NeMo Speech is NVIDIA’s open-source PyTorch framework for building speech AI: automatic speech recognition, text-to-speech and speech LLMs. It provides production-grade model architectures, pre-trained checkpoints and training recipes so researchers and developers can create, customize and deploy state-of-the-art voice models. The toolkit, formerly the speech collection of NVIDIA NeMo, became a dedicated repository and reached version 3.0 in 2026, with about 18,600 stars on GitHub.
NeMo Speech ships the model families behind NVIDIA’s best-known open checkpoints: Parakeet and Nemotron streaming ASR models that trade latency against accuracy, Canary models for recognition and translation across 25 European languages, and MagpieTTS for multilingual speech synthesis. Models are built from modular, configurable components on PyTorch Lightning-style training, with support for speaker diarization and recognition, speech translation and punctuation restoration. Distribution is flexible: install from source with uv, from PyPI, or run the official NGC container, and checkpoints are published openly on Hugging Face with demos. Recent releases even include streaming voice-chat systems built on Nemotron language models.
NeMo Speech is completely free and open source under the Apache-2.0 license. NVIDIA also offers hosted NIM microservices and enterprise platforms separately, but the toolkit and checkpoints themselves cost nothing.
NeMo Speech is for speech researchers and product teams building transcription, voice assistants, dubbing or call-center AI who want top-tier open models and full control over training.