What it does
NVIDIA NeMo Speech is a code repository for researchers and PyTorch developers who work on speech models: automatic speech recognition, text to speech and speech-focused large language models. It is meant to help people create, customize and deploy new models by building on existing code and pre-trained checkpoints. The project page lists recent model releases, including streaming speech recognition models covering many languages and a multilingual text-to-speech model, and points to a Hugging Face collection of open-weight checkpoints and demos. It requires Python 3.12 or newer and PyTorch 2.7 or newer, plus an NVIDIA GPU with CUDA for training (recommended for inference). It can be installed from source with uv, with pip as the nemo-toolkit package, or through a prebuilt NGC container run with Docker. The code is released under the Apache-2.0 licence.
Capabilities
- Training and customizing speech recognition and text-to-speech models
- Pre-trained checkpoints to build on instead of starting from scratch
- Support for speech-focused large language models
- Topic tags include speaker recognition, diarization and speech translation
- Install from source with uv, with pip, or as a Docker container
When to use it
- You are training or customizing speech recognition or text-to-speech models in PyTorch
- You want to start research from published speech checkpoints
- You need a GPU-based speech pipeline you can run in a container
