OpenAI open-sources Whisper
Trained on 680,000 hours of web-scraped audio and released under the MIT licence, an unusual openness for a company otherwise moving toward closed models.
- Open weights & ecosystem
- Models & capabilities
- Notable
OpenAI released Whisper, an automatic speech recognition system, and published its code and trained model weights on GitHub under the MIT licence — unusual openness for a company that had otherwise been moving toward keeping its models closed. Introducing Whisper described a model trained on 680,000 hours of audio collected from the web, roughly a third of it in languages other than English.
Unlike systems trained on curated, professionally transcribed corpora, Whisper’s training data was “weakly supervised” — noisy transcripts scraped alongside audio rather than produced under controlled conditions. OpenAI said the resulting model was markedly more robust to accents, background noise and technical vocabulary than prior systems, with English transcription accuracy approaching that of human transcribers. Performance varied considerably by language, with higher error rates for languages underrepresented in the training data, and the accompanying paper, “Robust Speech Recognition via Large-Scale Weak Supervision,” noted the model could hallucinate words that were never spoken — a consequence of training on noisy web audio rather than clean transcripts.
Several model sizes were released, trading accuracy against inference speed. OpenAI said its primary intended users were researchers studying robustness and generalisation rather than developers building production transcription tools, though it was quickly adopted for exactly that: folded into open-source captioning tools, dictation software and other companies’ speech products within weeks.
The release sat awkwardly against the company’s name. Publishing weights rather than a hosted API meant Whisper could be run offline, fine-tuned and embedded in other products without OpenAI’s involvement or ongoing revenue. It remains one of the few OpenAI models released this way, cited in later arguments about the company’s departure from openness as evidence that the shift, whatever its cause, was neither total nor applied uniformly across product lines.