Transcribe Audio Locally With Whisper
OpenAI's Whisper command-line package can transcribe a local recording. The initial setup and model download need network access. Once the dependencies and model are available, the command can process the recording locally.
Worked example
English test sentence to record
Mina booked room B-12 for nine thirty.
CPU transcription command
whisper standup.mp3 --model base.en --language en --device cpu --fp16 False --output_format txt --output_dir transcripts
Steps
- Follow the Whisper project's setup instructions and put FFmpeg on your command path. If your computer has multiple Python versions, install openai-whisper into the environment you plan to use.
- Record a short English sample and save it as standup.mp3 next to your terminal's working directory. For a WAV or M4A recording, substitute that filename in the command.
- Run the command and open transcripts/standup.txt. Play the original recording beside it and check names, quantities and speech during background noise.
What to check
- The text should preserve Mina, B-12 and the time. Accept equivalent punctuation or number formatting if it fits the destination.
- For audio in another language, choose a multilingual model and supply the language actually spoken.