Show HN: Parkiet – Fine-tune a large TTS model for any language under $100 (github.com)

Parkiet is an open-weights Dutch TTS system (Parakeet-based) ported from Dia to JAX with a full TRAINING.md walkthrough showing how to fine-tune the model for any language affordably on Google Cloud TPUs (author claims under one hundred dollars). The repo includes pretrained checkpoints on Hugging Face, demo scripts (JAX and PyTorch), and samples demonstrating multi‑speaker dialogues (up to four speakers), emotion/tone control, nonverbal sounds (laughter), stuttering/disfluencies, and low-data voice cloning. Guidance on prompt formatting (always start with [S1], alternate speakers, prefer lowercase and punctuation, use "..." to slow, and tags like (laughs)) helps reduce hallucination and keeps output consistent; the author also provides a human-comparison blog against ElevenLabs. Technically, the JAX port gives the best audio quality and scales on TPUs, though it requires extra setup and has longer first-run compile times; a PyTorch conversion exists but reportedly hallucinates more and shows artefacts due to small attention-kernel differences. The repo includes quick-start commands for downloading weights and running inference (uv sync/run workflows) and a complete TPU training guide. The project is targeted at research and education and explicitly bans identity misuse, deceptive or illegal applications — a reminder that democratized, low-cost expressive TTS increases both access and responsibility for the AI community.