I ran Whisper on a five-year-old laptop

Seven minutes of lecture, transcribed in 31 seconds, 99% right, on an M1 Pro from 2021. Nothing left the machine.

Seven minutes of lecture. Thirty-one seconds to transcribe it. Ninety-nine percent of the words right. On an M1 Pro from 2021, and nothing left the machine.

Speech-to-text is solved. It runs on a laptop. And most people are still uploading their voice to someone else's server to get a transcript.

The setup

OpenAI's Whisper turbo model, running through whisper.cpp, which uses the Mac's GPU. The whole thing is four commands:

brew install whisper-cpp
curl -LO https://huggingface.co/ggerganov/whisper.cpp/resolve/main/ggml-large-v3-turbo.bin
ffmpeg -i lecture.m4a -ar 16000 -ac 1 -c:a pcm_s16le lecture.wav
whisper-cli -m ggml-large-v3-turbo.bin -f lecture.wav -l en -otxt

The model is a 1.5 GB download. That's it. No account, no API key, no subscription.

I didn't have a lecture recording I could publish, so I wrote one. Seven minutes on how neural networks learn, full of jargon like backpropagation and softmax, read out loud by my Mac's built-in voice. That way I know every word that was actually said, and I can score it.

The numbers

The first run took a minute, most of it the GPU setting itself up. Every run after that: 31 seconds.

That's 14 times faster than real time. A 90-minute lecture would be done in about six and a half minutes, while you pack your bag.

It used about 2 GB of memory. My laptop has 32.

Out of 1,166 words, it got 11 wrong, once you ignore British versus American spelling. The mistakes are the fun part:

  • "loss landscape" became "lost landscape"
  • "Take ten examples" became "Click 10 examples"
  • "Take a step, and repeat" became "Take a step, take a step, and repeat"
  • "the key idea of the whole course" became "the key of the idea of the whole course"

To be fair to real lectures: a computer voice in a quiet room is the easy case. A real lecturer, a bad microphone and a room full of coughing will do worse. But a few years ago, the easy case was hard too.

Then I made it think

The obvious next step: hand the transcript to a reasoning model and let it fix the mistakes. So I ran Qwen3 8B through Ollama, on the same laptop, and asked it to fix transcription errors and change nothing else.

It thought for about 1,800 words, took 3 minutes and 16 seconds, and fixed a few of the mistakes.

Three minutes of thinking for a few fixes. Whisper took 31 seconds to get everything else right.

Money

Running it locally isn't about money. OpenAI's hosted Whisper costs $0.006 a minute at the time of writing. A 90-minute lecture is 54 cents. A whole term of lectures is about the price of a textbook.

Do it locally for the other reasons. The audio never leaves your machine. It works on a train. And you know exactly what every step is doing.

Verdict

Whisper turbo on whisper.cpp: yes. Free, private, and done before you've left the room.

The reasoning pass: only if the exact words really matter. The small fast model did 99% of the work. The big thinking model polished the last 1%, slowly.

Sources

Next: Software is about to be free