Long audio file, need help,.

[artyomboyko]

Hello all. I'm new to audio and Whisper. Can anyone give me some pointers. How to process a long audio file in parts? Is there any mechanism that allows slicing audio in Python code (something like batches)?

As I understand it, this piece of code trims the audio to a length of 30 seconds:

``python
audio = whisper.pad_or_trim(audio)

How can I process a long audio file correctly?

[glangford]

I haven't had to do this myself, but you could identify periods of silence and split at those times, using for example ffmpeg directly or a library like pydub. Would that work for your audio? How long is it?

See for example
https://askubuntu.com/questions/1264779/how-to-split-an-mp3-file-by-detecting-silent-parts

[artyomboyko]

This option allows you to split the file when feeding it into the pipeline, during model inference.

[mattymatty]

Solution found.

from transformers import pipeline

# Определим наличие графического ускорителя. Если доступен - используем.
pytorch_device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print("Current PyTorch device is set to", pytorch_device)


audio_file_path = "/content/CS25 I Stanford Seminar - Transformers United: DL Models that have revolutionized NLP, CV, RL/audio.mp3"

pipe = pipeline(model="openai/whisper-small", device=pytorch_device)
predictions = pipe(files["audio_path"], chunk_length_s=20, stride_length_s=(5))

predictions = predictions["text"]
print(predictions)

Where do you put this code?

[Funatiker]

Well. You put it in a Python file and execute it. However, it has two flaws:

1. The lines audio_file_path = … and predictions = … don't match. First the path to the audio file is defined, then something else is used. So you should replace files["audio_path"] by audio_file_path. Also you probably want to define a different path to an audio file.
2. import torch is missing

Besides that, the shown code converts an audio file to a text that is then printed in English. (If you want output in the original languages, set the kwargs for language to the original language and task to transcribe instead of translate. See https://huggingface.co/openai/whisper-large-v3 )

[artyomboyko]

@Funatiker Glad you figured out how to adapt someone else's code to solve your problem.

[Funatiker]

Yeah. Quite awesome that you posted your solution. So I could answer @mattymatty's question.