NVIDIA Magpie Multilingual TTS is a 364M-parameter open-weights model supporting 12 languages for real-time speech generation. Designed for developers who need control over latency, privacy, and customization, it fits into a cascaded voice pipeline where TTS is the final stage. The model achieves 32 ms Time to First Audio (TTFA) on a B200 GPU for a single stream, and 239 ms TTFA under 64 concurrent streams while sustaining throughput at 320× real time. On H100, DGX Spark, and A100, single-stream TTFA ranges from 47–79 ms. These measurements come from the NVIDIA NIM optimized inference container running on the user’s own infrastructure.

The latest release adds Modern Standard Arabic, Korean, and Brazilian Portuguese, and improves quality across existing languages through updated training data and model refinements. Character error rates dropped notably for French (2.70%→1.54%) and Spanish (1.14%→0.60%), while speaker similarity improved across several languages. The architecture uses frame stacking (decoder predicts two audio frames per step) combined with a local transformer that models dependencies between codebook tokens to maintain naturalness. These techniques are described in the paper “Frame-Stacked Local Transformers for Efficient Multi-Codebook Speech Generation” (ICASSP 2026).

Open weights allow deployment in private or air-gapped environments, fine-tuning with NeMo for domain-specific pronunciation and voices, and independent benchmarking and optimization of each pipeline stage. The model is licensed under the NVIDIA Open Model License. Magpie TTS is also integrated into the NVIDIA Nemotron Voice Agent Developer Example, a reference implementation combining speech recognition (Nemotron Speech), Magpie TTS, language models, and NIM microservices. This example includes patterns for real-time interruptible conversations, multimodal understanding, multi-agent orchestration, and multilingual interaction with sub-second end-to-end latency.

For developers building customer support agents, healthcare assistants, enterprise copilots, or translation systems, Magpie provides a deployable foundation with measurable latency that they control end to end.