NVIDIA TTS NIM Support Matrix#

This page describes the software and hardware that the NVIDIA TTS NIM supports.

Hardware#

The NVIDIA TTS NIM requires an NVIDIA GPU with a Compute Capability of 8.0 or higher and at least 16 GB of VRAM. Ensure that the models you deploy do not exceed the available GPU memory.

GPU	Precision
A30, A100	FP16
H100	FP16
A2, A10, A16, A40	FP16
L4, L40, GeForce RTX 40xx	FP16
GeForce RTX 50xx	FP16
Blackwell RTX 60xx	FP16
DGX Spark	FP16

Note

Magpie TTS Multilingual, Magpie TTS Zeroshot, and Chatterbox TTS Multilingual support the DGX Spark platform (Multilingual support extended in the Riva TTS NIM Release 1.10.0). On DGX Spark, Chatterbox TTS Multilingual provides only the batch_size=8 profile; refer to Chatterbox TTS Multilingual model profiles.

Software#

Linux operating systems (Ubuntu 22.04 or later recommended)
NVIDIA Driver >= 535
NVIDIA Docker >= 23.0.1

Supported Models#

The NVIDIA TTS NIM microservice supports the following Nemotron TTS models.

The microservice requires an NVIDIA GPU with Compute Capability >= 8.0. It automatically downloads a pre-built model when available for the target GPU or generates an optimized model on-the-fly using the RMIR model.

Model	Languages Supported	Inference Mode	Publisher	WSL Support	Model Format
Magpie TTS Multilingual	English (en-US), Spanish (es-US), French (fr-FR), German (de-DE), Mandarin (zh-CN), Vietnamese (vi-VN), Italian (it-IT), Hindi (hi-IN), Japanese (ja-JP), Korean (ko-KR), Arabic (ar-XA), Portuguese (pt-BR)	Streaming & Offline	NVIDIA	No	Prebuilt, RMIR
Magpie TTS Zeroshot	English (en-US), Spanish (es-US), French (fr-FR), German (de-DE), Mandarin (zh-CN), Vietnamese (vi-VN), Italian (it-IT), Hindi (hi-IN), Japanese (ja-JP), Arabic (ar-XA), Portuguese (pt-BR), Korean (ko-KR)	Streaming & Offline	NVIDIA	No	Prebuilt, RMIR
Chatterbox TTS Multilingual	Arabic (ar-SA), Danish (da-DK), German (de-DE), Greek (el-GR), English (en-US), Spanish (es-ES), Finnish (fi-FI), French (fr-FR), Hebrew (he-IL), Hindi (hi-IN), Italian (it-IT), Japanese (ja-JP), Korean (ko-KR), Malay (ms-MY), Dutch (nl-NL), Norwegian (nb-NO), Polish (pl-PL), Portuguese (pt-BR), Russian (ru-RU), Swedish (sv-SE), Swahili (sw-KE), Turkish (tr-TR), Mandarin (zh-CN)	Streaming & Offline	Resemble AI (community)	No	Prebuilt, RMIR

Note

All models use FP16 precision.

Note

No TTS model currently supports Windows Subsystem for Linux (WSL2). For WSL2 deployments, refer to the ASR support matrix for the models that do.

Magpie TTS Multilingual#

This model supports text to speech in English (en-US), Spanish (es-US), French (fr-FR), German (de-DE), Mandarin (zh-CN), Vietnamese (vi-VN), Italian (it-IT), Hindi (hi-IN), Japanese (ja-JP), Korean (ko-KR), Arabic (ar-XA), and Portuguese (pt-BR).

To use this model, set CONTAINER_ID to magpie-tts-multilingual. Then, set NIM_TAGS_SELECTOR to one of the values from the following table as required. For more information, refer to Launching the TTS NIM.

Model Profiles#

Model Profile (Selected using `NIM_TAGS_SELECTOR`)	Inference Mode	Host Memory (GiB)	GPU Memory (GiB)
`batch_size=8`	offline and streaming	5.182	12.58
`batch_size=32`	offline and streaming	5.208	41.46
`batch_size=64`	offline and streaming	5.258	74.74

Note

If NIM_TAGS_SELECTOR is not specified, the default model profile is batch_size=8.

