app.py

import tempfile
import gradio as gr
import subprocess
import os, stat
import uuid
from googletrans import Translator
from TTS.api import TTS
import ffmpeg
from faster_whisper import WhisperModel
from scipy.signal import wiener
import soundfile as sf
from pydub import AudioSegment
import numpy as np
import librosa
from zipfile import ZipFile
import shlex
import cv2
import torch
import torchvision
from tqdm import tqdm
from numba import jit

os.environ["COQUI_TOS_AGREED"] = "1"

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

#Whisper
model_size = "small"
model = WhisperModel(model_size, device="cuda", compute_type="int8")

def check_for_faces(video_path):
    face_cascade = cv2.CascadeClassifier(cv2.data.haarcascades + 'haarcascade_frontalface_default.xml')
    cap = cv2.VideoCapture(video_path)

    while True:
        ret, frame = cap.read()
        if not ret:
            break

        gray = cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
        faces = face_cascade.detectMultiScale(gray, 1.1, 4)

        if len(faces) > 0:
            return True

    return False
    
def process_video(radio, video, target_language):
    if target_language is None:
        return gr.Error("Please select a Target Language for Dubbing.")
        
    run_uuid = uuid.uuid4().hex[:6]
    output_filename = f"{run_uuid}_resized_video.mp4"
    ffmpeg.input(video).output(output_filename, vf='scale=-2:720').run()

    video_path = output_filename
    
    if not os.path.exists(video_path):
        return f"Error: {video_path} does not exist."

    # Move the duration check here
    video_info = ffmpeg.probe(video_path)
    video_duration = float(video_info['streams'][0]['duration'])

    if video_duration > 60:
        os.remove(video_path)  # Delete the resized video
        return gr.Error("Video duration exceeds 1 minute. Please upload a shorter video.")

    ffmpeg.input(video_path).output(f"{run_uuid}_output_audio.wav", acodec='pcm_s24le', ar=48000, map='a').run()

    #y, sr = sf.read(f"{run_uuid}_output_audio.wav")
    #y = y.astype(np.float32)
    #y_denoised = wiener(y)
    #sf.write(f"{run_uuid}_output_audio_denoised.wav", y_denoised, sr)

    #sound = AudioSegment.from_file(f"{run_uuid}_output_audio_denoised.wav", format="wav")
    #sound = sound.apply_gain(0)
    #sound = sound.low_pass_filter(3000).high_pass_filter(100)
    #sound.export(f"{run_uuid}_output_audio_processed.wav", format="wav")

    shell_command = f"ffmpeg -y -i {run_uuid}_output_audio.wav -af lowpass=3000,highpass=100 {run_uuid}_output_audio_final.wav".split(" ")
    subprocess.run([item for item in shell_command], capture_output=False, text=True, check=True)

    segments, info = model.transcribe(f"{run_uuid}_output_audio_final.wav", beam_size=5)
    whisper_text = " ".join(segment.text for segment in segments)
    whisper_language = info.language
    print(whisper_text)
    del model
    torch.cuda.empty_cache()
    
    language_mapping = {'English': 'en', 'Spanish': 'es', 'French': 'fr', 'German': 'de', 'Italian': 'it', 'Portuguese': 'pt', 'Polish': 'pl', 'Turkish': 'tr', 'Russian': 'ru', 'Dutch': 'nl', 'Czech': 'cs', 'Arabic': 'ar', 'Chinese (Simplified)': 'zh-cn'}
    target_language_code = language_mapping[target_language]
    translator = Translator()
    try:
        translated_text = translator.translate(whisper_text, src=whisper_language, dest=target_language_code).text
        print(translated_text)
    except AttributeError as e:
        print("Failed to translate text. Likely an issue with token extraction in the Google Translate API.")
        translated_text = "Translation failed due to API issue."

    tts = TTS("tts_models/multilingual/multi-dataset/xtts_v1")
    tts.to('cuda')
    tts.tts_to_file(translated_text, speaker_wav=f"{run_uuid}_output_audio_final.wav", file_path=f"{run_uuid}_output_synth.wav", language=target_language_code)
    del tts
    torch.cuda.empty_cache()
    
    pad_top = 0
    pad_bottom = 15
    pad_left = 0
    pad_right = 0
    rescaleFactor = 1

    video_path_fix = video_path

    has_face = check_for_faces(video_path)

    if has_face:
        cmd = f"python Wav2Lip/inference.py --checkpoint_path 'Wav2Lip/checkpoints/wav2lip_gan.pth' --face {shlex.quote(video_path)} --audio '{run_uuid}_output_synth.wav' --pads {pad_top} {pad_bottom} {pad_left} {pad_right} --resize_factor {rescaleFactor} --nosmooth --outfile '{run_uuid}_output_video.mp4'"
        subprocess.run(cmd, shell=True)
    else:
        # Merge audio with the original video without running Wav2Lip
        cmd = f"ffmpeg -i {video_path} -i {run_uuid}_output_synth.wav -c:v copy -c:a aac -strict experimental -map 0:v:0 -map 1:a:0 {run_uuid}_output_video.mp4"
        subprocess.run(cmd, shell=True)

    if not os.path.exists(f"{run_uuid}_output_video.mp4"):
        raise FileNotFoundError(f"Error: {run_uuid}_output_video.mp4 was not generated.")

    output_video_path = f"{run_uuid}_output_video.mp4"

    # Cleanup: Delete all generated files except the final output video
    files_to_delete = [
        f"{run_uuid}_resized_video.mp4",
        f"{run_uuid}_output_audio.wav",
        f"{run_uuid}_output_audio_final.wav",
        f"{run_uuid}_output_synth.wav"
    ]
    for file in files_to_delete:
        try:
            os.remove(file)
        except FileNotFoundError:
            print(f"File {file} not found for deletion.")

    return output_video_path
    
    
def swap(radio):
    if(radio == "Upload"):
        return gr.update(source="upload")
    else:
        return gr.update(source="webcam")
        
video = gr.Video()
radio = gr.Radio(["Upload", "Record"], value="Upload", show_label=False)
iface = gr.Interface(
    fn=process_video,
    inputs=[
        radio,
        video,
        gr.Dropdown(choices=["English", "Spanish", "French", "German", "Italian", "Portuguese", "Polish", "Turkish", "Russian", "Dutch", "Czech", "Arabic", "Chinese (Simplified)"], label="Target Language for Dubbing", value="Spanish")
    ],
    outputs=gr.Video(),
    live=False,
    title="AI Video Dubbing",
    description="""This tool was developed by [@artificialguybr](https://twitter.com/artificialguybr) using entirely open-source tools. Special thanks to Hugging Face for the GPU support. Thanks [@yeswondwer](https://twitter.com/@yeswondwerr) for original code.""",
    allow_flagging=False
)
with gr.Blocks() as demo:
    iface.render()
    radio.change(swap, inputs=[radio], outputs=video)
    gr.Markdown("""
    **Note:**
    - Video limit is 1 minute. It will dubbling all people using just one voice.
    - Generation may take up to 5 minutes.
    - The tool uses open-source models for all models. It's a alpha version.
    - Quality can be improved but would require more processing time per video. For scalability and hardware limitations, speed was chosen, not just quality.
    - If you need more than 1 minute, duplicate the Space and change the limit on app.py.
    """)
demo.queue(concurrency_count=1, max_size=15)
demo.launch()