Tiny, Base, Small or Medium: choosing a Whisper model for your phone

Whisper is OpenAI’s open speech recognition model, released under the MIT license. It comes in several sizes, and on a phone the size you pick decides how much storage it takes, how much memory it needs and how long you wait.

The file is only part of the cost

The download size is the number everyone sees, but a model also needs working memory while it runs. EchoLocal uses whisper.cpp, a C/C++ port of Whisper built to run efficiently on ordinary hardware. Its figures show the gap clearly.

Memory is usually what rules a model out. Your phone may have room for a 1.5 GB file and still struggle to spare about 2.1 GB of memory alongside everything else it’s doing. That’s why EchoLocal marks a model “Not recommended” when it looks too large for your phone, and tells you why.

Bigger is slower

OpenAI publishes rough speed figures alongside each model’s size. Treat the ratios as a guide to the trend only. They were measured on a data center GPU transcribing English, so your phone’s real speed will be different.

Where each one fits

A ten-minute test

• Pick a hard clip. Choose two or three minutes of the kind of audio you actually record, including the noise and the names.
• Run it on Tiny. It’s already installed. Time it and note what it gets wrong.
• Run it on Base. It’s free too. If the errors now don’t matter to you, stop here.
• Consider Small or Medium if errors still cost you time. These are part of Pro. Compare the time they’d save in editing with the longer wait.

Try it in your own language too. Accuracy varies by language, voice and model, so a result in English may not carry over to Urdu or Korean.

You can compare the models and see their sizes in EchoLocal. Tiny and Base are free, and models you’ve downloaded keep working without a connection.