NVIDIA Magpie TTS

Spec sheet

What it does

Magpie TTS Multilingual is a neural text-to-speech model for generating speech in twelve languages.Read 15 Oct 2026 · build.nvidia.com

Platforms

web,linux,api,self_hostedRead 20 Sep 2026 · docs.api.nvidia.com

Free plan

YesRead 20 Sep 2026 · docs.api.nvidia.com

Headquarters

Santa Clara, California, USARead 28 Sep 2026 · build.nvidia.com

Founded

1993Read 28 Sep 2026 · build.nvidia.com

Alignment features

Attention priors, classifier-free guidance and GRPO are used to improve alignment and generation robustness.Read 15 Oct 2026 · build.nvidia.com

API access

YesRead 20 Sep 2026 · docs.api.nvidia.com

API protocols

NVIDIA states that Nemotron Speech supports both gRPC and HTTP APIs.Read 15 Oct 2026 · build.nvidia.com

Architecture

It uses transformer encoder and decoder networks with local-transformer refinement and multi-codebook audio-token prediction.Read 15 Oct 2026 · build.nvidia.com

Cloud integrations

NVIDIA documents deployment through Azure, AWS, GCP, OCI and other NVIDIA cloud-service-provider partners.Read 15 Oct 2026 · docs.api.nvidia.com

Commercial use

NVIDIA marks this release as ready for commercial use under its stated terms.Read 15 Oct 2026 · build.nvidia.com

Company founded

LinkedIn lists NVIDIA as founded in 1993.Read 15 Oct 2026 · linkedin.com

Company headquarters

LinkedIn lists NVIDIA headquarters in Santa Clara, California.Read 15 Oct 2026 · linkedin.com

Company legal name

LinkedIn identifies the maker as NVIDIA, a public computer-hardware company.Read 15 Oct 2026 · linkedin.com

Company profile

Crunchbase describes NVIDIA as designing graphics processors and accelerated-computing hardware for AI and high-performance computing.Read 15 Oct 2026 · crunchbase.com

Company size

LinkedIn places NVIDIA in the 10,001-plus employee size category.Read 15 Oct 2026 · linkedin.com

Developer access complaint

Forum users reported that applications for Magpie TTS Zeroshot access were rejected or asked for an NVIDIA contact and business email, with one user later reporting access was granted.Read 15 Oct 2026 · forums.developer.nvidia.com

Documentation depth

NeMo documentation covers training manifests, Hydra configuration, DPO and GRPO preference optimization, inference parameters and long-form generation.Read 15 Oct 2026 · docs.nvidia.com

Export formats

WAVRead 20 Sep 2026 · docs.api.nvidia.com

HTTP API

The hosted model exposes HTTP endpoints for listing voices and synthesizing speech, authenticated with an NVIDIA API key.Read 15 Oct 2026 · build.nvidia.com

Independent implementation

The community magpie-tts.cpp project provides a C++17/ggml implementation with a CLI, C++ library and flat C API, and reports CPU inference without Python, PyTorch, NeMo or CUDA.Read 15 Oct 2026 · github.com

Languages

12Read 20 Sep 2026 · docs.api.nvidia.com

Languages

Supported languages are English, Spanish, German, French, Italian, Vietnamese, Mandarin Chinese, Hindi, Japanese, Modern Standard Arabic, Brazilian Portuguese and Korean.Read 15 Oct 2026 · build.nvidia.com

Licence

Use is governed by the NVIDIA Open Model License, while the NIM container is governed by NVIDIA software and product-specific AI terms.Read 15 Oct 2026 · build.nvidia.com

Limit: crackling reports

An NVIDIA developer-forum report describes intermittent popping or crackling near the start or end of Magpie segments on Jetson AGX Thor; NVIDIA said newer Riva releases fixed the reproduced case.Read 15 Oct 2026 · forums.developer.nvidia.com

Limit: duplicated audio bug

An NVIDIA NeMo GitHub issue reports intermittent duplicated or repeated audio at the end of generations.Read 15 Oct 2026 · github.com

Limit: hardware dependence

The independent review says the high-performance path relies on NVIDIA TensorRT, limiting AMD, Intel and Apple Silicon options.Read 15 Oct 2026 · singularitymoments.com

Limit: offline duration

NVIDIA TTS troubleshooting states that offline output is capped at 20 seconds per request and recommends streaming for longer text.Read 15 Oct 2026 · docs.nvidia.com

Limit: pipeline latency

The independent review notes that Magpie remains one stage in a cascaded ASR, language-model and TTS pipeline, so other stages can dominate latency.Read 15 Oct 2026 · singularitymoments.com

Limit: streaming format

The streaming HTTP endpoint returns raw LPCM without a WAV header, so clients must convert it before normal playback.Read 15 Oct 2026 · docs.nvidia.com

Limit: voice names

Voice identifiers are model-specific and invalid or misspelled names cause synthesis errors.Read 15 Oct 2026 · docs.nvidia.com

Long-form synthesis

The model supports batch inference and long-form inference using sliding windows for very long text.Read 15 Oct 2026 · build.nvidia.com

Model size

The model has approximately 364 million parameters.Read 15 Oct 2026 · build.nvidia.com

Output format

Generated audio is mono 16-bit PCM WAV sampled at 22.05 kHz.Read 15 Oct 2026 · build.nvidia.com

Praise: local control

An independent editorial review praises open weights, self-hosting and keeping audio data under the developer's deployment control.Read 15 Oct 2026 · singularitymoments.com

Praise: multilingual support

The same review highlights native coverage of twelve languages as a strength for voice applications.Read 15 Oct 2026 · singularitymoments.com

Release date

The model-card release date for Build.NVIDIA.com and NGC is June 22, 2026.Read 15 Oct 2026 · build.nvidia.com

Runtime integration

NVIDIA lists Riva 2.15.0 as the runtime integration.Read 15 Oct 2026 · build.nvidia.com

Security and data terms

API trial terms restrict production use without a subscription and prohibit submitting confidential, health, payment-card or other sensitive data unless expressly allowed.Read 15 Oct 2026 · assets.ngc.nvidia.com

Self-hosting

NIM can be downloaded as NGC or Docker containers and deployed on cloud, data-center or workstation infrastructure.Read 15 Oct 2026 · docs.api.nvidia.com

Supported hardware

Listed GPU compatibility includes NVIDIA A100, A30, A10, H100, L4 and L40 GPUs.Read 15 Oct 2026 · build.nvidia.com

Supported operating systems

The model card lists Linux and Linux for Tegra as supported operating systems.Read 15 Oct 2026 · build.nvidia.com

Voice cloning

YesRead 20 Sep 2026 · docs.api.nvidia.com

Voice cloning input

Optional zero-shot voice cloning accepts mono 16-bit PCM WAV audio at 22.05 kHz, with a recommended reference length of 3 to 10 seconds.Read 15 Oct 2026 · build.nvidia.com

Voice coverage

The model provides at least one male and one female speaker for every supported language.Read 15 Oct 2026 · build.nvidia.com

Voice mixing

The voice documentation says Magpie can combine a voice from one locale with text in another language to create accented speech.Read 15 Oct 2026 · docs.nvidia.com

Plans and prices

Every NVIDIA Magpie TTS plan and price →

Where Specifiction places it

• Text-to-speech tools#1 of 2339.4
• Text-to-Speech Software#14 of 2239.0