Context

The company wanted to reuse short Douyin and TikTok videos for its markets abroad. Doing that by hand means downloading, transcribing, translating, recording a voice-over that fits the original timing and remuxing, for every video and every language.

What I did

I built this alone (all commits in the repository are mine).

• Restructured the code into a package with separate config, downloaders, services and media modules, and one workspace folder for all output.
• Extended the pipeline from Vietnamese only to 12+ target languages, including Taglish, Japanese, Korean, Indonesian and Swiss variants of German, French and Italian.
• Loaded the voice list from Edge-TTS at runtime and added a fallback when a locale has no voice (for example it-CH falls back to it-IT).
• Added an all-in-one tab that runs every step automatically, a FastAPI web UI next to the CustomTkinter desktop app, and fixed a 44.1 kHz audio resampling bug.

How it works

• Download the source video from Douyin or TikTok.
• Transcribe locally with Whisper into an SRT file.
• Translate with Gemini, keeping a one-to-one mapping to the original timeline.
• Synthesize speech with Edge-TTS, adjusting speaking rate and grouping short lines so the voice fits each segment.
• Mux the new audio with FFmpeg; optionally run Wav2Lip and GFPGAN for lip-sync.

Results

• 12+ target languages, up from one.
• Two interfaces: desktop (CustomTkinter) and web (FastAPI).
• Each step starts automatically when the previous one succeeds, and a running job can be cancelled.

What I'd do next

I did not measure output quality: there is no translation accuracy or voice-quality score, and the project has only one test file. Before calling it production-ready I would add tests for each pipeline stage and a small human-rated sample per language to see where timing or translation breaks down.