NVIDIA Magpie TTS enhances multilingual voice agents

NVIDIA's Magpie TTS Multilingual model allows for the deployment of voice agents that support 12 languages, including English, Spanish, French, German, Italian, Vietnamese, Mandarin, Hindi, Japanese, and newly added Modern Standard Arabic, Korean, and Brazilian Portuguese. The model provides male and female speaker voices through a shared multilingual representation, enhancing flexibility with code-switching support for Hindi and Japanese.

Magpie TTS can be deployed within users' own infrastructure, allowing them to manage latency and optimize performance directly. Key metrics include a Time to First Audio (TTFA) of 32 to 79 milliseconds for a single stream on NVIDIA GPUs, with B200 reaching 32 milliseconds. For 64 concurrent streams, B200 achieves 239 milliseconds, delivering audio 320 times faster than playback.

The model demonstrates improved synthesis quality with reduced character error rates and higher speaker similarity, particularly for French and Spanish. Open weights enable developers to deploy on private environments, customize pronunciation and voices, and maintain control over sensitive data. Magpie TTS also integrates into the NVIDIA Nemotron Voice Agent Developer Example, providing a comprehensive solution for building voice agents.