODEL=openbmb/VoxCPM2       # or openbmb/VoxCPM1.5 (lighter)
VOXCPM_CFG=2.0                     # 1.5-3.0, higher = closer to reference voice
VOXCPM_STEPS=10                    # 5-20, lower = faster

# Translation backend
OLLAMA_URL=http://localhost:11434

# UI behavior
TACHIDUBB_OPEN_BROWSER=1           # 0 to disable auto-open
TACHIDUBB_QA_THRESHOLD=0.4         # stricter (lower) = more re-rolls on bad TTS

If you use Claude Code, copy .claude/skills/tachidubb/SKILL.md into your global skills folder (~/.claude/skills/tachidubb/). After that, just say:

• "Dub this YouTube short into French and German"
• "Make a showcase reel of this clip in 5 languages"
• "Re-dub job 5038e404 into Japanese and Italian"
• "What's the status of my dub?"

The skill teaches Claude which tool to call, what arguments to use, how to poll, how to recover from errors, and when to suggest a comparison vs a showcase. Read SKILL.md for the full trigger map.

Works with any MCP-compatible agent — Cursor, Cline, Continue, custom agents. The MCP tool schema is auto-discovered.

Is this really free? Yes. MIT licensed. The only "cost" is your electricity and GPU. No telemetry, no phone-home.

Do I need an NVIDIA GPU? For reasonable speeds, yes. CPU works but a 1-minute dub takes ~30 minutes instead of ~2.

Does it work on Apple Silicon (M1/M2/M3)? Yes via CPU + MPS fallback. Expect about 4-8× slower than a discrete GPU. PyTorch MPS support for VoxCPM2 is experimental — edge-tts fallback is reliable.

Can I voice-clone a specific person? Yes — drop a 5-30 second clean WAV/MP3 into presets/voices/ and pick it as the reference. Please don't do this without that person's consent. See SECURITY.md.

What's the quality vs ElevenLabs? On clean source audio, VoxCPM2 is genuinely close. On noisy / multi-speaker content, ElevenLabs still wins (their diarization is better). For 95% of one-speaker YouTube content, you won't tell the difference.

Does it preserve emotion / tone? Partially. VoxCPM2 picks up energy and pacing from the reference. It doesn't model fine emotion the way some closed models do. If the source is a calm explainer, the dub is calm; if it's a hype reel, the dub is hype.

Can I run multiple dubs in parallel? The server queues GPU work serially (one VoxCPM2 invocation at a time) to avoid OOM. CPU stages (download, transcribe with CPU Whisper, ffmpeg) overlap automatically.

Does it work for animated content / games / non-real voices? Yes — anything VoxCPM2 can fit as a reference (usually 5+ s of clean speech) clones fine. Singing is not supported.

Why VoxCPM2 instead of XTTS / OpenVoice / F5-TTS? VoxCPM2 has the best cross-lingual cloning quality we tested at the 5 GB weight class. The architecture is swappable — pipeline/synthesizer.py has a base class; PRs for other backends welcome.

Can agents trigger this without my approval? Each MCP tool call requires user confirmation by default (per the MCP spec). Tachidubb doesn't bypass that.

• MCP server + CLI
• Stitched multilingual showcase reels
• Multi-speaker diarization
• Background music preservation
• Deterministic voice across cross-lingual segments
• Subtitle burn-in toggle (currently SRT sidecar only)
• Speaker labelling UI (assign names to detected speakers)
• Browser-only mode (no Ollama dependency, use llama.cpp WASM)
• Batch processing folder watcher
• Docker image with everything pre-baked
• Hardware-accelerated diarization (NVIDIA NeMo)
• Apple Silicon MLX backend

Vote / suggest features in Discussions.

Voice cloning is powerful and easily misused. TachiDUBB is built for legitimate creators dubbing their own content or content they have rights to. Please:

• Don't clone someone's voice without their explicit, informed consent.
• Don't impersonate real people (politicians, celebrities, your boss) for deception, fraud, or harassment.
• Disclose AI-generated speech when publishing — most platforms now require this, and it's the right thing to do.
• Comply with your local laws on synthetic media (EU AI Act, US state laws, etc.).

We refuse to add features that defeat watermarking, anti-cloning safeguards, or platform AI-disclosure requirements. See SECURITY.md for the threat model and how to report abuse.

PRs welcome. See CONTRIBUTING.md for setup, code style, and the modular pipeline design — most contributions are a single drop-in file in pipeline/.

Good first issues:

• Add a TTS backend (XTTS, F5-TTS, OpenVoice)
• Add a translation backend (OpenAI-compatible HTTP, vLLM, mlx_lm)
• New language voices in the edge-tts fallback map
• Improve the duration-matching prompt for hard language pairs

Built by TachikomaRed and smolemaru — in collaboration with Claude (Anthropic).

Follow the build on X: @smolekoma · @smolemaru

Standing on shoulders:

• VoxCPM2 — voice cloning TTS (Apache-2.0)
• faster-whisper — ASR (MIT)
• pyannote.audio — diarization (MIT)
• Ollama — local LLM serving (MIT)
• yt-dlp — universal downloader (Unlicense)
• edge-tts — cloud TTS fallback (GPL-3.0)
• audio-separator — stem separation (MIT)
• Model Context Protocol — agent integration (Anthropic)

MIT — see LICENSE. VoxCPM2 is Apache-2.0. edge-tts is GPL-3.0; using it doesn't require this project to be GPL because it's a runtime dependency invoked as a process.