Whisper Large v3: improved multilingual ASR trained on 5 million hours
In one sentence Whisper Large v3 reduces error rates on low-resource languages, improves timestamp accuracy and adds new language support, remaining the most widely deployed open-source ASR model.
Whisper is OpenAI's open-source speech recognition system that changed the industry standard. The large v3 version is a refined evolution that brings concrete improvements where it mattered most.
The most significant jump concerns low-resource languages — those languages that have few training data available, such as African languages, regional dialects, or idioms with little online presence. On these languages the error rate drops noticeably compared to the previous version.
Timestamp accuracy also improves: when Whisper v3 says "this word was spoken at second 3.42", you can trust it more than before. For those building automatic subtitles or speech search systems, this precision is fundamental.
The model was trained on a whopping 5 million hours of audio, an impressive amount that explains its robustness across different accents, noisy contexts and linguistic variations.
Despite the arrival of faster competitors, Whisper Large v3 remains the reference point for open-source ASR quality — the model against which everything is compared.
Companies
OpenAI
Tools
—
Tags
Sources