NVIDIA NeMo

NeMo is the open-research surface for NVIDIA's speech work. If your team can train and ship Parakeet-class models, you can match top commercial vendors on quality, and the underlying recipes are all in the repository for free.

For ML engineering teams that have GPU compute and the in-house ML chops to fine-tune transformer models, NeMo is the most credible open path to building a custom ASR system that competes with Deepgram or AssemblyAI on accuracy. The framework is heavy.

PyTorch Lightning under the hood, full training recipes for Parakeet TDT 1.1B and Canary multilingual models, with the kind of configuration surface area that lets you train on your own data, swap encoder architectures, and customise decoding behaviour. None of that is the right toolkit for someone who just wants an API; it's the right toolkit for a team building speech infrastructure as a core differentiator.

The Apache 2.0 licence on the models means you can use them commercially without licensing complications, but in practice the production deployment path runs through NVIDIA's Riva platform, which adds AI Enterprise licensing for support. Training requirements are GPU-heavy: fine-tuning a Parakeet model meaningfully requires multiple H100s and meaningful elapsed time, not something you do on a laptop.

For the right team it's the open-source path to commercial-grade ASR. For everyone else it's overkill.