Magpie TTS Zeroshot#

This model supports text to speech in English (en-US), Spanish (es-US), French (fr-FR), German (de-DE), Mandarin (zh-CN), Vietnamese (vi-VN), Italian (it-IT), Hindi (hi-IN), Japanese (ja-JP), Arabic (ar-XA), Portuguese (pt-BR), and Korean (ko-KR) languages. It synthesizes speech from input text using either a built-in voice or an audio prompt of three to ten seconds for zero-shot voice cloning.

To use this model, set CONTAINER_ID to magpie-tts-zeroshot. Then, set NIM_TAGS_SELECTOR to one of the values from the following table, based on your requirements. For more information, refer to Launching the TTS NIM.

Note

Access to Magpie TTS Zeroshot model is restricted. Apply for access

Model Profiles#

Model Profile (Selected using `NIM_TAGS_SELECTOR`)	Inference Mode	CPU Memory (GB)	GPU Memory (GB)
`batch_size=8`	offline and streaming	4.00	13.06
`batch_size=32`	offline and streaming	7.08	41.30

Note

If NIM_TAGS_SELECTOR is not specified, the default model profile is batch_size=8.

Available Voices#

Magpie TTS Zeroshot provides built-in default voices and also supports zero-shot voice cloning from an audio prompt. The same voices are available across all supported languages.

Built-in voices

Magpie-ZeroShot-Multilingual.Female  (Default female speaker)
Magpie-ZeroShot-Multilingual.Male    (Default male speaker)

Audio prompt (voice cloning)

Magpie-ZeroShot-Multilingual

Chatterbox TTS Multilingual#

Chatterbox TTS Multilingual is an end-to-end multilingual text-to-speech model that generates expressive and natural speech. The model uses a T3 architecture (text-to-speech token generation) paired with an S3Gen diffusion-based decoder using flow matching for high-fidelity audio generation. At 500M parameters, it supports Arabic, Danish, German, Greek, English, Spanish, Finnish, French, Hebrew, Hindi, Italian, Japanese, Korean, Malay, Dutch, Norwegian, Polish, Portuguese, Russian, Swedish, Swahili, Turkish, and Chinese.

Features include emotion exaggeration control for adjusting expressiveness intensity.

Note

Chatterbox TTS Multilingual is a community model hosted on NVIDIA NIM for accelerated inference. It is not owned or developed by NVIDIA. The model is licensed under the MIT License; use of this API is governed by the NVIDIA API Trial Service Terms of Use.

To use this model, set CONTAINER_ID to chatterbox-tts-multilingual. Then, set NIM_TAGS_SELECTOR to the value from the following table. For more information, refer to Launching the TTS NIM.

Model Profiles#

Model Profile (Selected using `NIM_TAGS_SELECTOR`)	Inference Mode	Supported GPUs	Host Memory (GiB)	GPU Memory (GiB)
`batch_size=8`	offline and streaming	A100, H100, L40S, DGX Spark	4.86	44.61
`batch_size=32`	offline and streaming	A100, H100, L40S	5.40	46.84
`batch_size=64`	offline and streaming	A100, H100, L40S	5.54	49.72

Note

If NIM_TAGS_SELECTOR is not specified, the default model profile is batch_size=8.

Note

On DGX Spark, only the batch_size=8 profile is available. Setting NIM_TAGS_SELECTOR to batch_size=32 or batch_size=64 on DGX Spark causes deployment to fail because no matching profile exists. To list the profiles that are compatible with your system, run the container with the nim_list_model_profiles entrypoint.

Supported Locales#

Locale	Language
`ar-SA`	Arabic
`da-DK`	Danish
`de-DE`	German
`el-GR`	Greek
`en-US`	English
`es-ES`	Spanish
`fi-FI`	Finnish
`fr-FR`	French
`he-IL`	Hebrew
`hi-IN`	Hindi
`it-IT`	Italian
`ja-JP`	Japanese
`ko-KR`	Korean
`ms-MY`	Malay
`nl-NL`	Dutch
`nb-NO`	Norwegian
`pl-PL`	Polish
`pt-BR`	Portuguese
`ru-RU`	Russian
`sv-SE`	Swedish
`sw-KE`	Swahili
`tr-TR`	Turkish
`zh-CN`	Mandarin

Known Limitations#

English quality is strongest. Non-English languages have reported issues with pronunciation, mixed-language output, hallucinated words, and speed-up artifacts. See upstream issues #268, #254, #271, #327.