generate_caption.py

import os
import time 
from pathlib import Path
import torch
from transformers import pipeline
torch.cuda.empty_cache()


if __name__ == "__main__":

    folder_path = "./part_1"

    for roots, dirs, files in os.walk(folder_path):
        for file in files:

            if file.lower().endswith('.en.srt') | file.lower().endswith('.en-orig.srt'):
                print('[srt] file are found.')

                
            else:
                if file.lower().endswith('.mp4'):
                    video_path = Path(roots+'/'+file)
                    
                    if video_path.exists():
                        split_video_path = str(video_path).split('.mp4')[-2]
                        create_srt_path = split_video_path.__add__('.en.srt')
                        create_srt_path = Path(create_srt_path)
                        # if create_srt_path.exists() and create_srt_path.read_text(encoding="utf-8"):
                        if create_srt_path.exists() and create_srt_path.stat().st_size > 0:
                            # content = create_srt_path.read_text(encoding="utf-8")
                            print(f"content are already found in [srt] file. filepath: >>>> {create_srt_path}")
                            
                        else:
                            

                            # Initialize the pipeline (handles chunking intelligently)
                            device = "cuda" if torch.cuda.is_available() else "cpu"
                            pipe = pipeline(
                                "automatic-speech-recognition",
                                model="openai/whisper-large-v3",
                                torch_dtype=torch.float16,
                                device=device,
                            )

                            # Generate with built-in chunking and stride
                            # chunk_length_s=30: Process 30s at a time
                            # stride_length_s=5: Overlap chunks by 5s to connect sentences correctly
                            outputs = pipe(
                                str(video_path),
                                chunk_length_s=30,
                                stride_length_s=5,
                                generate_kwargs={"task": "translate", "language": "hindi"}
                            )

                            # print(outputs["text"])
                            print(">>>>>>>>>>>>>>>>>>> :", create_srt_path)

                            torch.cuda.empty_cache()

                            # Save it so you don't lose it!
                            with open(create_srt_path, "w") as f:
                                f.write(outputs["text"])

                    else:
                        print('video_path are not working...')

            # ---------------------------------------------------------------