Tutorial for Whisper

I am sharing below a tutorial I made for using OpenAI’s Whisper software, which allows you to transcribe interviews.

To download it as a presentation (in French), click here

Whisper

• Artificial intelligence trained on several language tasks.

• Based on Python.

“Whisper’s performance varies widely depending on the language. The figure shows a WER (Word Error Rate) breakdown by languages of the Fleurs dataset using the large-v2 model (The smaller the numbers, the better the performance).”

Google Colab

• So as not to make the computer run too hard and faster.

• Detailed instructions here: https://bytexd.com/how-to-use-whisper-a-free-speech-to-text-ai-tool-by-openai.

• Follow: https://colab.research.google.com/#create=true.

• Use GPU (Graphics Processing Unit): Runtime > Change runtime and in the drop-down menu Hardware Accelerator choose GPU, then save.

Using Whisper

• Install Whisper: !pip install git+https://github.com/openai/whisper.git. And ffmpeg: !sudo apt update && sudo apt install ffmpeg. Then crtl+enter.

• Type !whisper fichier.format. E.g.: !whisper test.mp3.

• If there is a space in the file name, enclose it in commas. E.g.: !whisper "my favorite interview.mp3".

• For the list of commands: !whisper --help.

• Automatic detection of the original language, but can be specified in advance. E.g.: --language Spanish. NB: all commands come after the file name. E.g.: !whisper test.mp3 --language Spanish.

• To translate, in English only: --task translate.

• Generates various formats at the end of the process, including .txt and .srt.

• Difficulty recognising the full stop and certain technical words and proper nouns.

Models

• Arbitration accuracy/transcription time. Small by default. Command to change: --model medium.

Local

• If confidentiality is a concern, it can be run locally.

• You need a terminal and a package manager. To install a manager:
  • MacOS: https://brew.sh
  • Windows: https://scoop.sh or https://chocolatey.org/
  • Linux: already installed.
• Installing Python and its dependencies: pip install git+https://github.com/openai/whisper.git. Then ffmpeg:

# on Ubuntu
sudo apt update && sudo apt install ffmpeg
# on MacOS with Homebrew (https://brew.sh/)
brew install ffmpeg
# on Windows with Scoop (https://scoop.sh/)
scoop install ffmpeg

• Command without exclamation mark. E.g.: whisper test.mp3 --language Spanish --model medium --task translate