too slow

[zalastone]

I love how accurate this implementation of whisper but I am shocked at how slow it is to process one episode of tv show
I am using an RTX 3070
and it takes hours to finish a 150-minute video. I am getting an average of 99.53frames/s

whisper_timestamped 'video.mp4' --language ko --model large-v2 --task transcribe --output_format vtt --output_dir . --recompute_all_timestamps True --temperature 0.3 --best_of 2 --device cuda

can I improve it ?