Hugging Face and NVIDIA introduced Magpie Multilingual TTS, an open-weights model designed to give developers full deployment control and low latency for multilingual voice agents. Spanning 364 million parameters, Magpie supports 12 languages, including English, Spanish, French, German, Italian, Vietnamese, Mandarin, Hindi, Japanese, and newly added Modern Standard Arabic, Korean, and Brazilian Portuguese. Each language provides male and female speaker voices through a shared multilingual representation, alongside enhanced code-switching support for Hindi and Japanese using IPA grapheme-to-phoneme processing and custom dictionaries. Magpie addresses key latency bottlenecks in voice pipelines through two architectural innovations: frame stacking, which predicts two audio frames per decoding step to cut iterations in half, and a local transformer, which models dependencies to recover audio quality. According to performance documentation, Magpie delivers Time to First Audio between 32ms and 79ms on single streams across NVIDIA GPUs, while B200 hardware achieves 239ms TTFA under 64 concurrent streams with throughput at 320 times real time. Available via open Hugging Face checkpoints for research and tuned NVIDIA NIM containers for production, the model aims to help developers build responsive enterprise copilots, customer support agents, and healthcare assistants on their own infrastructure.
Build Low-Latency Multilingual Voice Agents: Open Weights & Full Deployment Control with NVIDIA Magpie TTS
NVIDIA has released Magpie Multilingual TTS, a 364M-parameter open-weights text-to-speech model designed for low-latency voice agents. Supporting 12 languages with male and female voices, the model features frame-stacked local transformers to optimize generation time and maintain natural speech quality on self-hosted infrastructure.
Why it matters
Self-hosted, open-weights speech models allow enterprises to retain control over data residency, domain fine-tuning, and end-to-end latency in real-time voice applications.