Checkout our
for the latest open weight checkpoints and demos!
Updates
NeMo Speech 3.0 is now available as
and in the
26.07.00 NeMo Speech NGC container
. The final NeMo release before the repository split was
, available in the
.
2026-07:
has been released with support for 3 new languages (Ar, Ko, Pt) + 9 existing languages (En, Es, De, Fr, Vi, It, Zh, Hi, Ja). Try out
!
2026-06:
Nemotron-3.5-ASR-Streaming-0.6B
has been released with 40 languages supported, controllable latency 80ms-1s, and 240-2400 1xH100 concurrent streams. Built on cache-aware Fastconformer architecture.
2026-04:
has been released with high-quality offline and streaming (with a minimum latency of 160ms) inference in one model for English language with punctuation and capitalization support.
2026-03:
is now released in Early Access. Built on the Nemotron Nano v2 LLM backbone with Nemotron speech and TTS decoder, VoiceChat delivers full-duplex, natural, interruptible conversations with low latency. Try out
and apply for
.
2026-03:
Nemotron-Speech-Streaming v2603
has been updated. It has been trained on a larger and more diverse corpus, resulting in lower WER across all latency modes. Try out
and check out
.
2026-03:
has been released with support for 9 languages (En, Es, De, Fr, Vi, It, Zh, Hi, Ja).
2026-01: Nemotron-Speech-Streaming was released: One checkpoint that enables users to pick their optimal point on the latency-accuracy Pareto curve!
2026-01:
was released.
2026: This repo has pivoted to focus on audio, speech, and multimodal LLMs. For the final pre-split NeMo release with support for additional modalities, see
.
2025-08:
and
have been released with speech recognition and translation support for 25 European languages.
2025-06:
has been released with record-setting 5.63% WER on English Open ASR Leaderboard.
Introduction
NVIDIA NeMo Speech is built for researchers and PyTorch developers working on Speech models including Automatic Speech Recognition (ASR), Text to Speech (TTS), and Speech LLMs. It is designed to help you efficiently create, customize, and deploy new AI models by leveraging existing code and pre-trained model checkpoints.
For technical documentation, please see the
NeMo Speech Developer Documentation
.
Requirements
NeMo Speech works with the Python, PyTorch, and CUDA versions of your choosing:
Python 3.12 or above
PyTorch 2.7 or above (CPU, CUDA, etc. — your choice)
NVIDIA GPU + CUDA (required for training; recommended for inference)
If you already have a Python/PyTorch/CUDA stack that satisfies those minimums, NeMo Speech installs on top of it without replacing it, so your existing PyTorch build is kept (see the install options below). The versions pinned in uv.lock and shipped in the official container — Python 3.13, PyTorch 2.11 with CUDA 12.9 or PyTorch 2.12 with CUDA 13.2 — are simply the combinations we actively test and support. They make setup turnkey and reproducible, but they are not a hard requirement.
As of
, torch.load defaults to using weights_only=True. Some model checkpoints may require using weights_only=False. In this case, you can set the env var TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1 before running code that uses torch.load. However, this should only be done with trusted files. Loading files from untrusted sources with more than weights only can have the risk of arbitrary code execution.
Developer Documentation
VersionDescription3.0.0 (latest release)
NeMo Speech 3.0.0 documentation
Nightly
Documentation for the latest main branch
Install NeMo Speech
The recommended way to install NeMo Speech is from source with
, which reproduces our actively-tested stack from the committed uv.lock. If you need different Python/PyTorch/CUDA versions, NeMo Speech also installs over your existing environment via pip — see the
below.
From source with uv (recommended)
git clone https://github.com/NVIDIA-NeMo/Speech.git cd Speech uv sync --extra all --extra cu13 # CUDA 13.x (recommended) — use --extra cu12 for CUDA 12.xThis installs our supported stack (Python 3.13, PyTorch 2.12, CUDA 13.2) into .venv/ with NeMo Speech editable. Add --group test for the test suite or --group docs to build the docs; run tools via uv run <cmd> or activate with source .venv/bin/activate. On Linux, cu12 and cu13 are mutually exclusive — pass exactly one (cu13 is the default). For the exact container baseline, add --locked --python 3.13 (the path the Dockerfile and CI use).
SpeechLM2 / Automodel: the Automodel backend runs without any compiled dependencies. It can optionally benefit from dedicated accelerated backends (Transformer Engine, FlashAttention, Mamba, grouped-GEMM/MoE, DeepEP) for better performance — these source-built kernels come from the compiled (Hopper/Blackwell) or compiled-a100 (A100) extras, built by docker/Dockerfile (GPU_TARGET=h100plus / a100). See the
for the full list and build details.
Docker (turnkey, our supported stack)
The latest prebuilt NeMo Speech image is the
:
docker pull nvcr.io/nvidia/nemo-speech:26.07.00 docker run --rm -it --gpus all -v "$PWD:/workspace" nvcr.io/nvidia/nemo-speech:26.07.00 bashTo build the container from source (CUDA 13 / H100+ by default):
git clone https://github.com/NVIDIA-NeMo/Speech.git cd Speech docker buildx build -f docker/Dockerfile -t nemo-speech .# CUDA 13 / H100+ (default) docker run --rm -it --gpus all -v "$PWD:/workspace" nemo-speech bashFor A100, set GPU_TARGET=a100 — A100 works with both CUDA 12 and CUDA 13 (CUDA 13, the default base image, is recommended; the CUDA 12 base is a convenience). See the header of
for all build arguments (BASE_IMAGE, GPU_TARGET).
From PyPI with pip (fallback — bring your own versions)
Prefer your own Python/PyTorch/CUDA? Install your PyTorch first (any version ≥ 2.7 for your CPU/CUDA/etc. target — see the
), then add NeMo Speech and it keeps your build. uv pip (uv's fast, pip-compatible installer) works like pip:
uv pip install 'nemo-toolkit[asr,tts]'# or plain: pip install 'nemo-toolkit[asr,tts]'⚠️ Do not use uv sync --locked for a bring-your-own stack — it applies uv.lock and replaces your Python/PyTorch/CUDA with the supported baseline. Use uv pip/pip here; reserve uv sync --locked for reproducing our stack.
To instead pull our pinned PyTorch build, add the CUDA extra and the matching wheel index (pip/uv pip do not read uv's project index config, so --extra-index-url is required):
pip install 'nemo-toolkit[asr,tts,cu13]' --extra-index-url https://download.pytorch.org/whl/cu132 # CUDA 13.x pip install 'nemo-toolkit[asr,tts,cu12]' --extra-index-url https://download.pytorch.org/whl/cu129 # CUDA 12.xContribute to NeMo Speech
We welcome community contributions! Please refer to
for the process.
Licenses
NeMo Speech is licensed under the
.