app.py

import gradio as gr
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
from subprocess import run
from faster_whisper import WhisperModel
import json
import tempfile
import os
import ffmpeg
from zipfile import ZipFile
import stat

ZipFile("ffmpeg.zip").extractall()
st = os.stat('ffmpeg')
os.chmod('ffmpeg', st.st_mode | stat.S_IEXEC)

with open('language_codes.json', 'r') as f:
    lang_codes = json.load(f)

tokenizer = AutoTokenizer.from_pretrained("facebook/nllb-200-distilled-600M")
model = AutoModelForSeq2SeqLM.from_pretrained("facebook/nllb-200-distilled-600M")
whisper_model = WhisperModel("large-v2", device="cuda", compute_type="float16")

def process_video(Video, target_language):
    print("Iniciando process_video")
    audio_file = tempfile.NamedTemporaryFile(suffix=".wav").name
    print("Executando FFmpeg para extração de áudio")
    run(["ffmpeg", "-i", Video, audio_file])
    print("Iniciando transcrição com Whisper")
    segments, _ = whisper_model.transcribe(audio_file, beam_size=5)
    segments = list(segments)
    temp_transcript_file = tempfile.NamedTemporaryFile(delete=False, suffix=".srt")
    with open(temp_transcript_file.name, "w", encoding="utf-8") as f:
        counter = 1
        for segment in segments:
            start_minutes = int(segment.start // 60)
            start_seconds = int(segment.start % 60)
            start_milliseconds = int((segment.start - int(segment.start)) * 1000)
            end_minutes = int(segment.end // 60)
            end_seconds = int(segment.end % 60)
            end_milliseconds = int((segment.end - int(segment.end)) * 1000)
            formatted_start = f"{start_minutes:02d}:{start_seconds:02d},{start_milliseconds:03d}"
            formatted_end = f"{end_minutes:02d}:{end_seconds:02d},{end_milliseconds:03d}"
            f.write(f"{counter}\n")
            f.write(f"{formatted_start} --> {formatted_end}\n")
            f.write(f"{segment.text}\n\n")
            counter += 1
    flores_code = lang_codes.get(target_language, "eng_Latn")
    temp_translated_file = tempfile.NamedTemporaryFile(delete=False, suffix=".srt")
    with open(temp_transcript_file.name, "r", encoding="utf-8") as infile, open(temp_translated_file.name, "w", encoding="utf-8") as outfile:
        for line in infile:
            if line.strip().isnumeric() or "-->" in line:
                outfile.write(line)
            elif line.strip() != "":
                inputs = tokenizer(line.strip(), return_tensors="pt")
                translated_tokens = model.generate(**inputs, forced_bos_token_id=tokenizer.lang_code_to_id[flores_code], max_length=100)
                translated_text = tokenizer.batch_decode(translated_tokens, skip_special_tokens=True)[0]
                outfile.write(translated_text + "\n")
            else:
                outfile.write("\n")
    if os.path.exists(temp_transcript_file.name):
        print(f"Arquivo de legenda criado: {temp_transcript_file.name}")
        if os.path.getsize(temp_transcript_file.name) > 0:
            print("O arquivo de legenda contém texto.")
        else:
            print("O arquivo de legenda está vazio.")
    else:
        print("Arquivo de legenda não foi criado.")
        
    output_video = "output_video.mp4"
    result = run(["ffmpeg", "-i", Video, "-vf", f"subtitles={temp_translated_file.name}", output_video])
    if result.returncode == 0:
        print("FFmpeg executado com sucesso.")
    else:
        print(f"FFmpeg falhou com o código de retorno {result.returncode}.")
    os.unlink(temp_transcript_file.name)
    os.unlink(temp_translated_file.name)
    print("process_video concluído com sucesso")
    return output_video

iface = gr.Interface(
    fn=process_video,
    inputs=[
        gr.Video(),
        gr.Dropdown(choices=list(lang_codes.keys()), label="Target Language for Dubbing", value="English"),
    ],
    outputs=gr.Video(),
    live=False,
    title="VIDEO TRANSCRIPTION AND TRANSLATION"
)

iface.launch()