NVIDIA Magpie TTS: Open-Weight Multilingual Voice Generation with Full Deployment Control
NVIDIA
NVIDIA releases Magpie TTS Multilingual, an open-weights model supporting 12 languages, including new additions like Modern Standard Arabic, Korean, and Brazilian Portuguese. The model offers low latency via frame stacking and local transformer techniques, with deployment flexibility through NVIDIA NIM and Hugging Face checkpoints.
NVIDIA introduces Magpie TTS Multilingual, a 364M-parameter open-weights text-to-speech model supporting 12 languages, with new additions of Modern Standard Arabic, Korean, and Brazilian Portuguese. The model is designed for low-latency voice agents, achieving a time-to-first-audio (TTFA) of 32ms on B200 GPUs and offering high throughput. Architectural improvements include frame stacking to reduce decoder iterations and a local transformer to maintain audio quality. The release highlights improved quality metrics, such as reduced character error rates (CER) and higher speaker similarity (SSIM) for French and Spanish. Developers can deploy the model on their own infrastructure using NVIDIA NIM or Hugging Face checkpoints, customize with NVIDIA NeMo, and integrate with the Nemotron Voice Agent Developer Example for building complete voice agents.
- Abbreviations
- TTS = Text-to-Speech — синтез речи
- TTFA = Time to First Audio — время до первого аудио
- CER = Character Error Rate — коэффициент символьных ошибок
- SSIM = Structural Similarity Index Measure — индекс структурного сходства
- ASR = Automatic Speech Recognition — автоматическое распознавание речи
- LLM = Large Language Model — большая языковая модель
- NIM = NVIDIA Inference Microservice — микросервис инференса NVIDIA
Source: Hugging Face blog —
original