whisper-ct2

Whisper for CTranslate2

A port of Open-AI Whisper for CTranslate2.

Repositories:

• Whisper original: https://github.com/openai/whisper
• Faster Whisper: https://github.com/guillaumekln/faster-whisper

Tested with:

• Arm64 - Debian 11 - Python 3.9
• x86_64 - Debian 11 - Python 3.9

Tweaking results

• Use python3 test.py -h to check-out the possible arguments
• Language is set to auto by default, which will use the code of the test files (en_...) if given. If not it will use the Whisper detector. Use --lang en etc. to skip detection.
• Beam-size is 1 by default which is technically the fastest decoding strategy, but it might be worth to play around with it (e.g. --beamsize 4 or --beamsize 64) to see how it affects speed and accuracy.
• There is a magical Whisper parameter called 'initial_prompt' that will prime the model in some unpredictable way ^^. It can influence accuracy and style of the transcription. Example: --init-prompt "Hey SEPIA, set a timer.