Whisper

Whisper is a family of open-source speech recognition models released by OpenAI. It transcribes spoken audio into text and can translate speech from many languages directly into English. The weights are published under the permissive MIT license, so you can download them and run them on your own hardware.

The current open release is large-v3-turbo, a distilled model introduced in October 2024. It keeps almost all of the accuracy of the full large-v3 model while running about 8x faster, which makes it far more practical for batch jobs and near-real-time use. The full large-v3 checkpoint is still available when you want the highest possible quality.

Alongside the open models, OpenAI now offers hosted transcription through its API. The classic endpoint is whisper-1, and it has since been joined by GPT-4o-based models: gpt-4o-transcribe and the cheaper gpt-4o-mini-transcribe, which add options like speaker diarization. According to reports, a streaming variant, gpt-realtime-whisper, arrived around May 2026 and returns transcript deltas live as you speak.

Pricing splits into two tracks. The open-source models are free: you pay only for your own compute. The hosted API is usage-based: whisper-1 and gpt-4o-transcribe are $0.006 per minute ($0.36/hr), gpt-4o-mini-transcribe is $0.003 per minute ($0.18/hr), and the live gpt-realtime-whisper runs about $0.017 per minute.

Accuracy varies a lot by language. Whisper was trained on around 99 languages, but the quality is strongest for high-resource languages like English and drops noticeably for less common ones. Running the larger models locally also needs a capable GPU, and the model can occasionally hallucinate text during long silences.

Whisper is a good fit if you want a free, self-hostable transcription engine, need broad multilingual coverage, or want a managed API without building your own pipeline. Choose the hosted GPT-4o models if you need diarization or live streaming out of the box.