Whisper Large v3

Whisper Large v3 is an Audio automatic speech recognition (ASR) model from OpenAI that transcribes and translates audio across 99 languages with state-of-the-art accuracy, available completely free under the Apache 2.0 license. It is designed for developers, researchers, and ML engineers who need a powerful, open-weight ASR foundation for building transcription pipelines.

Released on November 7, 2023 and hosted on Hugging Face, Whisper Large v3 has been downloaded over 118 million times all-time and roughly 4.8 million times per month, with more than 5,600 likes from the community. The model was trained on 1 million hours of weakly labeled audio plus 4 million hours of pseudo-labeled audio generated by Whisper Large v2, for 2.0 epochs over the mixture dataset. Compared to Large v2, it delivers a 10% to 20% reduction in errors across a wide variety of languages, and it scores a 7.44 average word error rate on the Open ASR Leaderboard benchmark. Key architectural changes include a 128 Mel frequency bin spectrogram input (up from 80) and an added language token for Cantonese, extending coverage to 99 languages.

The model supports both transcription (source language → same language text) and translation-to-English via a simple task flag, and it can output sentence-level or word-level timestamps. It natively supports audios up to 30 seconds and handles longer files via sequential (sliding-window) or chunked (parallel) long-form algorithms through the Hugging Face Transformers pipeline. Based on our analysis of 870+ AI tools in the aitoolsatlas.ai directory, Whisper Large v3 stands out as the most-downloaded open-weight ASR model available, and unlike hosted alternatives such as AssemblyAI, Deepgram, or Rev.ai, it can be self-hosted on your own GPU with zero per-minute usage fees. It is accessible through three inference providers on Hugging Face (Replicate, hf-inference, and fal-ai) for teams that prefer a managed API, while still offering full weights for on-prem deployment.