VoxCPM Studio
Voice-cloning narration studio, mastered to podcast standards.
- Role
- Developer
- Context
- Narration for PackTalk Cambodia 2026
- Stack
- Python · PyTorch · VoxCPM2
VoxCPM Studio clones a voice from a short reference recording, reads a timed script with it, and masters the result to podcast standards. I built it on OpenBMB’s open VoxCPM2 model to produce the PackTalk Cambodia 2026 narration — in English and Khmer — fitted exactly to the video’s timing.
What I did
- 01
Built a web studio (Next.js) and a FastAPI service around one shared voice-cloning toolkit, so the interface and the notebook can never drift apart.
- 02
Rendered scripts scene by scene — each fitted to its timecode and seeded — so one bad line can be regenerated without redoing the rest.
- 03
Fitted narration to an exact length: the PackTalk video narration lands at 2:02.4.
- 04
Mastered every render to podcast loudness (−16 LUFS stereo, ≤ −1 dBTP) through high-pass, mud cut, presence, de-ess and air stages.
- 05
Kept the reference voice exactly as recorded, after measuring that resampling it weakened the clone.
- 06
Packaged it with Docker on a GPU; only the web interface is exposed, never the model API.
Highlights
Voice cloning
From 3–20 seconds of clean speech.
Exact timing
Each scene fitted to its timecode.
Podcast master
Loudness and EQ to published practice.
Khmer & English
Same voice, both languages.