Korean TTS Workbench — (1) Starting With Qwen3-TTS on a 4GB GPU
A solo side project that ran from planning to cloning to model retirement — Part 1 of 5
Apr 27, 202615 min read3
Command Palette
Search for a command to run...
Series
Korean TTS Workbench
A 5-part series on building, troubleshooting, and eventually retiring a Korean voice-cloning TTS workbench.
Korean TTS Workbench — (2) Korean-Specific Tricks and Running on 4GB VRAM
From phone-number normalization to seed determinism gotchas — the small decisions behind running the workbench. Part 2 of 5
Apr 28, 202617 min readKorean TTS Workbench — (3) The Last Syllable Sounds Cut Off — and the Cause Was the Model Itself
Tracing a one-line symptom — the last syllable getting cut off — and why diagnostic logging paid back more than any single fix. Part 3 of 5
Apr 29, 202616 min read2Korean TTS Workbench — (4) When the Workarounds Failed, a Hardcoded Line Inside the Library
Tail anchor and EOS suppression both collapsed; what surfaced inside the library was a single hardcoded line — and patching it caused a runaway. Part 4 of 5
Apr 30, 202617 min read4Korean TTS Workbench — (5) Qwen3-TTS → VoxCPM2, Swapping the Model Out
The reasoning behind retiring the project, the migration to a new model, and five retrospectives from the troubleshooting arc. Final part of 5
May 6, 202614 min read3