I made a modified Google Colab notebook with more options (line width, line count, speaker labels, etc.)

[BlohoJo]

https://colab.research.google.com/drive/1k6MXAdmBgj1-cBGFmPQB9shwjnY0m98L [Download]

# Name of the audio file
audio_path = "Commentary.wav"

# Whether to enable music removal from speech, helps increase diarization quality but uses alot of ram
enable_stemming = True

# (choose from 'tiny.en', 'tiny', 'base.en', 'base', 'small.en', 'small', 'medium.en', 'medium', 'large-v1', 'large-v2', 'large-v3', 'large', 'turbo')
whisper_model_name = "large-v3"

# replaces numerical digits with their pronounciation, increases diarization accuracy
suppress_numerals = False
batch_size = 8
language = None  # autodetect language
device = "cuda" if torch.cuda.is_available() else "cpu"

# --- Subtitle Formatting Options ---
MAX_LINE_WIDTH = 42            # Max characters per line of subtitle
MAX_LINE_COUNT = 2             # Max lines per subtitle block
MAX_DURATION_SECONDS = 8.0     # Max time any subtitle is displayed
MIN_SPLIT_SENTENCE_LENGTH = 42 # Balance sentence chunks so orphans aren't left alone!

# --- Speaker Label Options ---
INCLUDE_SPEAKER_IN_LENGTH = True      # True: "Speaker X: " characters count towards max line width and min sentence splits
ONLY_LABEL_SPEAKER_CHANGES = True     # True: Only show "Speaker X:" when a new speaker starts talking

# --- Speaker Diarization Options ---
MAX_SPEAKERS = 8                      # Maximum number of speakers the diarization model can detect

The notebook has some "auto balancing" code for lines so line breaks don't result in a second subtitle line that is only one or two words.

It also tries to balance sentences so that you don't have one subtitle that is two lines long, then the next subtitle has only one one or two words. You can control this using the MIN_SPLIT_SENTENCE_LENGTH variable, so if you want, you can have subtitles that end sentences that are shorter. But I recommend leaving it at the defaults for the best balance.

Of course it's not perfect, you have to go through the SRT and look for glitches and quirks.

If your speaker says a lot of years ("...in 1976, we were..."), I recommend keeping suppress_numerals set to False.

Finally, I had problems using large-v2 where it would miss entire sentences at certain points, especially when the speaker is about to quote someone. It will tend to miss the quote, then pick back up at the next sentence, or maybe even a few sentences later. Sometimes the subtitles will be misaligned at these points, but will pick back up later. Anyway, that went away when I switched to large-v3.

I haven't thoroughly tested the ONLY_LABEL_SPEAKER_CHANGES functionality yet. Try it out! If it messes up, let me know.