Whisper (transcription audio)
OpenAI's audio transcription model, now the reference for turning speech into text. You give it an audio or video file, it returns a clean, punctuated, multilingual transcript. It's the brick that turns a podcast, a call or a voice memo into data the rest of your AI chain can use. It also returns each sentence with its timestamp: that is what lets an agent turn it into a subtitle file synced to the voice, then have FFmpeg burn it in with your font and colour, the CapCut subtitle step gone ([Louise Mourey](https://www.instagram.com/reel/Dd9gruJqyEn/)).
Strengths
- Solid transcription quality across many languages and noise conditions
- Available as a hosted API OR locally (open-source weights) depending on your privacy need
Limitations
- Diarisation (who speaks when) is not native, it needs an extra layer
- Running locally requires a GPU to stay fast on long files
- On a stretch with no speech it invents plausible text instead of returning nothing: cut the silences before transcribing (voice activity detection)
Best for
- Transcribing meetings, interviews or videos to make them searchable and summarisable
- Feeding a multimodal agent with text from audio sources