Can Hard Coded Hyperparameters be moved to a config file?

[morsczx]

Disclaimer - new to github & had a doubt In whisperx/audio.py - hyperparameters are hardcoded. can or should we move these to a config file so that we can edit these as required? specifically the chunk_length ? if so what file should be made? I can raise a PR basis suggestions.

SAMPLE_RATE = 16000
N_FFT = 400
HOP_LENGTH = 160
CHUNK_LENGTH = 30
N_SAMPLES = CHUNK_LENGTH * SAMPLE_RATE  # 480000 samples in a 30-second chunk
N_FRAMES = exact_div(N_SAMPLES, HOP_LENGTH)  # 3000 frames in a mel spectrogram input

N_SAMPLES_PER_TOKEN = HOP_LENGTH * 2  # the initial convolutions has stride 2
FRAMES_PER_SECOND = exact_div(SAMPLE_RATE, HOP_LENGTH)  # 10ms per audio frame
TOKENS_PER_SECOND = exact_div(SAMPLE_RATE, N_SAMPLES_PER_TOKEN)  # 20ms per audio token

[kevdawg94]

@morsczx did you resolve this issue? I was also trying to fine tune my segmentation performance and wondering whether chunk size would have an impact. Agree that these could be moved to config file.

[morsczx]

@kevdawg94 - no I did not resolve the issue - thought I. might have been wrong in the suggestion. I had a lot of audio clips with 10 second intervals of languages - which were being incorrectly identified - hence wanted control over the chunk size. will take this up

[kevdawg94]

I found you can already adjust chunk size as follows:

transcription = model.transcribe(audio, batch_size=batch_size, chunk_size=chunk_size)

[morsczx]

gotcha, thanks a lot!