NVIDIA expands Magpie TTS to 12 languages with open weights for real-time voice agents

NVIDIA published an updated release of Magpie TTS Multilingual on Hugging Face on August 10, 2026, expanding its open-weights text-to-speech model to 12 languages and publishing latency and accuracy numbers aimed squarely at teams building real-time, cascade-style voice agents.

What's new

The 364-million-parameter model now covers English, Spanish, French, German, Italian, Vietnamese, Mandarin, Hindi, Japanese, Modern Standard Arabic, Korean, and Brazilian Portuguese, adding the last three languages in this release. NVIDIA also reports quality gains on existing languages from updated training data and architecture changes — French character error rate dropped from 2.70% to 1.54%, and Spanish improved from 1.14% to 0.60%.

On throughput, NVIDIA published per-GPU figures for the release: on a B200 GPU, time to first audio is 32 milliseconds with a real-time factor of 319.81x across 64 concurrent streams; on H100, 47ms TTFA at 290.79x RTFX; on A100, 79ms TTFA at 197x RTFX. The model ships as open weights under the NVIDIA Open Model License, alongside a production NVIDIA NIM inference container, with live demos on NVIDIA Build and Hugging Face Spaces.

NVIDIA frames the release around the deployment constraints that come with running voice agents at scale: "Open models change what's possible" for teams that need to control latency, cost, and customization rather than call a hosted API. The model is designed as a speech-generation layer that plugs into an existing pipeline — turning an LLM's text output into real-time audio in a cascade voice-agent architecture — rather than a standalone conversational product.

Context

Magpie TTS is part of NVIDIA's broader open-model push for building blocks in agentic and voice AI, alongside its NIM microservices and NeMo tooling. Multilingual TTS has been a fast-moving area this year as labs compete on both raw voice quality and the latency needed to feel conversational in real time — NVIDIA is positioning Magpie specifically on the low-latency, self-hosted end of that spectrum rather than chasing maximal expressiveness.

Why it matters

Most competitive real-time TTS today is only available as a hosted, closed API, which forces teams building voice agents to accept a vendor's latency, pricing, and data-handling terms. An open-weights model with published sub-50-millisecond first-audio latency on mainstream data-center GPUs gives teams a self-hosted alternative for the specific case that matters most in voice UX — how quickly the agent starts talking back — without giving up control over where the model runs or what happens to the audio data.