app.py

import os
import re
import gradio as gr
import stable_whisper
import moviepy.editor as mp
from transformers import AutoModelForSeq2SeqLM
from transformers import AutoTokenizer
import torch
import time
from tqdm import tqdm 

def load_assets(TOKEN): 
  device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
  model = AutoModelForSeq2SeqLM.from_pretrained("InferciaNLP/NMT-SubGen", token=TOKEN)
  tokenizer = AutoTokenizer.from_pretrained("InferciaNLP/NMT-SubGen", token=TOKEN)
  model = model.to(device)
  return model, tokenizer


def translate_text(model, tokenizer, text):
  device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
  inputs = tokenizer(text, return_tensors="pt").input_ids.to(device)
  outputs = model.generate(inputs, num_beams=5,
                         num_return_sequences=1, max_new_tokens=64,
                         forced_bos_token_id=tokenizer.lang_code_to_id["fa_IR"])
  return tokenizer.decode(outputs[0], skip_special_tokens=True)
  

def generate_subtitles(model, tokenizer):
  with open("source_transcription.srt", "r") as f:
    srt = f.read()
  segments = re.split("\n\n", srt)
  translated_segments = []
  for segment in tqdm(segments):
    lines = segment.split("\n")
    index, timecode, text = lines[0], lines[1], " ".join(lines[2:])
    translated_text = translate_text(model, tokenizer, text)
    translated_segment = "\n".join([index, timecode, translated_text])
    translated_segments.append(translated_segment)
  translated_srt = "\n\n".join(translated_segments)
  with open(f"target_transcription.srt", "w") as f:
    f.write(translated_srt)
  
def transcribe_audio():
  model = stable_whisper.load_model("medium")
  result = model.transcribe("audio_sample.ogg", language='en')
  srt = result.to_srt_vtt(word_level=False, segment_level=True)  
  with open(f"source_transcription.srt", "w") as f:
    f.write(srt)

def extract_audio(video):
  clip = mp.VideoFileClip(video)
  audio = clip.audio
  audio.write_audiofile("audio_sample.ogg")


def merge_srt(video, subtitle): 
  subtitles = mp.TextClip(subtitle, font="Arial", fontsize=24, color="white")
  subtitles = subtitles.set_position(("center", "bottom")).set_duration(video.duration)
  final = mp.CompositeVideoClip([video, subtitles])
  final.write_videofile("target_video.mp4")
  return final



def init(video):
  init_time = time.time()
  print('Starting Proccess')
  # TOKEN = os.environ.get('INFERCIA_TOKEN') 
  TOKEN = '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'
  model, tokenizer = load_assets(TOKEN)
  print(f"Assets Loaded, Time Taken: {time.time() - init_time}"); init_time = time.time()
  extract_audio(video)
  print(f"Audio Extacted from Video, Time Taken: {time.time() - init_time}"); init_time = time.time()
  transcribe_audio()
  print(f"Audio Transcribed using Whisper, Time Taken: {time.time() - init_time}"); init_time = time.time()
  generate_subtitles(model, tokenizer)
  print(f"Transcriptions Translated, Time Taken: {time.time() - init_time}"); init_time = time.time()
  # translated_video = merge_srt(video, 'target_transcription.srt')
  # print(f"Subtitle Added, Time Taken: {time.time() - init_time}"); init_time = time.time()
  return 'target_transcription.srt'


video_input = gr.Video(type="file", label="Video")
srt_outputs = gr.File(label="Subtitle", file_types=[".srt", ".vtt"])
interface = gr.Interface(fn=init, inputs=video_input, outputs="video", title="Subtitle Generator")
interface.launch(share=True)


# WITH BUFFERS
# import io
# def extract_audio(video):
#   clip = mp.VideoFileClip(video)
#   audio = clip.audio
#   buffer = io.BytesIO()
#   audio.write_audiofile(buffer)
#   return buffer
# def transcribe_audio(buffer, model, tokenizer):
#     model = whisper.load_model("medium")
#     buffer.seek(0) # reset the buffer position to the beginning
#     result = model.transcribe(buffer, language='en')