NVIDIA Releases 364M Magpie TTS Model With 12-Language Support
The open-weight speech model adds Arabic, Korean and Brazilian Portuguese voices, plus a production NIM for teams deploying voice agents on their own infrastructure.
NVIDIA released Magpie TTS Multilingual on August 10, expanding its speech-generation model to 12 languages and making the 364-million-parameter weights available for developers to deploy and customise. The NVIDIA release article also introduces a production-ready NVIDIA NIM, giving teams a packaged inference route alongside direct access to the model weights.
Three new languages widen the deployment map
The release adds Modern Standard Arabic, Korean and Brazilian Portuguese. NVIDIA says each supported language includes male and female speaker voices. It has also expanded code-switching support for Hindi and Japanese, targeting conversations that move between languages rather than staying inside a single locale.
Those additions matter for voice-agent design because language coverage is more than a translation checklist. A useful spoken interface also needs appropriate pronunciation, timing and switching behaviour for the people it serves. The release broadens the model's available voice set, but the primary source does not establish equal quality across every accent, dialect, domain or mixed-language conversation. Teams should test those conditions with representative speakers before treating nominal language support as production readiness.
Open weights give operators more deployment control
NVIDIA positions Magpie TTS Multilingual for voice assistants, customer-service systems, accessibility tools and other applications that need generated speech. Open weights allow a team to run the model inside its own infrastructure and adapt it to a specific domain instead of relying exclusively on a hosted speech endpoint.
That control can be important when an application handles private conversations, specialised vocabulary or a predictable volume of requests. It does not remove the operational work. A self-hosted deployment still needs secure input handling, capacity planning, abuse controls, monitoring and an evaluation set that covers the languages and speaking styles the product will encounter. Operators also need clear consent and disclosure rules when a generated voice could be mistaken for a person.
The production NIM is intended to reduce some packaging and serving work. It provides an NVIDIA-supported inference container rather than requiring every team to assemble the runtime from the weights alone. The source describes that deployment route as production-ready; AI Signal did not independently test its hardware requirements, throughput or integration behaviour in this review.
A smaller model targets interactive voice workloads
At 364 million parameters, Magpie TTS Multilingual is compact relative to many general-purpose language models. NVIDIA presents that size as part of a low-latency voice-agent proposition. Interactive speech systems are especially sensitive to delay because a pause between a user's turn and the generated reply quickly makes a conversation feel unnatural.
The company publishes latency and speech-quality comparisons in its announcement. Those measurements are vendor-provided and were not independently reproduced for this article, so they should be read as evaluation targets rather than settled cross-platform results. Actual performance will depend on the GPU, serving configuration, input length, chosen voice and concurrent load.
For developers, the more durable significance of the release is the combination of multilingual weights and a packaged deployment path. A team can inspect and adapt the model, keep inference within its chosen environment and evaluate the full voice experience against its own users. That makes Magpie TTS Multilingual a practical new option for controlled voice-agent experiments, while leaving the burden of safety, localisation and real-world quality testing with the deployer.
Status
Confirmed model release from NVIDIA. Internal confidence is high because the primary actor documents the model size, language coverage, open-weight availability and deployment route. Performance and quality measurements remain vendor claims pending independent reproduction.
Sources
Update note: Last reviewed 2026-08-11. We will revise this article when independent multilingual evaluations or material deployment changes become available.
Sources
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.