SPACE-RANDOM/RANDOM-SPACES/Blue-Archive-TTS-v2.0/app.py

import os
import glob
import json
import traceback
import logging
import gradio as gr
import numpy as np
import librosa
import torch
import asyncio
import edge_tts
import sys
import io
import wave
import re
from datetime import datetime
from fairseq import checkpoint_utils
from fairseq.data.dictionary import Dictionary
from lib.infer_pack.models import (
    SynthesizerTrnMs256NSFsid,
    SynthesizerTrnMs256NSFsid_nono,
    SynthesizerTrnMs768NSFsid,
    SynthesizerTrnMs768NSFsid_nono,
)
from vc_infer_pipeline import VC
from config import Config
import soundfile as sf
import threading
import queue

config = Config()
logging.getLogger("numba").setLevel(logging.WARNING)
logging.getLogger("fairseq").setLevel(logging.WARNING)

# OPTIMASI: Cache untuk model yang sudah dimuat
model_cache = {}
hubert_loaded = False
hubert_model = None

spaces = True 

if spaces:
    audio_mode = ["Upload audio", "TTS Audio"]
else:
    audio_mode = ["Input path", "Upload audio", "TTS Audio"]

f0method_mode = ["pm", "harvest"]
if os.path.isfile("rmvpe.pt"):
    f0method_mode.insert(2, "rmvpe")

def clean_title(title):
    title = re.sub(r'^Blue Archive\s*-\s*', '', title, flags=re.IGNORECASE)
    return re.sub(r'\s*-\s*\d+\s*epochs', '', title, flags=re.IGNORECASE)

# OPTIMASI: Audio processing yang lebih cepat
def _load_audio_input(vc_audio_mode, vc_input, vc_upload, tts_text, spaces_limit=20):
    temp_file = None
    try:
        if vc_audio_mode == "Input path" and vc_input:
            # Gunakan librosa untuk loading
            audio, sr = librosa.load(vc_input, sr=16000, mono=True)
            return audio.astype(np.float32), 16000, None
            
        elif vc_audio_mode == "Upload audio":
            if vc_upload is None:
                raise ValueError("Mohon upload file audio terlebih dahulu!")
            sampling_rate, audio = vc_upload
            
            # Konversi ke float32
            if audio.dtype != np.float32:
                audio = audio.astype(np.float32) / np.iinfo(audio.dtype).max
            
            if len(audio.shape) > 1:
                audio = np.mean(audio, axis=0)
                
            if sampling_rate != 16000:
                audio = librosa.resample(audio, orig_sr=sampling_rate, target_sr=16000, res_type='kaiser_fast')
                
            return audio.astype(np.float32), 16000, None
            
        elif vc_audio_mode == "TTS Audio":
            if not tts_text or tts_text.strip() == "":
                raise ValueError("Mohon masukkan teks untuk TTS!")
            
            temp_file = "tts_temp.wav"
            # Async TTS dengan timeout
            async def tts_task():
                return await edge_tts.Communicate(tts_text, "ja-JP-NanamiNeural").save(temp_file)
            
            # Jalankan dengan timeout
            try:
                asyncio.run(asyncio.wait_for(tts_task(), timeout=10))
            except asyncio.TimeoutError:
                raise ValueError("TTS timeout! Silakan coba lagi.")
            
            audio, sr = librosa.load(temp_file, sr=16000, mono=True)
            return audio.astype(np.float32), 16000, temp_file
            
    except Exception as e:
        if temp_file and os.path.exists(temp_file):
            os.remove(temp_file)
        raise e
        
    raise ValueError("Invalid audio mode or missing input.")

def adjust_audio_speed(audio, speed):
    if speed == 1.0:
        return audio
    # Gunakan metode yang lebih cepat untuk time stretching
    return librosa.effects.time_stretch(audio.astype(np.float32), rate=speed)

# OPTIMASI: Fungsi preprocessing audio yang lebih efisien
def preprocess_audio(audio):
    # Normalize audio
    if np.max(np.abs(audio)) > 1.0:
        audio = audio / np.max(np.abs(audio)) * 0.9
    return audio.astype(np.float32)

# OPTIMASI: Pipeline inferensi yang lebih cepat
def create_vc_fn(model_key, tgt_sr, net_g, vc, if_f0, version, file_index):
    def vc_fn(
        vc_audio_mode, vc_input, vc_upload, tts_text,
        f0_up_key, f0_method, index_rate, filter_radius, 
        resample_sr, rms_mix_rate, protect, speed,
    ):
        temp_audio_file = None
        try:
            # Clear GPU cache sebelum memulai
            if torch.cuda.is_available():
                torch.cuda.empty_cache()
                
            # Preload model ke GPU
            net_g.to(config.device)
            
            yield "Status: 🚀 Memproses audio...", None
            
            # Load audio dengan optimasi
            audio, sr, temp_audio_file = _load_audio_input(vc_audio_mode, vc_input, vc_upload, tts_text)
            
            # Preprocess audio
            audio = preprocess_audio(audio)
            
            # Konversi ke tensor dengan optimasi memory
            audio_tensor = torch.FloatTensor(audio).to(config.device)
            
            times = [0, 0, 0]
            
