NVIDIA's Free, Open Voice AI Now Speaks 12 Languages and Responds in Under a Tenth of a Second

Key points

• NVIDIA Magpie Multilingual TTS, a free open-weights text-to-speech model, now supports 12 languages after adding Arabic, Korean and Brazilian Portuguese in its latest release.
• On NVIDIA's fastest current chip, the B200, the model produces its first spoken audio in just 32 milliseconds, well inside the threshold where humans notice a delay.
• Developers can download and run the model entirely on their own computers or servers, with no data sent to NVIDIA's cloud.
• French character error rate, a measure of how accurately the model pronounces words, dropped from 2.70% to 1.54% compared with the previous version.
• The model's architecture and benchmarks are documented in a paper accepted to ICASSP 2026, a leading audio research conference.

When you ask a voice assistant a question, a small clock starts ticking the moment you stop speaking. Your words get transcribed, an AI works out a reply, and then a text-to-speech engine, software that reads written text aloud in a synthetic voice, turns that reply into sound. The last step is the one your ears judge most harshly. If the voice takes too long to start talking, the whole conversation feels sluggish.

NVIDIA wants to fix that with Magpie Multilingual TTS, a text-to-speech model it is releasing as open weights, meaning the full trained model files are free to download and run anywhere. The latest update, first detailed on Hugging Face, expands the model to 12 languages and improves how natural it sounds across several existing ones.

What did NVIDIA actually change?

Three new languages arrive: Modern Standard Arabic, Korean and Brazilian Portuguese, bringing the total to 12. The model is 364 million parameters, a rough measure of its complexity, and handles all 12 languages in a single download rather than requiring a separate model per language.

Quality improved measurably for existing languages too. French pronunciation accuracy tightened sharply, and Spanish followed closely behind.

Source: Magpie TTS Multilingual model card. Lower error rate is better; higher voice match score is better.

How fast does it actually speak?

Fast enough that most people will not notice a pause. On an NVIDIA B200 chip, the kind found in high-end data centres, the model starts producing audio 32 milliseconds after being given text. That is roughly a thirtieth of a second. Even on an NVIDIA A100, an older and more widely deployed chip, that first-audio delay is 79 milliseconds.

For context, humans begin to perceive a conversation as awkward when responses take longer than roughly 200 milliseconds. Magpie's speech step leaves most of that budget free for the other parts of the pipeline.

Two engineering choices drive the speed. First, the model predicts two chunks of audio at once instead of one during each processing step, halving the number of steps needed. Second, a small additional network called a local transformer cleans up any quality loss that comes from doing two chunks at a time, keeping the voice sounding natural.

Why does running it yourself matter?

Most commercial text-to-speech services are cloud-based: your text travels to a remote server, gets converted to audio, and travels back. Every one of those trips costs time and means a third party handles your data.

Because Magpie's weights are open, a hospital, bank or any organisation with strict privacy rules can run the model on its own servers. Audio never leaves the building. Developers can also fine-tune the model on their own voice samples or specialist vocabulary, something a locked commercial API will not allow.

For everyday users, the practical effect is voice assistants in apps built on Magpie that feel more responsive and can handle more languages without the developer paying per-character fees to a cloud provider.