NVIDIA releases Magpie TTS Multilingual, an open-weights text-to-speech model for 12 languages with low-latency deployment

NVIDIA released Magpie TTS Multilingual, an open-weights text-to-speech model designed for low-latency, multilingual voice agents. The model has 364M parameters and supports 12 languages: English, Spanish, French, German, Italian, Vietnamese, Mandarin, Hindi, Japanese, Modern Standard Arabic, Korean, and Brazilian Portuguese. Each language includes male and female speaker voices through a shared multilingual speaker representation.

Magpie TTS is available as an open Hugging Face checkpoint for research and fine-tuning, and via NVIDIA NIM as a tuned serving stack for production deployments. Both run on infrastructure controlled by the user, enabling direct benchmarking, tuning, and scaling for specific workloads.

The model introduces architectural improvements aimed at real-time speech generation. Frame stacking reduces decoder iterations by predicting two audio frames per step, and a local transformer refines generated audio to maintain quality. These changes target Time to First Audio (TTFA), a critical latency metric for conversational AI.

In on-prem benchmarks across NVIDIA GPUs, Magpie TTS achieved TTFA between 32ms (B200) and 79ms (A100) on a single stream. At 64 concurrent streams, the B200 delivered 239ms TTFA with throughput at 320× real time. The reported figures are averages from three trials on NVIDIA TTS NIM Performance documentation (v26.07).

The release expands multilingual coverage with Modern Standard Arabic, Korean, and Brazilian Portuguese, and improves synthesis quality across several existing languages. Compared to the previous release, Magpie shows reduced character error rates (CER) and higher speaker similarity (SSIM) on French, Spanish, and German. Baseline CER for the newly added languages is 1.62% for Arabic, 2.69% for Korean, and 2.91% for Brazilian Portuguese.