How to use with faster whisper?

#1
by kristijanv - opened

Hello, how to use that model with faster whisper? https://github.com/SYSTRAN/faster-whisper

Worked using the command: ct2-transformers-converter --model primeline/distil-whisper-large-v3-german --output_dir primeline/distil-whisper-large-v3-german --copy_files preprocessor_config.json --quantization float16
But unfortunately this model is worse then the original. almost every third word is mistranslated

Wow! That's really damn fast! 550 tokens/sec! (RTX 4090). With WhisperX I only got 160 t/s with the standard large-3 model. With the destil model it was much faster (about 430 t/s but the german audio was output in english. No way to let it output in german. As I was told that model isn't able to do so. Happy that I found this repo!
GPU usage at 56% in maximum.

Worked using the command: ct2-transformers-converter --model primeline/distil-whisper-large-v3-german --output_dir primeline/distil-whisper-large-v3-german --copy_files preprocessor_config.json --quantization float16
But unfortunately this model is worse then the original. almost every third word is mistranslated

Is it bad at translating or transcribing?

Sign up or log in to comment