app.py

import os
import torch
import gradio as gr
from openvoice import se_extractor
from openvoice.api import BaseSpeakerTTS, ToneColorConverter

# Auto-download checkpoints if not exist (from OpenVoice repo)
if not os.path.isdir("checkpoints"):
    from openvoice.utils import download_checkpoints_v2
    print("Downloading OpenVoice V2 checkpoints (~1.5GB)...")
    download_checkpoints_v2()

# ------------------- Setup -------------------
device = "cuda" if torch.cuda.is_available() else "cpu"
output_dir = 'outputs'
os.makedirs(output_dir, exist_ok=True)

# English Base Speaker (V1 style)
ckpt_base_en = 'checkpoints/base_speakers/EN'
base_speaker_tts_en = BaseSpeakerTTS(f'{ckpt_base_en}/config.json', device=device)
base_speaker_tts_en.load_ckpt(f'{ckpt_base_en}/checkpoint.pth')

# Chinese Base Speaker
ckpt_base_zh = 'checkpoints/base_speakers/ZH'
base_speaker_tts_zh = BaseSpeakerTTS(f'{ckpt_base_zh}/config.json', device=device)
base_speaker_tts_zh.load_ckpt(f'{ckpt_base_zh}/checkpoint.pth')

# Tone Color Converter (shared)
ckpt_converter = 'checkpoints/converter'
tone_color_converter = ToneColorConverter(f'{ckpt_converter}/config.json', device=device)
tone_color_converter.load_ckpt(f'{ckpt_converter}/checkpoint.pth')

# Pre-load source style embeddings
source_se_en_default = torch.load(f'{ckpt_base_en}/en_default_se.pth', map_location=device)
source_se_en_whisper = torch.load(f'{ckpt_base_en}/en_style_se.pth', map_location=device)
source_se_zh_default = torch.load(f'{ckpt_base_zh}/zh_default_se.pth', map_location=device)


# ------------------- Main Function -------------------
def voice_clone(
    reference_audio,
    text,
    language="English",
    style="default",
    speed=1.0
):
    if reference_audio is None:
        return None, "Please upload a reference voice (5-30 seconds)"

    if not text.strip():
        return None, "Please enter some text"

    # Extract target speaker embedding
    target_se, _ = se_extractor.get_se(
        reference_audio, tone_color_converter, target_dir='processed', vad=True
    )

    # Choose base TTS and source SE
    if language == "English":
        tts = base_speaker_tts_en
        if style == "whispering":
            source_se = source_se_en_whisper
            speed = 0.9
        else:
            source_se = source_se_en_default
    else:  # Chinese
        tts = base_speaker_tts_zh
        source_se = source_se_zh_default
        style = "default"

    # Generate base speech
    src_path = f"{output_dir}/tmp.wav"
    tts.tts(text, src_path, speaker=style, language=language, speed=speed)

    # Convert to cloned voice
    save_path = f"{output_dir}/output_cloned.wav"
    encode_message = "@MyShell"
    tone_color_converter.convert(
        audio_src_path=src_path,
        src_se=source_se,
        tgt_se=target_se,
        output_path=save_path,
        message=encode_message
    )

    return save_path, f"Success! Cloned in {language} ({style})"


# ------------------- Gradio Interface -------------------
with gr.Blocks(title="OpenVoice Voice Style Control Demo") as demo:
    gr.Markdown("# OpenVoice Voice Style Control Demo")
    gr.Markdown("Upload any voice → Choose language & style → Generate cloned speech instantly!")

    with gr.Row():
        ref_audio = gr.Audio(
            label="Reference Voice (5-30s, clear speech)",
            type="filepath",
            sources=["upload"]
        )

    with gr.Row():
        text_input = gr.Textbox(
            label="Text to Speak",
            value="This audio is generated by OpenVoice.",
            lines=3
        )

    with gr.Row():
        language = gr.Dropdown(
            ["English", "Chinese"], 
            value="English", 
            label="Language"
        )
        style = gr.Dropdown(
            ["default", "whispering"], 
            value="default", 
            label="Style (English only)"
        )

    generate_btn = gr.Button("Generate Cloned Voice", variant="primary")

    with gr.Row():
        output_audio = gr.Audio(label="Cloned Output")
        status = gr.Textbox(label="Status")

    generate_btn.click(
        fn=voice_clone,
        inputs=[ref_audio, text_input, language, style],
        outputs=[output_audio, status]
    )

    gr.Markdown("""
    **Tech for good**: All outputs contain @MyShell watermark.  
    Made with [OpenVoice by MyShell.ai](https://github.com/myshell-ai/OpenVoice)
    """)

demo.launch()