Nvidia Magpie starts speech in 32 milliseconds

Nvidia’s 364-million-parameter Magpie TTS now supports 12 languages and starts one-stream speech in a reported 32 milliseconds on a B200. The open-weight release adds Arabic, Korean, and Brazilian Portuguese, while Nvidia reports 47 milliseconds to first audio on an H100 and 79 milliseconds on an A100. For builders weighing the self-hosted path described in Omilia’s hybrid voice-AI pilot, Magpie makes the speech layer a measurable slice of the latency budget rather than an opaque API delay.