NVIDIA releases Magpie multilingual text-to-speech model with 12-language open weights and 32 ms latency

NVIDIA has released Magpie Multilingual TTS, a 364-million-parameter open-weight text-to-speech model that supports 12 languages and is aimed at low-latency voice applications. The new release adds Modern Standard Arabic, Korean and Brazilian Portuguese to the previous line-up, and improves quality across the earlier-supported English, Spanish, French, German, Italian, Vietnamese, Mandarin, Hindi and Japanese with updated training data.

Each language ships with male and female speaker voices produced through a shared multilingual speaker representation. This release also improves code-switching for Hindi and Japanese, using IPA grapheme-to-phoneme processing and custom pronunciation dictionaries to handle names, technical terminology and mixed-language content, according to NVIDIA.

In a voice pipeline, text-to-speech is the last step before the user hears anything, so the delay between generation beginning and the first audio arriving — called Time to First Audio, or TTFA — is one of the most-noticed latency metrics. NVIDIA reports Magpie's TTFA on a single stream at 32 milliseconds on a B200 GPU, 47 ms on an H100, 53 ms on DGX Spark and 79 ms on an A100. At 64 concurrent streams, the B200 measurement rises to 239 ms while delivering throughput at 320 times real time.

The low latency comes from two architectural changes, NVIDIA said. Frame stacking has the decoder predict two audio frames per decoding step rather than one, which halves the number of decoder iterations. Because that would introduce dependencies between simultaneously generated codebook tokens and reduce audio quality, a 'local transformer' then models those dependencies and refines the output to recover the quality frame stacking would otherwise sacrifice.

The model is available both as open weights on Hugging Face for research and fine-tuning and as an NVIDIA NIM optimised serving container for production. NVIDIA said both run on hardware the developer controls, so the latency measured is server-side latency that can be tuned rather than a managed-service round-trip.