Magpie TTS

Magpie TTS is NVIDIA's open-weights text-to-speech model aimed squarely at builders of multilingual voice agents who want full deployment control instead of per-character API billing. The 357M-parameter model (causal transformer encoder-decoder with a NanoCodec decoder, 22.05 kHz mono output) launched with 9 languages — English, Spanish, German, French, Italian, Vietnamese, Mandarin, Hindi, Japanese — and the July 2026 v2607 update added Modern Standard Arabic, Korean, and Brazilian Portuguese, plus better Hindi/Japanese code-switching via IPA grapheme-to-phoneme processing. Each language ships male and female voices drawn from five baked speakers (Aria, Jason, John, Leo, Sofia) through a shared multilingual speaker representation; there is deliberately no zero-shot cloning, though the checkpoints are fine-tunable for custom voices. NVIDIA packages the full inference stack — model, NanoCodec decoder, tokenizer, and G2P dictionaries — so the C++ build runs without Python, PyTorch, NeMo, or a CUDA toolkit. On the provider's own model card it reports 0.37% English CER — an NVIDIA evaluation, not a listening test. Weights are under the NVIDIA Open Model License, which permits commercial use.

• Provider-reported 0.37% English CER on the v2607 model card (NVIDIA evaluation, not listening tests)
• No independent benchmarks found