Loading page
Whisper V3 Large is OpenAI's automatic speech recognition model, used for multilingual audio transcription and speech translation into English.
Prices updated daily. Last check: Sep 6, 2026
Whisper V3 Large fits workloads where transcription accuracy matters more than latency: transcribing podcasts, interviews, lectures, and recorded meetings; generating subtitles and closed captions with segment timestamps; producing searchable text archives from large audio or video libraries; and processing multilingual voice notes or support calls where the input language is not known in advance. Its translation mode makes it useful for turning non-English recordings into English transcripts in a single step. It is also frequently used as the speech-to-text stage in voice assistant and agent pipelines, though builders targeting conversational latency often pair it with chunked streaming logic or choose a smaller Whisper checkpoint or a purpose-built streaming ASR service instead. For tasks needing speaker labels, sentiment, or summarization, Whisper handles the transcription step and downstream models handle the rest.
Pricing depends on the provider and the billing model — speech-to-text is usually charged per minute or per hour of audio processed, and some hosts also offer batch or self-hosted GPU options priced by compute time. Because the weights are open, the same model is offered by multiple providers at different rates. Check the pricing table on this page for current per-provider pricing.
Batch transcription of recorded audio where accuracy is the priority: podcasts, meetings, interviews, lectures, and media archives, plus subtitle generation and multilingual transcription or speech-to-English translation.
Whisper V3 Large is the largest checkpoint in the family. It generally produces more accurate transcripts, especially on accented, noisy, or non-English audio, at the cost of slower inference and higher compute requirements than the tiny, base, small, and medium variants.
It was designed for file-based transcription over short fixed audio windows, not native streaming. Near-real-time use is possible by feeding overlapping audio chunks, but latency and stability depend on the serving implementation rather than the model itself.
No — diarization is not part of the model. Speaker labels require an additional diarization model or a provider pipeline that combines Whisper output with speaker segmentation.
Yes. OpenAI released Whisper weights openly, and the model runs through Hugging Face Transformers as well as optimized runtimes such as faster-whisper/CTranslate2 and whisper.cpp, including quantized CPU builds. Self-hosting shifts the cost to GPU or CPU rental, which you can compare against per-minute API rates.