NVIDIA expands Magpie TTS to 12 languages with open weights
NVIDIA released an updated version of Magpie TTS Multilingual, its open-weights text-to-speech model, adding Modern Standard Arabic, Korean, and Brazilian Portuguese. That brings the model's language coverage to 12: English, Spanish, French, German, Italian, Vietnamese, Mandarin, Hindi, Japanese, plus the three new additions. The model itself has 364M parameters, and every language gets both a male and a female speaker voice through a shared multilingual speaker representation. The release also widens code-switching support for Hindi and Japanese, using IPA grapheme-to-phoneme processing and custom pronunciation dictionaries so names and technical terms embedded in mixed-language text are pronounced correctly.
The headline metric is Time to First Audio (TTFA), the delay between speech generation starting and the first audio reaching the user, which NVIDIA calls one of the most important latency figures in a voice pipeline. Served through NVIDIA's NIM inference container running on-prem, Magpie reaches 32ms TTFA on a single stream on a B200 GPU, and 32 to 79ms across NVIDIA GPUs more broadly. Under load, at 64 concurrent streams on B200, TTFA rises to 239ms while throughput holds at 320 times real time, meaning the model still generates audio well over 300 times faster than it plays back. NVIDIA credits two architectural changes for the speed: frame stacking, where the decoder predicts two audio frames per decoding step instead of one, cutting the number of decoder iterations in half, paired with a local transformer that models the dependencies this introduces between codebook tokens and recovers the audio quality that frame stacking alone would otherwise cost. The design is described in a paper, "Frame-Stacked Local Transformers for Efficient Multi-Codebook Speech Generation," accepted at ICASSP 2026.
Quality moved alongside speed. NVIDIA reports lower character error rates (CER) and higher speaker similarity (SSIM) across several existing languages compared with the previous release, with French and Spanish showing the clearest gains, though the source gives exact figures only for the three newly added languages: 1.62% CER for Arabic, 2.69% for Korean, and 2.91% for Brazilian Portuguese, which NVIDIA describes as baseline numbers it expects to improve in future releases.
The weights are open under the NVIDIA Open Model License, distributed as a Hugging Face checkpoint for research and fine-tuning, with the NIM container as the production serving stack built from the same model. Developers can fine-tune it with NVIDIA NeMo for their own brand voice, domain vocabulary, or speaker data. NVIDIA positions Magpie TTS as one piece of its Nemotron Voice Agent reference architecture, meant to be combined with Nemotron Speech for streaming recognition and Nemotron language models for reasoning and tool calling, aimed at customer support, healthcare, enterprise, and translation use cases.
Key facts
- Magpie TTS Multilingual is a 364M-parameter open-weights model; the new release adds Modern Standard Arabic, Korean, and Brazilian Portuguese, bringing total language coverage to 12.
- On a single stream, Time to First Audio is 32ms on NVIDIA's B200 GPU and 32 to 79ms across NVIDIA GPUs generally, served via the NIM inference container on-prem.
- At 64 concurrent streams, B200 TTFA rises to 239ms while throughput stays at 320 times real time.
- The new-language quality baselines are 1.62% CER for Arabic, 2.69% for Korean, and 2.91% for Brazilian Portuguese.
- Weights are released under the NVIDIA Open Model License, deployable through NVIDIA NIM and fine-tunable with NeMo, as part of the Nemotron Voice Agent reference architecture.
Why it matters
Text-to-speech is the last step in a voice pipeline and the one users feel most directly: after audio capture, transcription, an LLM call and context retrieval have already spent milliseconds, a slow TTS step makes the whole interaction feel slow. NVIDIA is pitching Magpie as the answer for teams that build their own cascaded pipeline (separate ASR, LLM, and TTS components) rather than a single bundled speech API, arguing that owning each component lets you tune latency and swap models independently. The core claim is that TTFA as low as 32ms leaves most of a sub-200ms end-to-end latency budget free for the ASR and LLM steps that come before it.
Who it affects
Developers building production voice agents: customer support bots, healthcare assistants, enterprise copilots, translation systems, and other conversational AI applications that need to serve users in more than one language without maintaining a separate TTS model per region.
How to use it
The Hugging Face checkpoint is open under the NVIDIA Open Model License and is the path for research and fine-tuning; NVIDIA NIM is the optimized production container built from the same model, and NVIDIA NeMo is offered for fine-tuning pronunciation, brand voice, or domain vocabulary. NVIDIA also publishes recommended inference settings for the model (cfg_scale 2.5, temperature 0.6, top_k 80, attention prior enabled). Magpie is positioned as one component of the broader Nemotron Voice Agent reference architecture, alongside Nemotron Speech for recognition and Nemotron language models for reasoning and tool calling, which NVIDIA says developers can clone and deploy as a starting point.
How solid is it
The latency and quality numbers come from NVIDIA's own benchmarks (the TTFA figures cite NVIDIA's TTS NIM performance documentation, version 26.07, averaged over three trials on-prem; the CER and speaker-similarity figures cite the Magpie model card) rather than independent third-party testing. The underlying architecture change, frame stacking paired with a local transformer, is documented in a paper accepted at ICASSP 2026, which gives the latency claims an external, peer-reviewed basis even though the specific benchmark numbers in this post are NVIDIA's.
Risks and caveats
The source names no competing TTS model or vendor for comparison, so the latency and quality figures cannot be checked against a rival system from the same text. No pricing or licensing cost is given for running the NIM container in production, only that the model weights themselves are open. The latency and quality comparison tables referenced in the post are rendered as images in the source, so only the specific figures repeated in the surrounding prose could be confirmed as text; the full per-GPU and per-language breakdowns were not verifiable here.