Whisper V3 Turbo

Whisper V3 Turbo (also distributed as whisper-large-v3-turbo) is an automatic speech recognition model from OpenAI and part of the Whisper family of encoder-decoder transcription models. It is derived from Whisper large-v3 by substantially reducing the number of decoder layers while keeping the same audio encoder, which lowers decoding time per audio segment. OpenAI released the weights publicly under a permissive license, so the model is served both through hosted inference APIs and run locally via frameworks such as Hugging Face Transformers, faster-whisper/CTranslate2, and whisper.cpp. The model takes audio input and produces text output. Like other Whisper checkpoints, it processes audio in 30-second windows with log-Mel spectrogram features, and longer recordings are handled by chunking or sequential long-form decoding in the serving library. It supports multilingual transcription across the same broad language set as large-v3, plus speech translation into English, and can emit segment-level timestamps. Language can be auto-detected or pinned with a decoding parameter, and an initial prompt can be supplied to bias vocabulary and formatting. In practice Turbo is chosen when throughput and latency matter more than squeezing out the last increment of accuracy: batch transcription of media archives, captioning pipelines, voice-note ingestion, and near-real-time streaming front-ends. Compared with the full large-v3 checkpoint it trades some accuracy — most noticeably on harder or lower-resource languages and on speech translation — for faster decoding, while remaining considerably more capable than the small and base Whisper checkpoints. Hosted providers typically bill audio transcription per minute or per hour of audio rather than per token, so the cost model differs from chat LLMs; see the pricing table on this page for current provider rates.