NVIDIA has released Magpie Multilingual TTS, an open-weights text-to-speech model that supports 12 languages and is built for low-latency voice agents. The 364M-parameter model is designed for developers who want full control over deployment, from private servers to air-gapped environments. The release, announced August 10, 2026, pairs the model with NVIDIA NIM microservices for production inference and NeMo for fine-tuning.
Voice AI is becoming multilingual by default. Magpie TTS Multilingual covers English, Spanish, French, German, Italian, Vietnamese, Mandarin, Hindi, Japanese, and adds three new languages: Modern Standard Arabic, Korean, and Brazilian Portuguese. Each language includes male and female speaker voices through a shared multilingual speaker representation. The model also expands code-switching support for Hindi and Japanese, enabled via IPA grapheme-to-phoneme processing and custom pronunciation dictionaries.
Why Open Weights Matter for Voice Agents
Integrated speech models trade control for simplicity. They bundle speech recognition, synthesis, and reasoning into one system, which is easy to deploy but hard to tune. Magpie TTS takes the opposite path. It is built for a cascaded architecture, where purpose-built ASR, TTS, and LLM components work together. This design allows independent tuning and deployment of each piece.
Open weights give developers capabilities from owning deployment. The model runs on your own infrastructure, so you can benchmark and tune it. That control extends to latency, a critical factor in voice interaction. The full latency budget includes audio capture, transcription, LLM reasoning, retrieval, and response generation. TTS is the final step and most noticeable to users.
"Latency you can measure is useful; latency you can control is even better," the NVIDIA authors wrote. Maryam Motamedi, Mikyas Desta, Jason Li, and Jason Roche co-authored the announcement. The team emphasized that open weights let enterprises keep sensitive audio data on-premises, a requirement for many industries.
The model is available under the NVIDIA Open Model License. Developers can download it from Hugging Face, where the model card is listed as nvidia/magpie_tts_multilingual_357m. The Hugging Face page shows 0.2B parameters, was updated 5 days ago, and has racked up 10.6k downloads and 170 likes. A companion Hugging Face Space, "Magpietts Demo," has 28 likes for hands-on testing.
Performance Benchmarks Across GPUs
NVIDIA published latency and throughput figures from its TTS NIM Performance documentation, version v26.07. The numbers are an average of three trials, measured on-prem. Time to first audio, or TTFA, is the key metric for perceived responsiveness.
On a B200 GPU with a single stream, Magpie delivers first audio in 32 ms and runs at 12.1x real time. At 64 concurrent streams on the same hardware, TTFA rises to 239 ms, while throughput jumps to 319.81x real time. That means the system generates audio more than 300 times faster than playback.
The H100 shows similar strength. Single-stream TTFA is 47 ms with 14.7x real-time throughput. At 64 streams, TTFA reaches 275 ms and throughput hits 290.79x real time. The A100, an older but still common data center GPU, manages 79 ms TTFA on a single stream and 12.2x real time. At 64 streams, it delivers 395 ms TTFA and 197x real-time throughput.
For edge or desktop deployment, NVIDIA tested the DGX Spark. Single-stream TTFA is 53 ms with 9.8x real time. At 64 streams, TTFA climbs to 962 ms, and throughput drops to 75.88x real time. Still, that is enough for many interactive applications.
The article claims sub-200ms end-to-end latency is possible with Magpie on B200. That total window, under 200 ms, is widely considered the threshold for natural conversation. The TTS component alone, at 32 ms on B200, leaves ample room for the rest of the pipeline.
Architecture Improvements Behind the Speed
Magpie's speed comes from two architectural changes. The first is frame stacking, where the decoder predicts two audio frames per step instead of one. The second is a local transformer that models dependencies between codebook tokens. Together, they cut inference time while maintaining output quality.
The architecture is described in a paper titled "Frame-Stacked Local Transformers for Efficient Multi-Codebook Speech Generation," presented at ICASSP 2026. The paper details how these changes reduce the number of decoding steps and improve token coherence.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
NVIDIA also recommends a specific inference configuration for best results: cfg_scale = 2.5, temperature = 0.6, top_k = 80, apply_attention_prior = True, and prior_epsilon = 0.1. These settings balance creativity and stability in the generated speech.
Objective metrics help measure progress, but speech quality is ultimately perceptual. Still, the numbers show clear gains. For French, character error rate dropped from 2.70% in the previous release to 1.54% now. Speaker similarity, measured by SSIM, improved from 0.703 to 0.747. Spanish improved from 1.14% CER to 0.60%, with SSIM rising from 0.715 to 0.793. German saw a slight CER increase from 0.66% to 0.80%, but SSIM improved from 0.626 to 0.742.
The new languages have solid baselines. Arabic starts at 1.62% CER, Korean at 2.69%, and Brazilian Portuguese at 2.91%. These figures come from the Magpie TTS Multilingual model card on Hugging Face.
Building a Complete Voice Agent
Magpie TTS is not a standalone tool. NVIDIA positions it as the speech output component of a full voice agent stack. The reference implementation, called the NVIDIA Nemotron Voice Agent Developer Example, shows how to assemble the pieces.
The example includes patterns for barge-in, where the system can interrupt its own speech, multimodal voice agents, multi-agent orchestration, and multilingual interactions. It also demonstrates sub-second latency across the entire pipeline. The stack pairs Magpie TTS with Nemotron Speech, a streaming speech recognition model, and Nemotron language and multimodal models for reasoning and understanding.
The article claims the Nemotron Voice Agent Developer Example can be deployed in hours. That speed comes from using NVIDIA NIM, production-ready inference microservices that are GPU-optimized. NIM handles the serving layer, while NeMo provides the framework for customization and fine-tuning. Developers can adapt Magpie TTS for domain-specific needs, such as medical terminology or legal jargon.
For production deployment, NVIDIA points to the NVIDIA Magpie Multilingual TTS NIM. For fine-tuning and training, the NVIDIA NeMo Speech toolkit is the recommended path. Additional resources include the NVIDIA voice-agent-examples repository on GitHub.
Trying the Model and Next Steps
Developers can hear Magpie TTS in action through two channels. NVIDIA Build, the company's platform for AI demos, features the model. The Hugging Face demo Space also offers interactive testing.
The open weights mean enterprises can deploy in private or air-gapped environments. That is a significant advantage for organizations with strict data residency rules. The model card on Hugging Face provides full documentation, including the recommended inference settings and quality metrics.
The release signals NVIDIA's continued push into speech AI. With 12 languages, sub-100 ms first audio on modern GPUs, and open weights, Magpie TTS Multilingual targets developers who want both performance and control. The combination of NIM for serving and NeMo for tuning gives a complete workflow, from research to production.
The quality improvements on French and Spanish are the clearest evidence of progress. Both languages show substantial drops in character error rate and gains in speaker similarity. German's slight CER increase is offset by better SSIM, suggesting the model produces more consistent voices even if pronunciation accuracy dipped marginally.
For teams building multilingual voice agents, the math is straightforward. A 364M-parameter model that runs at 320x real time on a B200, supports 12 languages, and can be fine-tuned with NeMo covers most enterprise use cases. The open license removes vendor lock-in, and the cascaded architecture allows swapping components as better models emerge.
The article's authors framed the release around a simple idea: voice agents should sound natural, respond quickly, and respect your infrastructure. Magpie TTS Multilingual delivers on all three counts, with measurable latency and quality data to back it up.