Using transcription options repetition_penalty, and no_repeat_ngram_size

[dgoryeo]

faster-whisper has couple of transcription options that can help with hallucination and repetition:

    repetition_penalty: float
    no_repeat_ngram_size: int

I wonder if those options can already be passed to fatser-whisper in the python call? Also whether it would be an idea to have those options as part of stable-ts (CLI) as well.

[jianfch]

You can pass it directly in python model.transcribe_stable(..., repetition_penalty=1.5).

Some option availabilities change depending on certain augments, so they are not hardcoded for CLI. But you can still use them via these CLI augments:

stable-ts/stable_whisper/whisper_word_level.py

Lines 1334 to 1348 in d51edb6

	parser.add_argument('--refine_option', '-ro', action="extend", nargs='+', type=str,
	help='Extra option(s) to use for refining timestamps; Replace True/False with 1/0; '
	'E.g. --refine_option "steps=sese" --refine_options "rel_prob_decrease=0.05"')
	parser.add_argument('--demucs_option', '-do', action="extend", nargs='+', type=str,
	help='Extra option(s) to use for demucs; Replace True/False with 1/0; '
	'E.g. --demucs_option "shifts=3" --demucs_options "overlap=0.5"')
	parser.add_argument('--model_option', '-mo', action="extend", nargs='+', type=str,
	help='Extra option(s) to use for loading model; Replace True/False with 1/0; '
	'E.g. --model_option "download_root=./downloads"')
	parser.add_argument('--transcribe_option', '-to', action="extend", nargs='+', type=str,
	help='Extra option(s) to use for transcribing/alignment/locating; Replace True/False with 1/0; '
	'E.g. --transcribe_option "ignore_compatibility=1"')
	parser.add_argument('--save_option', '-so', action="extend", nargs='+', type=str,
	help='Extra option(s) to use for text outputs; Replace True/False with 1/0; '
	'E.g. --save_option "highlight_color=ffffff"')

For your case:

stable-ts audio.mp3 --faster_whisper -to "repetition_penalty=1.5" "no_repeat_ngram_size=2" -o output.srt

You can use --debug to see all the python methods and arguments used in the background for the CLI.

[dgoryeo]

Thanks for the quick reply.
Meanwhile, wish you a great year of 2024 ahead.
And thanks for 2023 and all the good work with stable-ts. It has become so geature rich.