Using Whisper to Transcribe Podcasts

Prerequisites

these are onetime setup things

brew install ffmpeg # takes a while!
git clone https://github.com/ggerganov/whisper.cpp
cd whisper.cpp
make
cd models
./download-ggml-model.sh base.en or ./download-ggml-model.sh medium.en # see full model list here https://github.com/ggerganov/whisper.cpp/tree/master/models

Steps

1. convert your audio file to a 16khz .wav file: ffmpeg -i SOURCE_FILE.wav -ar 16000 output.wav
2. THEN you can do ./main -m models/ggml-medium.en.bin -f output.wav >> output.txt inside of the whisper.cpp directory, which pipes the transcription into output.txt

• at a rate of about 3 minutes of input: 2 minutes to transcribe (for the medium - 769M param model)
• or at a rate of about 12 minutes of input: 1 minute to transcribe (for the base - 74M param model)