centre-for-humanities-computing / whisper-transcription

This is a repo for anyone who wants to transcribe audio using whisper. The intention is for the transcriptions to be done on uCloud so the whole process is GDPR compliant.

Instructions on transcribing audio files

1. Make sure the data is on uCloud
2. Create an instance using JupyterLab 3.6.1 from the "Apps" tab on the lefthand-side panel
   • Give your instance a name (where it says job name)
   • Choose hours (this should roughly correspond to the number of hours of audio you have)
   • Choose the largest machine type (u1-standard-64 or u1-standard-32)
   • Scroll down to where it says "Select folders to use" and click the "Add folder" button
   • In the second dropdown select the "tools" folder and click the "Use this folder" button in the top righthand corner
   • Click "Add folder" again and select your folder from the second drop down then click the "Use this folder" button.
   • At this point you should have two folders that will be loaded: tools and the folder where your audio files are located.
   • In the top righthand corner, click on the button "Submit"
   • It will say "Your job is being prepared"
3. Click "open terminal" and run the following command where <folder_name> is the name of the folder. This will create csv files containing all the transcriptions
   • ./700623/transcribe.sh ./<folder_name>/data
4. If you also want docx files you can run the following command
   • ./700623/transcribe.sh ./<folder_name>/data --to_plain_text
5. If you also want docx files you can run the following command
   • ./700623/transcribe.sh ./<folder_name>/data --to_plain_text --to_docx
6. If you want docx and pdf files in addition to the csv then you can run the following:
   • ./700623/transcribe.sh ./<folder_name>/data --to_plain_text --to_docx --to_pdf