NVIDIA Magpie TTS Adds 12-Language Voice AI With 32 ms Latency
NVIDIA Magpie TTS adds Arabic, Korean and Brazilian Portuguese, open weights, 12-language support and 32 ms first audio on B200 GPUs for private voice agents.
Summary
Released on Hugging Face on August 10, 2026, NVIDIA Magpie TTS Multilingual is a 364M-parameter open-weights model under the NVIDIA Open Model License, with production NVIDIA NIM serving on private or air-gapped infrastructure. It supports male and female voices in 12 languages: English, Spanish, French, German, Italian, Vietnamese, Mandarin, Hindi, Japanese, and new Modern Standard Arabic, Korean and Brazilian Portuguese. IPA grapheme-to-phoneme processing and dictionaries expand Hindi and Japanese code-switching and custom pronunciation; NeMo enables voice and vocabulary fine-tuning.
In on-prem NVIDIA TTS NIM v26.07 tests averaged across three trials, one-stream first-audio latency and real-time throughput were 32 ms and 12.1x on B200, 47 ms and 14.7x on H100, 53 ms and 9.8x on DGX Spark, and 79 ms and 12.2x on A100. At 64 streams, B200 delivered 239 ms and 319.81x, H100 275 ms and 290.79x, DGX Spark 962 ms and 75.88x, and A100 395 ms and 197x. NVIDIA says B200's 32 ms leaves room for ASR and LLM work within a sub-200 ms conversational target. Frame stacking predicts two audio frames per decoder step to halve iterations; a local transformer restores codebook-token dependencies, as detailed in Frame-Stacked Local Transformers for Efficient Multi-Codebook Speech Generation, ICASSP 2026.
Against the previous release, French CER fell 2.70% to 1.54% and SSIM rose 0.703 to 0.747; Spanish improved 1.14% to 0.60% and 0.715 to 0.793; German CER worsened 0.66% to 0.80% while SSIM rose 0.626 to 0.742. New CER baselines are Arabic 1.62%, Korean 2.69% and Brazilian Portuguese 2.91%. The cloneable Nemotron Voice Agent example combines Nemotron Speech, Magpie, Nemotron language and multimodal models, NIM and NeMo for barge-in, vision, tools, multi-agent and multilingual operation with claimed sub-second latency. NVIDIA Build and Hugging Face host demos; recommended cfg_scale is 2.5, temperature 0.6, top_k 80, apply_attention_prior True and prior_epsilon 0.1.
Positives
- Magpie consolidates 12 languages and male and female voices in one 364M-parameter open-weights model.
- B200 produced first audio in 32 ms on one stream, the fastest result among the four tested NVIDIA systems.
- French CER fell from 2.70% to 1.54%, while Spanish CER dropped from 1.14% to 0.60%.
- Private and air-gapped NIM deployment gives enterprises control over data residency, latency, scaling and customization.
- The Nemotron Voice Agent example supplies cloneable patterns for barge-in, vision, tool use, multi-agent orchestration and multilingual interactions.
Risks & concerns
- German CER increased from 0.66% to 0.80%, despite SSIM improving from 0.626 to 0.742.
- All 64-stream TTFA results exceeded the cited sub-200 ms conversational window, ranging from 239 ms to 962 ms.
- New CER baselines vary from Arabic's 1.62% to Brazilian Portuguese's 2.91%, leaving room for future quality improvements.
- Performance figures average three on-prem NVIDIA NIM trials, while the Hugging Face checkpoint is positioned for research and fine-tuning.
- Magpie's cascaded approach requires separate ASR, LLM and TTS components, unlike a simpler one-call integrated speech model.