Whisper: OpenAI just gave away human-level speech recognition. Your meetings became data

🎧 Whisper: OpenAI just gave away human-level speech recognition. Your meetings became data.

Last week OpenAI open-sourced Whisper — a speech model trained on 680,000 hours of multilingual audio, approaching human-level robustness on English. The part that made my team sit up: it is MIT-licensed. Download the weights, run it on your own machines, pay nobody per minute.

Three reads on why this matters:

1. A priced capability just became free. Transcription has been a per-minute API business — metered, and your audio leaves the building. A free model inside your own boundary changes the economics and the data flow. Worth watching whether this pattern — labs giving away what vendors charge for — repeats. My bet: it will.

2. Your largest untapped dataset is spoken. Meetings, sales calls, support conversations — enterprises run on speech, and almost none of it becomes usable data. When every recorded conversation can become searchable text on your own servers, a whole layer of the organisation becomes analysable for the first time.

3. Accent robustness is the headline for us. Models trained on clean American English have always stumbled in Indian enterprise reality — accents, code-switching, jargon. A model trained on messy real-world audio is built for our conditions.

The cautions: not real-time out of the box; it occasionally transcribes words nobody said, so review before you trust; and a transcript is the most sensitive document type there is — names, numbers, candid opinions, all searchable. Decide the consent, retention and access rules before the transcription pipeline exists, not after.

The habit worth building: every time a capability drops from priced-and-hosted to free-and-yours, re-run the list of ideas you rejected as too expensive. Some of them just became projects.