End-to-end video dubbing pipeline. Download a video, transcribe it, translate the subtitles, clone a voice, and produce a fully dubbed audio or video — in one command.
Prerequisites: Python 3.10+ and ffmpeg on your PATH (apt install ffmpeg / brew install ffmpeg).
⚠️ Useuv— plainpipwill fail.mazinger[all]bundles packages with conflicting dependency pins that onlyuvcan resolve. Installuvonce:pip install uv
uv pip install "mazinger[all]"mazinger web --with-ollama --with-coherexA local URL opens in your browser. Paste a video link, pick a voice, and click Start. The flags install a free local LLM (Ollama) and pre-download CohereX, the transcription backend Studio defaults to — no API keys required.
No Hugging Face account? CohereX's weights are gated, so swap in
--with-faster-whisperand pick Faster Whisper in the UI — it needs no sign-in and covers any language.
The flags only move the wait to launch time: if you run plain
mazinger web, Studio installs Ollama and pulls the model by itself the first time you start a mission with the Ollama (Local — Free) provider. SetOLLAMA_HOSTto use an Ollama server running elsewhere.
💡 No GPU? Run it on a free Colab T4 in two clicks:
Prefer the command line or Python? Skip ahead to Common Tasks or the Python API.
mazinger[all] is a single, GPU-friendly install that includes everything needed for the full pipeline and the Studio web UI:
| Capability | Engine |
|---|---|
| Local transcription — 14 languages + a dedicated Arabic model | CohereX — Cohere Transcribe + wav2vec2 (Studio default) |
| Local transcription — any language, no sign-in | Faster Whisper (CLI default) |
| Cloud transcription (no GPU) | Deepgram Nova 3 — $200 free credit |
| Voice-cloned TTS | Qwen3-TTS, OmniVoice (24 languages) |
| Background-audio separation | Demucs |
| Web UI | Gradio (Mazinger Studio) |
| Local LLM (optional) | Ollama (auto-installed by mazinger web) |
CohereX needs a one-time Hugging Face sign-in — the Cohere models are gated. In the Studio, open 🤗 Hugging Face → Sign in with Hugging Face; from the CLI, set
HF_TOKEN. Not signed in, or working in a language CohereX does not cover? Switch the transcription method to Faster Whisper, which needs neither.
Need Chatterbox, MLX (Apple Silicon), or a lighter install? See the Installation Guide — mazinger[all] now bundles CohereX, which brings the pyannote stack with it (~50 extra packages).
No voice files needed. Mazinger picks the best 20–60 s of the source as the cloning reference.
mazinger dub "https://youtube.com/watch?v=VIDEO_ID" \
--target-language Spanish16 built-in themes — no files, no profile downloads.
mazinger dub "https://youtube.com/watch?v=VIDEO_ID" \
--voice-theme narrator-m \
--target-language SpanishThemes: narrator-m/f · young-m/f · deep-m/f · warm-m/f · news-m/f · storyteller-m/f · kid-m/f · teen-m/f
mazinger dub "https://youtube.com/watch?v=VIDEO_ID" \
--clone-profile abubakr \
--target-language ArabicAvailable out-of-the-box: abubakr · daheeh-v1 · 3b1b · italian-v1 · morgan-freeman · trump-v1 — full list in Voice Profiles.
mazinger dub "https://youtube.com/watch?v=VIDEO_ID" \
--voice-sample speaker.m4a \
--voice-script speaker_transcript.txt \
--target-language Spanishmazinger dub "https://youtube.com/watch?v=VIDEO_ID" \
--voice-theme narrator-m \
--target-language Arabic \
--output-type video \
--embed-subtitles \
--subtitle-google-font "Noto Sans Arabic"See Subtitle Styling for fonts, colors, positioning, and RTL options.
14 languages with word-level alignment, plus a dedicated Arabic model that is
selected automatically for Arabic sources. Included in mazinger[all], and the
default in Mazinger Studio — nothing to install separately.
The Cohere models are gated on HuggingFace, so it needs a one-time sign-in. In
the Studio, open 🤗 Hugging Face → Sign in with Hugging Face and follow the
link it shows — no token to copy. From the CLI, set HF_TOKEN instead.
The CLI still defaults to Faster Whisper; pass --transcribe-method coherex to
use CohereX there:
export HF_TOKEN=hf_your_token
mazinger dub "https://youtube.com/watch?v=VIDEO_ID" \
--transcribe-method coherex \
--source-language Arabic \
--voice-theme narrator-m \
--target-language English
⚠️ Cohere Transcribe cannot detect the source language on its own. Given one, it transcribes confidently in whatever language you name; given none, it probes each supported language first, which is slower and less reliable. Always pass--source-language(or pick one in the Studio).
export DEEPGRAM_API_KEY=your_key_here
mazinger dub "https://youtube.com/watch?v=VIDEO_ID" \
--transcribe-method deepgram \
--voice-theme narrator-m \
--target-language Englishmazinger download "https://youtube.com/watch?v=VIDEO_ID"
mazinger transcribe ./output/projects/my-video/source/audio.mp3 -o subs.srt
mazinger translate --srt subs.srt --target-language French -o translated.srt
mazinger subtitle video.mp4 --srt translated.srt -o output.mp4Every stage caches its output. Re-running resumes where it stopped. Full command list in the CLI Reference.
from mazinger import MazingerDubber
dubber = MazingerDubber(openai_api_key="sk-...", base_dir="./output")
proj = dubber.dub(
source="https://youtube.com/watch?v=VIDEO_ID",
voice_theme="narrator-m",
target_language="Spanish",
output_type="video",
)
print(proj.final_video) # ./output/projects/<slug>/tts/dubbed.mp4Full reference: Python API.
Mazinger chains ten resumable stages: Download → Transcribe → Thumbnails → Describe → Review → Translate → Re-segment → Speak → Assemble → Subtitle. Every stage runs standalone or as part of the full pipeline; completed stages and individual TTS segments are cached and skipped on re-runs.
See the Pipeline Overview for a diagram and the data flow between stages.
| Topic | What's inside |
|---|---|
| Installation | All install options, advanced extras, Apple Silicon, Colab, uv overrides |
| Quick Start | More copy-paste workflows |
| Pipeline Overview | The ten stages, data flow, resume behavior |
| CLI Reference | Every command, flag, and default |
| Python API | Classes, functions, parameters |
| Voice Profiles | Using, creating, uploading profiles |
| Subtitle Styling | Fonts, colors, positioning, RTL, Google Fonts |
| Configuration | Env vars, caching, tempo, LLM usage tracking |
| Project Structure | Output directory layout |
| YouTube Cookies | Cookies for age-restricted / region-locked videos |
MIT