How the pipeline actually works.
This is an early set of docs — CLI usage, pipeline stages, execution providers, and the model manifest. A full reference lands with v1.
The CLI runs the same pipeline the desktop app runs.
No separate headless engine, no feature gap — same stages, same models, scriptable for batch localization or CI. Included on every tier, Free included.
trackdub dub --media ./video.mp4 --target-language es
trackdub dub --media ./video.mp4 --target-language de \ --model asr:whisper-small --model tts:kokoro-onnx
trackdub dub --preset my-preset --input-dir ./videos
Re-running the same project skips already-completed stages — only what changed reruns.
Six stages. Each one a real, inspectable artifact.
- 01IngestProbe the container, extract audio, detect scenes and speech.
- 02TranscribeSource-language ASR with word-level timestamps.
- 03TranslatePer-line MT against an editable target script and glossary.
- 04DiarizeCluster speakers, attach a voice reference to each.
- 05VoicePer-speaker zero-shot TTS. Any single line can regenerate on its own.
- 06MixAlign to source timing, duck under preserved music/SFX, mux the export.
Full interactive walkthrough on the homepage.
Fallback is per-stage, not per-project.
A missing provider on one stage doesn't disable the rest of the pipeline — each stage picks the fastest provider your hardware supports and falls back on its own.
Auto-selected on supported GPUs.
Broadest Windows GPU coverage, incl. Intel Arc, AMD Radeon.
Used on non-RTX NVIDIA cards.
Neural Engine + GPU, auto-selected on M-series.
Always available. No GPU required to complete a project.
Every bundled model, declared — not implied.
bundled-models.manifest.json lists every model Trackdub ships: task, license, whether commercial use is allowed, and a checksum. No research-only or non-commercial-only checkpoint ships in any tier.
Full API and SDK reference lands with v1.
Building something against the pipeline now? Write to hello@trackdub.com and we'll loop you in as it ships.