Whisper Large v3

Whisper Large v3 is the third generation of OpenAI's Whisper family, introduced by Alec Radford and colleagues in the paper Robust Speech Recognition via Large-Scale Weak Supervision. It is designed for automatic speech recognition and speech translation, generalising zero-shot across many datasets and domains. The model retains the same Transformer encoder-decoder backbone used in the earlier large and large-v2 checkpoints, with only two small architectural changes: the input spectrogram now uses 128 Mel frequency bins instead of 80, and Cantonese was added as a new language token to broaden multilingual coverage.

Training scale is the main driver of v3's gains: OpenAI combined 1 million hours of weakly labelled audio with 4 million hours of pseudo-labelled audio produced by Whisper large-v2, and trained the model for 2.0 epochs over that mixture, building on the 5 million hours of labelled data lineage behind the family. The result is a 10% to 20% reduction in error rates across a wide variety of languages compared with large-v2, which makes it well suited to multilingual transcription, translation pipelines, and downstream audio analytics. Because the weights are openly published, developers can fine-tune or self-host the model, and it is also offered as a downloadable, DGX Cloud-accelerated NVIDIA NIM for production deployments.