Every major AI lab has now shipped an open-source large speech model for ASR, and picking the right one for production has become a real decision. This review puts the heavy hitters side by side: OpenAI's Whisper, Google's USM-adjacent releases, Meta's MMS and wav2vec 2.0 lineage, and Nvidia's NeMo family (Conformer, Parakeet, Canary). The goal is an honest comparison across the axes that actually matter to teams shipping speech products — accuracy on realistic audio, language coverage, latency, licensing, fine-tuning ergonomics, and the size of the community around each stack.

Rather than reciting leaderboard numbers, the segment focuses on where each model earns its keep and where it quietly falls down: long-form transcription, code-switching, domain adaptation, streaming, and the practical cost of running them at scale. It also flags the trade-offs between end-to-end Transformer decoders and CTC-style architectures when you need timestamps or hot-word biasing. If you're evaluating an open-source ASR foundation model for an in-house pipeline, this rundown is a solid starting map before you burn GPU hours benchmarking them all yourself.