NVIDIA Releases Magpie Multilingual TTS Open Weights

NVIDIA has launched Magpie Multilingual TTS, a 364-million-parameter open-weights text-to-speech model designed for low-latency voice applications. The release expands the model's language coverage to 12 languages by adding Modern Standard Arabic, Korean, and Brazilian Portuguese to its existing lineup of English, Spanish, French, German, Italian, Vietnamese, Mandarin, Hindi, and Japanese. It also introduces expanded code-switching capabilities for Hindi and Japanese through custom pronunciation dictionaries and International Phonetic Alphabet processing.

To achieve rapid speech generation, Magpie utilizes frame stacking to predict two audio frames per decoding step, cutting decoder iterations in half. It pairs this with a local transformer to refine the audio and maintain natural speech quality. In benchmark testing, the model achieved a Time to First Audio of 32 milliseconds on a single stream using an NVIDIA B200 GPU, with latency ranging from 32 to 79 milliseconds across various NVIDIA GPUs. Under a heavy concurrent load of 64 streams, the B200 GPU maintained a Time to First Audio of 239 milliseconds while delivering throughput at 320 times real-time speed.

The updated model also improves synthesis quality, showing reduced character error rates and higher speaker similarity, particularly in French and Spanish. The newly added languages establish strong baselines, with Arabic recording a 1.62 percent character error rate, Korean at 2.69 percent, and Brazilian Portuguese at 2.91 percent.

For AI practitioners, these open weights and the accompanying NVIDIA NIM optimized inference containers provide full deployment control. Instead of relying on closed APIs that introduce network round-trip latency, developers can run Magpie entirely on their own private or air-gapped infrastructure. This allows teams to optimize their latency budgets, customize pronunciation, and fine-tune voices using NVIDIA NeMo to match specific brand identities or domain vocabularies.