AKA Spoken audio to text.
I found myself in a position of recording an audio conference and then wanting to get a transcription of the discussion. On my Android phone, there seems to be a plethora of apps in the store you can use, but my content was an hour and a half long. The best I could get was 5 minutes of transcript for free, or buy minutes.
After a bit of browsing, I found there were products available on Linux as Free and Open Source Software. But how do I use them? I found some quite complicated. All I wanted was here’s the file, now give me text. Many required training and had command lines I could not be bothered to figure out.
I settled on openai-whisper from https://github.com/openai/whisper
First, you need to resample your audio to mono and 16k. To do this I used ffmpeg, but understand I could have used sox. Then I take the mono output .wav file and run it through the whisper.cpp-medium.en program.
Install ffmpeg and whisper in Manjaro using
pamac install whisper.cpp-medium.en ffmegConvert the audio file to mono, 16k .wav, and run the .wav through whisper.
ffmpeg -i input_file.m4a -ar 16000 -ac 1 output_file.wav
whisper.cpp-medium.en output_file.wav --output-txt > transcript.txtNow, this took a fair old time, probably as long as it takes to play the audio. Even on an Intel i9-9900k with 16-cores. But the output text file is super quality.