NVIDIA updates open-weights Magpie multilingual TTS for low-latency voice agents

NVIDIA has released an updated version of its Magpie multilingual text-to-speech model, adding Modern Standard Arabic, Korean, and Brazilian Portuguese to bring single-model language coverage to 12 languages. The company says the release also improves synthesis quality across many existing languages through updated training data and model improvements.

Magpie TTS Multilingual is a 364M-parameter open-weights model covering English, Spanish, French, German, Italian, Vietnamese, Mandarin, Hindi, and Japanese alongside the three new languages. Each language ships with male and female voices built on a shared multilingual speaker representation.

The update also expands code-switching support for Hindi and Japanese, using IPA grapheme-to-phoneme processing and custom pronunciation dictionaries to handle names, technical terms, and mixed-language content more accurately. NVIDIA reports character error rates dropping from 2.70% to 1.54% for French and from 1.14% to 0.60% for Spanish, with higher speaker similarity in both.

Latency is the centerpiece of the release. On a B200, time-to-first-audio (TTFA) reaches 32 ms on a single stream, while 64 concurrent streams deliver 239 ms TTFA at more than 300x real-time throughput; measured figures for H100, A100, and DGX Spark are published as well.

The low latency comes from two architectural changes: frame stacking lets the decoder predict two audio frames per step, halving the number of decoder iterations, while a local transformer models dependencies between codebook tokens to recover quality. The architecture is described in a paper accepted at ICASSP 2026.

Magpie is built for cascaded voice pipelines, with purpose-built ASR, TTS, and LLM layers running together rather than all-in-one speech models. Open weights plus NVIDIA NIM containers let enterprises deploy on their own or air-gapped infrastructure, keeping control over latency budgets, data residency, and pronunciation customization.

NVIDIA also positions Magpie inside its Nemotron Voice Agent developer example, combining Nemotron Speech streaming recognition, Magpie TTS, and Nemotron language and multimodal models into a reference implementation supporting interruptible conversations, multimodal understanding, and multilingual interaction.

For developers, the immediate next steps are downloading the open weights for fine-tuning or deploying the NIM containers for production. In customer support, healthcare assistants, and translation systems, Magpie gives self-hosted multilingual voice agents a concrete open-source starting point.