PiperBot

Discord TTS bot backed by the PiperHub voice catalog (~500 community Piper voices). TypeScript + discord.js, synthesis via the pipertts native module (persistent in-process ONNX session).

Commands

• /infer run - Synthesize speech: model (autocomplete), text (empty = demo text in the model's language), length_scale, noise_scale, noise_w, speaker (autocomplete for multi-voice models), sentence_silence, volume. A 📊 Stats button on the result reveals download/inference timings (ephemeral).
• /infer help - Parameter reference.
• /models - List and filter voices by language, name, quality.
• /dataset list|sources - Browse companion LJSpeech datasets.
• /help - Command overview.

Inference requests run through a FIFO queue (one synthesis at a time).

How it works

• Model metadata comes from models_mapping.csv / datasets_mapping.csv in the parent crawler repo (expected at ~/huggingface-crawler/).
• Model files (.onnx + .onnx.json) are downloaded from Hugging Face via pipertts hub resolution into tmpfs (/dev/shm/piperbot, RAM only, never disk).
• Voices stay cached for 5 minutes idle, then files are evicted. Warm inference is ~10x faster than spawning the Piper CLI per call (~260 ms vs ~2.5 s).
• Gated Hugging Face repos are not supported (pipertts downloads without auth).

Run

npm install
npm run build
# DISCORD_TOKEN in ~/huggingface-crawler/.env or environment
node dist/index.js

pm2 example: pm2 start ecosystem.config.cjs

License

MIT