Whisper Large v3

Whisper Large v3 is OpenAI's open-weight speech-recognition checkpoint — an encoder-decoder transformer with the same architecture as Whisper large-v2 bar two changes: 128 Mel frequency bins (up from 80) and a new Cantonese language token. It was trained on roughly 1 million hours of weakly-labelled audio plus a further 4 million hours pseudo-labelled by large-v2 itself.

This registry entry covers Groq's hosted inference of that open checkpoint, not a self-hosted deployment. It matters because OpenAI's own hosted transcription API, whisper-1, still runs the older large-v2 checkpoint — so reaching large-v3 through Groq is a meaningfully more accurate option within the same Whisper lineage, and at $0.00185 per minute it is the cheapest STT option in this registry, undercutting whisper-1 as well.

OpenAI's release notes report a 10–20% relative error-rate reduction over large-v2 across most languages, with the largest gains on accented and non-native speech; no single absolute WER figure was published the way large-v2's were. Groq's pricing page currently states 217× real-time throughput for its hosting, up from an earlier independently-benchmarked 164× as its serving stack has been optimised.

The tradeoffs are structural: this is batch / file-upload only — there is no streaming endpoint, no native speaker diarization, and no custom vocabulary support — and Groq bills a 10-second minimum per request.