Whisper

Whisper is OpenAI's open-source automatic speech recognition model, released in 2022 and trained on roughly 680,000 hours of multilingual and multitask audio scraped from the web. It transcribes speech to text, translates non-English speech into English, and detects language, handling accents, background noise, and domain-specific vocabulary noticeably better than the tools that preceded it. It comes in several sizes, from tiny to large, letting developers trade off speed against accuracy depending on whether they need real-time transcription on modest hardware or the highest possible accuracy offline.

Because the model weights are free to download and run locally, or call through OpenAI's paid API, Whisper has become the transcription engine underneath a huge number of other products rather than something end users interact with directly. Meeting-notes apps, podcast transcription tools, subtitle generators, and voice interfaces all build on it, and its open license has spawned faster community forks like whisper.cpp and faster-whisper that optimize it for edge devices and low-latency use.

Its main advantages over older speech-to-text systems are robustness to noisy or accented audio and strong multilingual coverage without needing separate models per language. It isn't built for streaming or diarization out of the box, so many downstream tools add their own layers for real-time transcription or speaker separation on top of it.