NVIDIA Magpie TTS: Open-Weight Multilingual Voice AI with Low Latency

NVIDIA has introduced Magpie Multilingual TTS, a text-to-speech model featuring 364M parameters and open weights, capable of supporting 12 languages: English, Spanish, French, German, Italian, Vietnamese, Mandarin, Hindi, Japanese, Modern Standard Arabic, Korean, and Brazilian Portuguese. This model offers both male and female voice options, enhances code-switching specifically for Hindi and Japanese, and is designed for real-time speech generation. Key techniques include frame stacking and a local transformer, as outlined in the paper 'Frame-Stacked Local Transformers for Efficient Multi-Codebook Speech Generation' (ICASSP 2026). It boasts a Time to First Audio (TTFA) of 32ms on the B200 GPU and is available on Hugging Face and as an NVIDIA NIM container, improving synthesis quality and lowering character error rates for new languages. Magpie is part of the NVIDIA Nemotron Voice Agent Developer Example.