NVIDIA Releases Magpie TTS Multilingual Update: 364M-Parameter Open-Weights Model Now Supports 12 Languages, Sub-50ms La

What happened

NVIDIA released an updated version of Magpie TTS Multilingual, an open-weights text-to-speech model, expanding language support from nine to twelve languages and improving synthesis quality across several existing ones. The 364-million-parameter model now covers English, Spanish, French, German, Italian, Vietnamese, Mandarin, Hindi, Japanese, Modern Standard Arabic, Korean, and Brazilian Portuguese — with the last three being new additions.

Each language ships with male and female voices built on a shared multilingual speaker representation. The release also expands code-switching support for Hindi and Japanese via IPA grapheme-to-phoneme processing and custom pronunciation dictionaries, aimed at improving accuracy on names, technical terms, and mixed-language text.

Performance numbers

According to NVIDIA's TTS NIM performance documentation (v26.07), single-stream time-to-first-audio (TTFA) ranges from 32ms on B200 GPUs to 79ms on A100 GPUs. At 64 concurrent streams, B200 delivers 239ms TTFA while sustaining throughput 319.81x real time. H100 reaches 275ms TTFA at 64 streams with 290.79x real-time throughput; DGX Spark reaches 962ms TTFA at 75.88x.

On quality, NVIDIA reports character error rate (CER) improvements on French (2.70% to 1.54%) and Spanish (1.14% to 0.60%), with corresponding speaker similarity (SSIM) gains from 0.703 to 0.747 and 0.715 to 0.793 respectively. German CER rose slightly from 0.66% to 0.80% while SSIM improved from 0.626 to 0.742. The three newly added languages establish baseline CER scores: Arabic at 1.62%, Korean at 2.69%, and Brazilian Portuguese at 2.91%. These figures come from NVIDIA's own model card and have not been independently verified.

Architecture

NVIDIA attributes the low latency to two techniques described in a paper accepted at ICASSP 2026, "Frame-Stacked Local Transformers for Efficient Multi-Codebook Speech Generation." Frame stacking has the decoder predict two audio frames per decoding step instead of one, halving decoder iterations. A local transformer then models dependencies between the simultaneously generated codebook tokens to recover audio quality that frame stacking alone would degrade.

Deployment model

Magpie is distributed as open weights on Hugging Face for research and fine-tuning, and as an NVIDIA NIM container for production serving — both running on infrastructure controlled by the deployer rather than through a managed API. NVIDIA positions this against integrated speech-to-speech APIs, arguing that self-hosting eliminates managed-service round-trip latency and allows fine-tuning with NeMo for custom vocabulary, brand voices, or air-gapped enterprise environments. Pricing is not applicable in the traditional per-token sense since the model is open weights; compute costs depend on the deploying organization's own GPU infrastructure.

What this means

This is an incremental but meaningful update to an existing open TTS model rather than a new architecture. The sub-50ms single-stream latency figures on modern NVIDIA GPUs are notable for real-time voice agent pipelines, where every millisecond in the TTS stage eats into the sub-200ms budget needed for conversation to feel natural. The tradeoff NVIDIA is selling — control and customization versus the simplicity of an all-in-one speech-to-speech API — will matter most to enterprises with data residency requirements or teams that need to fine-tune pronunciation for brand and domain vocabulary. Whether the reported latency numbers hold outside NVIDIA's own benchmarking environment, and how the model performs against competing open TTS systems like those from ElevenLabs or Coqui, remains for independent evaluation.