aiexpert
Home / News / Brief
Chips · Aug 10, 2026, 04:35 PM · 4 sources

NVIDIA releases Magpie TTS Multilingual with 12-language support, open weights for on-prem voice agents

<cite index="64-1">NVIDIA released Magpie TTS Multilingual on August 10, 2026, providing open-weights multilingual text-to-speech for voice agents and brand ambassadors</cite>. <cite index="61-1,61-2">The 364M-parameter model synthesizes speech across 12 languages—Arabic, Chinese, English, French, German, Hindi, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese—using transformer encoder-decoder architecture with multi-codebook prediction, classifier-free guidance, and Group Relative Policy Optimization</cite>. <cite index="65-2">Version 2602 added support for Hindi and Japanese with expanded code-switching capabilities</cite>.

<cite index="65-3,70-2">Magpie supports both streaming and offline inference; it generates speech by predicting discrete audio codec tokens via transformer encoder-decoder, with a downstream audio codec converting tokens to audible waveform</cite>. <cite index="61-3">The model plugs into existing AI pipelines for cascade voice-agent setups, where it converts LLM text into natural, real-time speech for user playback, without changing upstream language models or downstream audio handling</cite>. This enables low-latency multilingual voice agents that developers can deploy and tune on their own infrastructure.

<cite index="70-2,70-3">The NIM (NVIDIA Inference Microservice) version provides production latency: Magpie achieves natural German voice output with no direct playback required, and can handle long-form inputs via streaming mode or batch inference</cite>. For practitioners building multilingual voice agents, Magpie's open-source foundation + production NIM stack eliminates lock-in to managed speech APIs, enabling on-premises deployment with data residency guarantees and latency predictability—a critical control surface for voice-AI application builders targeting sub-200ms end-to-end response windows.

Sources

Everything this brief rests on
  1. 01 Primary source huggingface.co
  2. 02 huggingface.co huggingface.co “Magpie TTS Multilingual open-weights model; 364M parameters; supports 12 languages including new Arabic, Korean, Portuguese”
  3. 03 huggingface.co huggingface.co “B200 achieves 32ms time-to-first-audio (TTFA); 239ms TTFA at 64 concurrent streams with 320× real-time throughput”
  4. 04 huggingface.co huggingface.co “Cascaded architecture: deployed on-prem via NIM; frame stacking + local transformer optimize real-time latency”