NVIDIA releases MagpieTTS Multilingual 357M v2607, 12 languages with preset voices

NVIDIA's MagpieTTS Multilingual 357M reached version v2607 on 21 July 2026, per the model card, after earlier tags v2602 (March 2026) and v2512 (January 2026). It is a 364M-parameter encoder-decoder transformer covering 12 languages: Arabic, Chinese, English, French, German, Hindi, Italian, Japanese, Korean, Portuguese, Spanish and Vietnamese. It ships with five preset English speaker voices that speak all 12 languages; four are proprietary and one is public. Output is 22.05 kHz mono.

The card says zero-shot voice cloning was removed in this release "for security reasons". It adds built-in text normalisation, IPA phoneme input, long-form synthesis with a sliding window, and batched synthesis. Reported character error rates for v2607 include 0.37 percent for English, 0.60 for Spanish, 1.54 for French, 3.17 for Chinese and 2.91 for Portuguese.

The licence is the NVIDIA Open Model License, which the card says is ready for commercial use. Standard mode generates up to 20 seconds at a time, and the model is not suitable for languages outside the 12 listed.

A model that gets safer by losing a feature. At least the card says so plainly.

Sources

• NVIDIA model card on Hugging Face

Editor's first pass (AI-assisted)

• Real 4.5: Weights under the NVIDIA Open Model License, which the card says is ready for commercial use.
• Useful 3.5: Twelve languages and phoneme input suit products, but no cloning and only five voices.
• Notable 3.5: NVIDIA publishes per-language error rates in the card.
• Novel 2.5: A versioned update; the notable change is cloning being removed.