Whisper popularized the idea that a single multilingual model could be downloaded and run without calling a hosted API. Under the hood it maps audio spectrograms to text tokens, often in 30-second chunks for long files. That chunking is why long-form transcription can feel batch-like even when a product shows progress. Fine-tuned or distilled descendants exist, but the public family is still the reference people mean when they say they run Whisper locally. It is a model, not a dictation app: you still need capture, a hotkey, and a way to insert text.
Because the weights are open, vendors wrap them in different runtimes — some optimized for Apple silicon, some for CPUs, some for servers. Those wrappers change latency and battery use more than they change the core idea. Whisper is often slower than specialized streaming models on short bursts, and stronger on messy, multilingual, or long audio. Word error rate varies by domain: read speech scores better than overlapping conversation. If a product says it uses Whisper, ask whether it runs on-device, which size, and whether a cloud fallback is the actual default.