Pharos is taking you through this one. Free, and it runs right here. No signup.
Learn Whisper in 5 minutes.
Whisper is OpenAI's open speech-to-text model: feed it audio, get back text, in almost any language, with timestamps if you want them. Scholé uses it when agents narrate. Your five minutes go to running it on a real clip and judging the output the way a product would.
Lesson not loading? Open it in a new tab. It is the same free lesson either way.
Model size, language, and timestamps
Whisper is a family of models, not one. The choices that matter are how big a model you run, whether you tell it the language, and whether you need word timings.
Tiny to large
Small models run fast on a laptop and miss more words; large models are slow and accurate. Products usually pick the smallest one that is good enough on their audio.
Detect or declare
Whisper can guess the language, but telling it saves errors on short clips and mixed accents. It also translates to English if asked.
Text with a clock
Segment and word timestamps let you build captions, jump-to-quote, and karaoke-style highlighting. Ask for them only when you need them; they cost accuracy.
What you will do in the lesson
You answer and try things. Nobody demonstrates at you for five minutes.
- Transcribe a clean clip. You run a thirty-second clip and mark every word the model got wrong.
- Break it with noise. You feed a noisy clip and see which mistakes appear and which model size fixes them.
- Declare the language. You transcribe an accented clip with and without the language set and compare.
- Get timestamps. You request segment timings and turn them into a caption file.
- Pick a model for a product. You end with a latency budget and an accuracy floor and choose a size and settings.
Before you start
Is Whisper free?
The model weights are open and free to run yourself. OpenAI and others also sell it as an API, priced per minute of audio.
Can it run on my laptop?
Yes, the small and base models run fine on a modern laptop. Large models want a GPU or the API.
How does it compare with dictation apps?
Dictation apps like Wispr Flow are products built for typing by voice. Whisper is the engine you would build such a product on.
What languages does it handle?
Around a hundred, with accuracy highest for English and other well-represented languages. Test yours before relying on it.
Other tools people learn next
Same format, same five minutes, still free.
Learners from these organizations are already learning on 
