nemo-toolkit

What it is and what it does

nemo-toolkit is NVIDIA's framework for speech AI, built on PyTorch to help researchers and developers train and deploy models for automatic speech recognition, text-to-speech synthesis, and speech-based language models. It ships with pre-trained checkpoints (Parakeet, Canary, MagpieTTS, Nemotron-Speech) covering multiple languages and use cases, and supports both offline and streaming inference with configurable latency-accuracy tradeoffs.

The package installs as a pure-Python wheel over your existing PyTorch and CUDA setup without replacing them. It depends on 15 runtime packages including torch, numba, scikit-learn, huggingface_hub, and tensorboard. Training requires an NVIDIA GPU; inference can run on CPU but GPU is recommended. The repository is actively maintained (last commit 2026-08-14) and carries no known vulnerabilities.

Use it for:

• Build automatic speech recognition systems for English or 25+ European languages using pre-trained Parakeet or Canary checkpoints
• Deploy streaming ASR with configurable latency (80ms–1s) using Nemotron-3.5-ASR-Streaming for real-time transcription
• Generate multilingual speech synthesis using MagpieTTS with support for 9 languages including English, Spanish, French, and Mandarin
• Fine-tune or customize speech models on your own audio data using the modular PyTorch-based training pipeline
• Integrate speech recognition and translation into conversational AI applications using Nemotron VoiceChat or speech LLM components

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

nemo-toolkit provides a PyTorch framework for building, training, and deploying speech AI models including automatic speech recognition (ASR), text-to-speech (TTS), and speech-based language models.

Yes. nemo-toolkit is actively maintained, carries no known vulnerabilities, and offers low install friction. It is well-suited for researchers and developers building speech AI systems. Requires PyTorch 2.7+, Python 3.10+, and ideally an NVIDIA GPU; if your environment already meets these, installation is straightforward. The Apache 2.0 license permits commercial use. Install if you need to work with ASR, TTS, or speech-based language models; skip if you need only inference on pre-trained models without customization or if you lack GPU access.

Install

nemo-toolkit on PyPI

pip install nemo-toolkit

uv add nemo-toolkit

poetry add nemo-toolkit

Installing nemo-toolkit

Before you install

Low install friction with a pure-Python wheel. Actively maintained with a release 7 days old and 18128 repository stars. Requires PyTorch 2.7 or above and Python 3.10 or above; GPU with CUDA is required for training but optional for inference.

License in practice

Apache License 2.0 permits commercial use, modification, and distribution with attribution. You may use this in proprietary projects provided you include a copy of the license and document any changes to the source.

Quickstart

pip install nemo-toolkit
import nemo.collections.asr as nemo_asr
asr_model = nemo_asr.models.EncDecCTCModel.from_pretrained('nvidia/parakeet-ctc-0.6b')
transcriptions = asr_model.transcribe(['audio.wav'])

Requires PyTorch 2.7 or above and Python 3.10 or above. NVIDIA GPU with CUDA is required for training; for inference, CPU is possible but GPU is recommended. Some model checkpoints may require setting TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1 environment variable.

Verify before relying

• Whether all 15 runtime dependencies (including torch, numba, cuda-bindings) install without version conflicts in typical environments
• Performance characteristics and latency of inference on CPU versus GPU
• Compatibility matrix with specific CUDA versions beyond the stated 12.6/13.2 support

Package facts

Evidence: nemo_toolkit-3.0.0-py3-none-any.whl

Keywords: NLP, NeMo, deep, gpu, language, learning, learning, machine, nvidia, pytorch, speech, torch, tts

Tags

More Libraries packages

Further reading