app.py

import streamlit as st
import whisper
from TTS.api import TTS
from moviepy.editor import VideoFileClip, AudioFileClip, CompositeAudioClip
import os
from tempfile import NamedTemporaryFile
import torchaudio

# Page config
st.set_page_config(page_title="AI Voiceover Generator V2", layout="centered")
st.title("🎤 AI Voiceover V2: Replace One Speaker Only")

# Load models
@st.cache_resource
def load_whisper_model():
    return whisper.load_model("small")

@st.cache_resource
def load_tts_model():
    return TTS(model_name="tts_models/en/ljspeech/tacotron2-DDC", progress_bar=False, gpu=False)

whisper_model = load_whisper_model()
tts = load_tts_model()

# Upload video
video_file = st.file_uploader("Upload a short video clip (MP4 preferred)", type=["mp4", "mov", "avi"])

if video_file:
    with NamedTemporaryFile(delete=False, suffix=".mp4") as tmp_video:
        tmp_video.write(video_file.read())
        tmp_video_path = tmp_video.name

    st.video(tmp_video_path)

    # Extract audio
    video = VideoFileClip(tmp_video_path)
    audio_path = tmp_video_path.replace(".mp4", ".wav")
    video.audio.write_audiofile(audio_path)

    # Transcribe
    st.info("Transcribing using Whisper...")
    result = whisper_model.transcribe(audio_path)
    st.subheader("📝 Detected Speech")
    st.write(result["text"])

    # Custom voiceover input
    custom_text = st.text_area("Enter your custom voiceover text to replace one speaker:", result["text"])

    if st.button("Replace Only One Speaker's Voice"):
        # Generate new voiceover from custom text
        ai_voice_path = audio_path.replace(".wav", "_ai_voice.wav")
        tts.tts_to_file(text=custom_text, file_path=ai_voice_path)
        st.audio(ai_voice_path)

        # Load original audio
        original_audio, sr = torchaudio.load(audio_path)
        ai_audio, _ = torchaudio.load(ai_voice_path)

        # Trim or pad AI voice to match duration (for demo purposes)
        if ai_audio.shape[1] < original_audio.shape[1]:
            diff = original_audio.shape[1] - ai_audio.shape[1]
            ai_audio = torchaudio.functional.pad(ai_audio, (0, diff))
        else:
            ai_audio = ai_audio[:, :original_audio.shape[1]]

        # Mix original and AI audio (simulating voice replacement, basic blend)
        # NOTE: This does NOT perform speaker diarization — it's a placeholder
        mixed_audio = (original_audio * 0.4) + (ai_audio * 0.6)
        mixed_path = audio_path.replace(".wav", "_mixed.wav")
        torchaudio.save(mixed_path, mixed_audio, sr)

        # Final video
        final_video = video.set_audio(AudioFileClip(mixed_path))
        final_path = tmp_video_path.replace(".mp4", "_final_v2.mp4")
        final_video.write_videofile(final_path, codec="libx264", audio_codec="aac")

        with open(final_path, "rb") as f:
            st.download_button(label="📥 Download Final Video with Mixed Voiceover", data=f, file_name="final_ai_video_v2.mp4")