app.py

import gradio as gr
import numpy as np
import librosa
import soundfile as sf

# Import SV2TTS modules (from the Real-Time-Voice-Cloning repo you’ll add as dependency)
from encoder import inference as encoder
from synthesizer.inference import Synthesizer
from vocoder import inference as vocoder

# Load models (make sure weights are in /saved_models in your repo)
encoder.load_model("saved_models/encoder/saved_model.pt")
synthesizer = Synthesizer("saved_models/synthesizer/saved_model.pt")
vocoder.load_model("saved_models/vocoder/saved_model.pt")

def clone_voice(sample_wav, text, consent):
    if not consent:
        return "⚠️ You must confirm consent to use this voice.", None
    if sample_wav is None or text.strip() == "":
        return "Please upload a sample and enter text.", None

    wav, sr = librosa.load(sample_wav, sr=None)
    wav = encoder.preprocess_wav(wav, sr)
    embed = encoder.embed_utterance(wav)

    # Generate mel spectrogram
    specs = synthesizer.synthesize_spectrograms([text], [embed])
    # Vocoder to waveform
    generated_wav = vocoder.infer_waveform(specs[0])

    out_path = "out.wav"
    sf.write(out_path, generated_wav, synthesizer.sample_rate)
    return "✅ Done", out_path

with gr.Blocks() as demo:
    gr.Markdown("# 🎙️ SV2TTS Voice Cloning Demo")
    sample = gr.Audio(label="Upload speaker sample (5–10s)", source="upload", type="filepath")
    txt = gr.Textbox(label="Text to say", value="Hello, this is a test.")
    consent = gr.Checkbox(label="I confirm I have permission to clone this voice", value=False)
    status = gr.Textbox(label="Status")
    out_audio = gr.Audio(label="Generated audio")
    btn = gr.Button("Generate")
    btn.click(fn=clone_voice, inputs=[sample, txt, consent], outputs=[status, out_audio])

if __name__ == "__main__":
    demo.launch()