Upload 3 files

Dubbing_app/app.py

@@ -0,0 +1,138 @@
+import os
+import tempfile
+import gradio as gr
+from moviepy.editor import VideoFileClip, AudioFileClip
+import cv2
+import torch
+from facenet_pytorch import MTCNN
+from transformers import MarianMTModel, MarianTokenizer
+
+# Instalar os wheels (só na primeira execução)
+# pip install qwen3_asr_toolkit-1.0.4-py3-none-any.whl
+# pip install cosyvoice-2.0.0-py3-none-any.whl
+# pip install liveportrait-1.0.0-py3-none-any.whl
+
+from qwen3_asr_toolkit import transcribe_audio
+from cosyvoice.cli.cosyvoice import CosyVoice2
+from cosyvoice.utils.file_utils import load_wav
+from liveportrait import LivePortraitPipeline
+
+# === Configurações ===
+DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
+MODEL_DIR = "models"
+
+os.makedirs(MODEL_DIR, exist_ok=True)
+
+# === 1. Baixar vídeo do YouTube ===
+def download_youtube_video(url, output_path):
+    from yt_dlp import YoutubeDL
+    ydl_opts = {
+        'format': 'best[height<=720]',
+        'outtmpl': output_path,
+        'quiet': True,
+    }
+    with YoutubeDL(ydl_opts) as ydl:
+        ydl.download([url])
+    return output_path
+
+# === 2. Extrair áudio ===
+def extract_audio(video_path, audio_path):
+    clip = VideoFileClip(video_path)
+    clip.audio.write_audiofile(audio_path, fps=16000, verbose=False, logger=None)
+
+# === 3. Encontrar frame com rosto ===
+def extract_face_frame(video_path, output_image):
+    mtcnn = MTCNN(keep_all=False, device=DEVICE)
+    cap = cv2.VideoCapture(video_path)
+    while True:
+        ret, frame = cap.read()
+        if not ret:
+            break
+        boxes, _ = mtcnn.detect(frame)
+        if boxes is not None:
+            cv2.imwrite(output_image, frame)
+            cap.release()
+            return output_image
+    cap.release()
+    raise ValueError("Nenhum rosto encontrado no vídeo.")
+
+# === 4. Traduzir para pt-PT ===
+def translate_to_pt(text):
+    model_name = "Helsinki-NLP/opus-mt-en-pt"
+    tokenizer = MarianTokenizer.from_pretrained(model_name)
+    model = MarianMTModel.from_pretrained(model_name).to(DEVICE)
+    inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True).to(DEVICE)
+    translated = model.generate(**inputs)
+    result = tokenizer.decode(translated[0], skip_special_tokens=True)
+    return result
+
+# === 5. Pipeline principal ===
+def dub_youtube_video(youtube_url, target_lang="pt-PT"):
+    with tempfile.TemporaryDirectory() as tmpdir:
+        video_path = os.path.join(tmpdir, "input.mp4")
+        audio_path = os.path.join(tmpdir, "audio.wav")
+        face_image = os.path.join(tmpdir, "face.jpg")
+        dubbed_audio = os.path.join(tmpdir, "dubbed.wav")
+        output_video = os.path.join(tmpdir, "output.mp4")
+
+        # 1. Baixar vídeo
+        gr.Info("📥 A baixar vídeo do YouTube...")
+        download_youtube_video(youtube_url, video_path)
+
+        # 2. Extrair áudio
+        gr.Info("🔊 A extrair áudio...")
+        extract_audio(video_path, audio_path)
+
+        # 3. Transcrever
+        gr.Info("📝 A transcrever...")
+        text_en = transcribe_audio(audio_path)
+
+        # 4. Traduzir
+        gr.Info("🔄 A traduzir para português de Portugal...")
+        text_pt = translate_to_pt(text_en)
+
+        # 5. Clonar voz com CosyVoice
+        gr.Info("🗣️ A clonar voz e gerar áudio dublado...")
+        cosyvoice = CosyVoice2('pretrained_models/CosyVoice2-0.5B', load_jit=False, fp16=False)
+        prompt_audio = load_wav(audio_path, 16000)[:80000]  # primeiros 5s
+        for i, chunk in enumerate(cosyvoice.inference_zero_shot(
+            text=text_pt,
+            prompt_text=text_en[:100],  # trecho curto do original
+            prompt_speech_16k=prompt_audio,
+            stream=False
+        )):
+            torch.save(chunk['tts_speech'], dubbed_audio)  # simplificado
+            break
+
+        # 6. Extrair rosto
+        gr.Info("👤 A encontrar rosto...")
+        extract_face_frame(video_path, face_image)
+
+        # 7. Lip sync com LivePortrait
+        gr.Info("🎬 A gerar lip sync realista...")
+        lp = LivePortraitPipeline()
+        lp.generate(source_image=face_image, driving_audio=dubbed_audio, output_path=output_video)
+
+        return output_video
+
+# === Interface Gradio ===
+iface = gr.Interface(
+    fn=dub_youtube_video,
+    inputs=gr.Textbox(label="URL do YouTube", placeholder="https://www.youtube.com/watch?v=..."),
+    outputs=gr.Video(label="Vídeo Dublado (pt-PT)"),
+    title="🎥 Dublagem Automática com Voz Clonada + Lip Sync Realista",
+    description="""
+    Insira um link do YouTube. A app irá:
+    - Baixar o vídeo
+    - Transcrever e traduzir para **português de Portugal**
+    - Clonar a voz original
+    - Gerar novo vídeo com **lip sync realista**
+    """
+)
+
+if __name__ == "__main__":
+    # Baixar modelos na primeira execução
+    from modelscope import snapshot_download
+    if not os.path.exists("pretrained_models/CosyVoice2-0.5B"):
+        snapshot_download('iic/CosyVoice2-0.5B', local_dir='pretrained_models/CosyVoice2-0.5B')
+    iface.launch(server_name="0.0.0.0", server_port=7860)

Dubbing_app/requirements.txt

@@ -0,0 +1,14 @@
+gradio==4.40.0
+yt-dlp
+moviepy
+opencv-python
+facenet-pytorch
+torch>=2.0.0
+torchaudio>=2.0.0
+transformers>=4.36.0
+sentencepiece
+pyyaml
+tqdm
+librosa>=0.10.0
+soundfile
+modelscope

Dubbing_app/run.bat

@@ -0,0 +1,4 @@
+pip install -r requirements.txt
+pip install qwen3_asr_toolkit-1.0.4-py3-none-any.whl
+pip install cosyvoice-2.0.0-py3-none-any.whl
+pip install liveportrait-1.0.0-py3-none-any.whl