NVIDIA has released an update to its Magpie Multilingual TTS model, expanding support to twelve languages with the addition of Modern Standard Arabic, Korean, and Brazilian Portuguese. The release also includes quality improvements across existing languages through updated training data and architectural changes.

  • The 364M-parameter open-weights model now supports English, Spanish, French, German, Italian, Vietnamese, Mandarin, Hindi, Japanese, and the three new languages.
  • Expanded code-switching support for Hindi and Japanese is enabled via IPA grapheme-to-phoneme processing and custom pronunciation dictionaries.
  • Frame stacking and a local transformer reduce inference time while maintaining speech quality.
  • The model achieves Time to First Audio (TTFA) of 32ms on B200 GPUs, leaving latency budget for ASR and LLM processing.
  • Objective metrics show reduced character error rates and higher speaker similarity compared to the previous release.

The update allows developers to deploy multilingual speech synthesis within their own infrastructure, optimizing for low latency, data residency, and domain-specific customization.