openai/whisper is the official implementation of a general-purpose speech recognition model that turns speech in many languages into text. The model also performs translation into English and language identification (the task of determining which language was spoken). The repository was created on September 16, 2022, and the latest release published on GitHub is v20250625. The implementation language is Python, the license is MIT, and the star count is 110981.
Whisper is a general-purpose speech recognition model trained on a wide variety of speech data. The repository describes the model as robust speech recognition built through large-scale weak supervision. The README states that it handles speech recognition, translation, and language identification within a single transformer encoder-decoder architecture, and that it processes multiple speech tasks together in the form of token sequences (lists of text pieces). It ships trained speech models together with public weights, and notes that a command-line tool can transcribe audio files.
However, the README notes that the memory requirements by model size and the relative speed are measurements taken under conditions of transcribing English speech on an A100, and that actual performance may differ. Those figures cannot be read as a performance baseline for other environments. Of the published evaluation scores, only momentum is filled in, at 7.2 (out of 10), while the novelty, technical, and quality items are blank, so no basis for comparative evaluation is given. The README advises using the multilingual model rather than turbo when translating non-English speech into English. The performance information presented by this repository is limited to measurements under specific conditions.