Chatterbox fine-tuning probes new-language limits

Five hours from one speaker is enough to imitate timbre, but not enough to guarantee a clean new-language model. If the language is already supported, Chatterbox's zero-shot path is probably the better bet; if it isn't, expect a real adaptation project rather than an instant win.

• –The model card says Chatterbox Multilingual covers 23 languages and warns that mismatched reference clips can leak accent into the output, so data alignment matters as much as duration. [Hugging Face model card](https://huggingface.co/ResembleAI/chatterbox)
• –The GitHub README splits the family into English-only Turbo, 23+ language Multilingual, and English Chatterbox; it doesn't spell out a fine-tuning workflow, which suggests unsupported-language adaptation is DIY. [GitHub README](https://github.com/resemble-ai/chatterbox)
• –Cross-lingual TTS research shows speaker identity can transfer with very little adaptation data, but pronunciation quality still depends on the target language and the backbone's multilingual coverage. [arXiv paper](https://arxiv.org/abs/2111.09075)
• –For an unsupported language, five clean hours from one speaker is a decent prototype budget, but expect accent leakage and uneven prosody unless you add transcripts, phonemization, and a tight eval loop.
• –The Product Hunt launch for Chatterbox Turbo reinforces the family's inference-first positioning around speed, expressiveness, and watermarking.