NVIDIA Magpie TTS Multilingual Release

NVIDIA has released Magpie TTS Multilingual, a 364M-parameter open-weights text-to-speech (TTS) model supporting 12 languages. This release enables developers to deploy high-quality, multilingual speech synthesis on their own infrastructure to minimize Time to First Audio (TTFA) and maintain full control over data residency and model customization.

Multilingual Support and Capabilities

Magpie TTS Multilingual provides a single open foundation for voice applications across 12 languages, eliminating the need for separate models per region. The model supports:

Supported Languages: English, Spanish, French, German, Italian, Vietnamese, Mandarin, Hindi, Japanese, Modern Standard Arabic, Korean, and Brazilian Portuguese.
New Language Additions: Modern Standard Arabic, Korean, and Brazilian Portuguese are new to this release.
Voice Variety: Each language includes both male and female speaker voices via a shared multilingual speaker representation.
Code-Switching: The model features expanded support for code-switching in Hindi and Japanese, utilizing custom pronunciation dictionaries and IPA grapheme-to-phoneme processing to improve the pronunciation of technical terms and names.

Low-Latency Performance and Deployment

To ensure natural conversation, Magpie TTS focuses on reducing Time to First Audio (TTFA), the delay between the start of generation and the first audio reaching the user. By deploying via NVIDIA NIM (optimized inference containers) on-premises, developers avoid managed-service round-trip latency.

Performance Benchmarks

On a single stream, Magpie delivers first audio in 32–79ms across various NVIDIA GPUs. Under a concurrent load of 64 streams, the B200 GPU maintains a TTFA of 239ms while achieving a throughput of 319.81 × real time.

GPU	1-stream TTFA	1-stream RTFX	64-stream TTFA	64-stream RTFX
B200	32 ms	12.1×	239 ms	319.81×
H100	47ms	14.7×	275 ms	290.79×
DGX Spark	53 ms	9.8×	962 ms	275.88×
A100	79 ms	12.2×	395 ms	197×

Architectural Optimizations for Real-Time Speech

Magpie utilizes two key architectural improvements to reduce inference time without sacrificing audio quality, as detailed in the paper Frame-Stacked Local Transformers for Efficient Multi-Codebook Speech Generation (ICASSP 2026):

Frame Stacking: The decoder predicts two audio frames per decoding step instead of one, effectively halving the number of decoder iterations.
Local Transformer: This component models dependencies between simultaneously generated codebook tokens to refine the audio and recover quality lost during frame stacking.

Synthesis Quality Improvements

This release improves synthesis quality across existing languages, specifically reducing Character Error Rates (CER) and increasing Speaker Similarity (SSIM).

Language	CER (prev)	CER (this release)	SSIM (prev)	SSIM (this release)
French	2.70%	1.54%	0.703	0.747
Spanish	1.14%	0.60%	0.793	0.793
German	0.66%	0.80%	0.742	0.742

New languages (Arabic: 1.62% CER, Korean: 2.69% CER, and Brazilian Portuguese: 2.91% CER) establish the baseline for future iterations.

Enterprise Control and Integration

Because Magpie is an open-weights model, enterprises can deploy it in private or air-gapped environments, fine-tune it using NVIDIA NeMo for brand-specific voices or domain vocabulary, and scale the serving stack independently.

Magpie is integrated into the NVIDIA Nemotron Voice Agent developer example, which provides a reference architecture for building complete voice agents. This system combines:

Nemotron Speech: For streaming speech recognition (ASR).
Magpie TTS: For multilingual speech synthesis.
Nemotron Language/Multimodal Models: For reasoning and tool calling.
NVIDIA NIM: For GPU-optimized inference.
NeMo: For customization and fine-tuning.

This reference implementation supports production patterns such as real-time interruptible (barge-in) conversations and multimodal voice agents with vision understanding.