NVIDIA Magpie TTS adds 12-language support for low-latency voice agents
NVIDIA has released an updated Magpie TTS Multilingual text-to-speech model with open weights, support for 12 languages and deployment options designed for controlled, low-latency voice applications. The 364M-parameter model adds Modern Standard Arabic, Korean and Brazilian Portuguese, alongside English, Spanish, French, German, Italian, Vietnamese, Mandarin, Hindi and Japanese. It provides male and female voices through a shared multilingual speaker representation.
The release also expands code-switching support for Hindi and Japanese, using IPA grapheme-to-phoneme processing and pronunciation dictionaries. NVIDIA reports quality improvements in several existing languages; for example, French character error rate fell from 2.70% to 1.54%, while speaker similarity rose from 0.703 to 0.747.
Performance measurements for the NVIDIA NIM serving container, run on-premises, show Time to First Audio of 32–79 ms on a single stream across B200, H100, DGX Spark and A100 GPUs. On B200, 64 concurrent streams reached 239 ms TTFA and 319.81× real-time throughput. The model uses frame stacking and a local transformer to reduce decoding time while preserving speech quality. Developers can run it on their own infrastructure, fine-tune it with NeMo and integrate it with NVIDIA voice-agent components.
Written by Sakhanda Wire AI from reporting by Hugging Face. Facts, quotes and full context are in the original article below.