whisper.py

import torch
from transformers import pipeline
from datasets import load_dataset

device = "cuda:0" if torch.cuda.is_available() else "cpu"

pipe = pipeline(
  "automatic-speech-recognition",
  model="openai/whisper-tiny.en",
  chunk_length_s=30,
  device=device,
)

ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
sample = ds[0]["audio"]

prediction = pipe(sample.copy(), batch_size=8)["text"]

# we can also return timestamps for the predictions
prediction = pipe(sample.copy(), batch_size=8, return_timestamps=True)["chunks"]