KOOMPI TTS

Long-form Khmer text-to-speech powered by VoxCPM2. Generates natural, humanized audio of any length by splitting scripts at Khmer sentence boundaries, synthesizing each chunk, then stitching everything together with FFmpeg.

Live at: https://huggingface.co/spaces/koompi/koompi-tts

How it works

Input text
    ↓
split_text()          # Khmer sentence boundaries (།), max 500 chars/chunk
    ↓
VoxCPM2 API           # gradio_client → HF Space
  • chunk 1: voice description prompt → locks reference WAV
  • chunk 2+: controllable cloning from reference → consistent voice
    ↓
FFmpeg concat          # 400ms silence between sentences
    ↓
output.wav

Deploy

hf upload koompi/koompi-tts space/ . --repo-type space

Local dev

pip install -r space/requirements.txt
# Also requires ffmpeg on PATH
cd space && python app.py

Project structure

space/          HuggingFace Space (Gradio app)
  app.py        UI + generation logic
  requirements.txt
src/            CLI utilities (chunker, mixer, client)
generate.py     CLI entry point
examples/       Sample scripts