Backends

Audio version — Estimated duration: 2 min 37 sec

Both the LLM (text rewriting) and TTS (audio synthesis) stages have independent online and local backends.

LLM backends

Online — Cerebras API

• Model: gpt-oss-120b (configurable via --model)
• Requires: CEREBRAS_API_KEY environment variable
• Rate limit: 5 requests per 60 seconds (enforced client-side)

Local — llama.cpp

• Expects: server at http://localhost:8080/v1
• Any model compatible with llama.cpp server (GGUF format)
• No rate limiting
• Set --model to any value (passed through but mainly cosmetic)

TTS backends

Online — edge-tts

• Microsoft’s cloud TTS via edge_tts Python library
• Output: .mp3
• Voice: any Microsoft neural voice (default en-US-JennyNeural)
• Zero setup beyond pip install edge-tts

Local — Piper

• GPU-accelerated via onnxruntime
• Output: .wav (16-bit mono PCM)
• Models auto-downloaded from Hugging Face on first use
• Default voice: en_US-hfc_female-medium
• --cpu flag to force CPU inference

Piper model download

Models are cached in models/:

models/en_US-hfc_female-medium.onnx
models/en_US-hfc_female-medium.onnx.json