README.md

metadata

title: Tool Generate Text2Speech
emoji: 🗣️
colorFrom: blue
colorTo: purple
sdk: gradio
app_file: app.py
pinned: false
license: apache-2.0

TTS Dataset Generator (VoxCPM2)

Batch text-to-speech tool for preparing paired text/audio datasets to train ASR and TTS models — built on openbmb/VoxCPM2.

What it does

1. Upload a JSON file shaped like:

   [
     {"id": 1, "text": "សួស្តី! តើថ្ងៃនេះអ្នកសុខសប្បាយជាទេ?"},
     {"id": 2, "text": "..."}
   ]
   

   (up to 100 items per batch)
2. Optionally describe a voice (e.g. a calm young woman, clear and steady voice) — VoxCPM2's Voice Design applies it consistently across every item.
3. Click Generate batch. Progress and per-item status show live in the table.
4. Download the resulting .zip:

   tts_dataset.zip
   ├── audio/
   │   ├── 1.wav
   │   ├── 2.wav
   │   └── ...
   ├── metadata.csv     # id, text, filename, status
   └── metadata.json
   

Deploying this Space

This runs as a standard Gradio SDK Space (not Static / gradio-lite) because VoxCPM2 is a 2B-parameter PyTorch model — it needs real Python + (ideally) GPU execution, which a browser-only static Space cannot provide.

Push these three files (app.py, requirements.txt, README.md) to the Space repo:

git clone https://huggingface.co/spaces/phonsobon/Tool_Generate_Text2Speech
cd Tool_Generate_Text2Speech
cp /path/to/app.py /path/to/requirements.txt /path/to/README.md .
git add .
git commit -m "Batch TTS dataset generator (VoxCPM2)"
git push

Hardware notes

• Running on free CPU basic: generation is slow (roughly tens of seconds per sentence for a 2B-param model). Test with the "Limit" field set low (e.g. 5) before running a full batch of 100.
• If you later upgrade to HF PRO you can switch this Space to ZeroGPU hardware in the Space settings for a large speed-up — no code changes needed, just add a @spaces.GPU decorator around the generation call.
• Khmer is one of VoxCPM2's 30 officially supported languages.