Looking for complete guide for how to get best performance for transcribing Chinese TV show using Whisper

[vlashada]

I have the mp4 files for Peppa Pig with Chinese dub. I am trying to make a transcript using OpenAI Whisper, but I am struggling to get high-quality output. I experience frequent hallucinations or repeats of words.

For preprocessing, I am removing the into and outro of the video. For postprocessing, I am removing lines if they are really long, have very many repeated words or have many non-Hanzi characters. These are my current options for the transcript:

res = model.transcribe(
    filename,
    temperature = (0.0, 0.2, 0.4),
    compression_ratio_threshold = 1.3,
    logprob_threshold = -0.5,
    no_speech_threshold = 0.6,
    condition_on_previous_text = True,
    initial_prompt = "",
    word_timestamps = False,
    prepend_punctuations = "\"'“¿([{-",
    append_punctuations = "\"'.。,，!！?？:：”)]}、",
    language = "Chinese",
    sample_len = None,
    best_of = 20,
    beam_size = 5,
    patience = None,
    length_penalty = 0.1,
    prompt = None,
    prefix = None,
    suppress_tokens = "-1",
    suppress_blank = True,
    without_timestamps = False,
    max_initial_timestamp = 1.0,
    fp16 = True,
)

I have tried looking for a comprehensive guide into what all the parameters are, and what the best strategy is for getting the best performance, but I am struggling to find any good resources. I have a Quadro RTX 6000, so I am easily able to run large-v3.

[dgoryeo]

Hi @vlashada , I came across this old post. I'm keen to hear if you have found better strategies / settings than above parameter values.