AIBRIEF
The 364 million parameter speech synthesis model runs on a team's own hardware rather than as a cloud voice API, with vendor reported sub 50ms time to first audio output on NVIDIA's flagship data center GPUs (B200 and H100).
NVIDIA's Magpie text-to-speech model now supports twelve languages, adding Modern Standard Arabic, Korean, and Brazilian Portuguese to the nine it already covered, the company announced in a Hugging Face blog post.
The 364-million-parameter model ships as open weights under the NVIDIA Open Model License, letting product teams run speech synthesis on their own hardware rather than paying an integrated cloud voice API. A single shared speaker representation handles both male and female voices across all twelve languages.
Two architectural changes drive the latency gains: frame stacking that doubles the audio frames predicted per decoder step, and a local transformer that models dependencies between codebook tokens to recover quality lost from the speedup, per the team's ICASSP 2026 paper.
On NVIDIA's TTS NIM benchmarks, a single-stream Magpie run reaches 32 milliseconds time-to-first-audio on a B200 GPU and 47 ms on H100, with real-time factor scores above 9x on every tested accelerator. Those numbers are vendor-reported and measured on NVIDIA's own inference stack.
Quality comparisons in the blog are against NVIDIA's prior release, not against competing speech synthesis systems. A demo on NVIDIA's build site and the Nemotron Voice Agent developer example are also available. Independent benchmarks and production deployment data are not yet public.