            # OPTIMASI: Gunakan batch processing untuk audio yang panjang
            max_chunk_size = 16000 * 30  # 30 detik per chunk
            if len(audio) > max_chunk_size:
                chunks = []
                for i in range(0, len(audio), max_chunk_size):
                    chunk = audio[i:i + max_chunk_size]
                    chunk_tensor = torch.FloatTensor(chunk).to(config.device)
                    
                    chunk_opt = vc.pipeline(
                        hubert_model, net_g, 0, chunk_tensor, 
                        "chunk" if vc_input else "temp", times, 
                        int(f0_up_key), f0_method, file_index, index_rate,
                        if_f0, filter_radius, tgt_sr, resample_sr,
                        rms_mix_rate, version, protect, f0_file=None,
                    )
                    chunks.append(chunk_opt)
                
                audio_opt = np.concatenate(chunks)
            else:
                # Processing single chunk
                audio_opt = vc.pipeline(
                    hubert_model, net_g, 0, audio_tensor, 
                    vc_input if vc_input else "temp", times, 
                    int(f0_up_key), f0_method, file_index, index_rate,
                    if_f0, filter_radius, tgt_sr, resample_sr,
                    rms_mix_rate, version, protect, f0_file=None,
                )
            
            # Pastikan audio_opt dalam format float32
            audio_opt = audio_opt.astype(np.float32)
            
            # Apply speed adjustment
            if speed != 1.0:
                audio_opt = adjust_audio_speed(audio_opt, speed)
            
            # Normalize output dan pastikan float32
            if np.max(np.abs(audio_opt)) > 0:
                audio_opt = (audio_opt / np.max(np.abs(audio_opt)) * 0.9).astype(np.float32)
            
            # Return format yang sesuai untuk gradio.Audio
            yield "Status: ✅ Selesai!", (tgt_sr, audio_opt)
            
        except Exception as e:
            yield f"❌ Error: {str(e)}\n\n{traceback.format_exc()}", None
        finally:
            # Cleanup
            if temp_audio_file and os.path.exists(temp_audio_file):
                os.remove(temp_audio_file)
            
            # Kosongkan GPU cache
            if torch.cuda.is_available():
                torch.cuda.empty_cache()
                
            # Return model ke CPU untuk hemat memory (kecuali untuk cache)
            if model_key not in model_cache:
                net_g.to('cpu')
    
    return vc_fn

def load_model():
    categories = []
    if os.path.isfile("weights/folder_info.json"):
        with open("weights/folder_info.json", "r", encoding="utf-8") as f:
            folder_info = json.load(f)
        
        for category_name, category_info in folder_info.items():
            if not category_info.get('enable', True): 
                continue
                
            category_title, category_folder, description = (
                category_info['title'], 
                category_info['folder_path'], 
                category_info['description']
            )
            
            models = []
            model_info_path = f"weights/{category_folder}/model_info.json"
            
            if os.path.exists(model_info_path):
                with open(model_info_path, "r", encoding="utf-8") as f:
                    models_info = json.load(f)
                
                for character_name, info in models_info.items():
                    if not info.get('enable', True): 
                        continue
                    
                    model_title, model_name, model_author = (
                        info['title'], 
                        info['model_path'], 
                        info.get("author")
                    )
                    
                    # Buat key unik untuk cache
                    cache_key = f"{category_folder}_{character_name}"
                    
                    # Gunakan cache jika tersedia
                    if cache_key in model_cache:
                        tgt_sr, net_g, vc, if_f0, version, model_index = model_cache[cache_key]
                    else:
                        model_cover = f"weights/{category_folder}/{character_name}/{info['cover']}"
                        model_index = f"weights/{category_folder}/{character_name}/{info['feature_retrieval_library']}"
                        
                        # Load model weights
                        model_path = f"weights/{category_folder}/{character_name}/{model_name}"
                        cpt = torch.load(model_path, map_location="cpu")
                        
                        tgt_sr = cpt["config"][-1]
                        cpt["config"][-3] = cpt["weight"]["emb_g.weight"].shape[0]
                        if_f0, version = cpt.get("f0", 1), cpt.get("version", "v1")
                        
                        # Inisialisasi model
                        if version == "v1":
                            if if_f0 == 1:
                                net_g = SynthesizerTrnMs256NSFsid(*cpt["config"], is_half=config.is_half)
                            else:
                                net_g = SynthesizerTrnMs256NSFsid_nono(*cpt["config"])
                        else:
                            if if_f0 == 1:
                                net_g = SynthesizerTrnMs768NSFsid(*cpt["config"], is_half=config.is_half)
                            else:
                                net_g = SynthesizerTrnMs768NSFsid_nono(*cpt["config"])
                        
                        # Load weights
                        if hasattr(net_g, "enc_q"): 
                            del net_g.enc_q
                        net_g.load_state_dict(cpt["weight"], strict=False)
                        net_g.eval().to('cpu')  # Simpan di CPU dulu
                        
                        # Buat VC instance
                        vc = VC(tgt_sr, config)
                        
                        # Cache model
                        model_cache[cache_key] = (tgt_sr, net_g, vc, if_f0, version, model_index)
                    
                    models.append((
                        character_name, model_title, model_author, 
                        f"weights/{category_folder}/{character_name}/{info['cover']}", 
                        version, 
                        create_vc_fn(cache_key, tgt_sr, net_g, vc, if_f0, version, model_index)
                    ))
            
            categories.append([category_title, category_folder, description, models])
    
    return categories

def load_hubert():
    global hubert_model, hubert_loaded
    if hubert_loaded:
        return
        
    torch.serialization.add_safe_globals([Dictionary])
    models, _, _ = checkpoint_utils.load_model_ensemble_and_task(
        ["hubert_base.pt"], 
        suffix="",
    )
    hubert_model = models[0].to(config.device)
    hubert_model = hubert_model.half() if config.is_half else hubert_model.float()
    hubert_model.eval()
    hubert_loaded = True

def change_audio_mode(vc_audio_mode):
    is_input_path = vc_audio_mode == "Input path"
    is_upload = vc_audio_mode == "Upload audio"
    is_tts = vc_audio_mode == "TTS Audio"
    
    return (
        gr.Textbox.update(visible=is_input_path),
        gr.Checkbox.update(visible=is_upload),
        gr.Audio.update(visible=is_upload),
        gr.Textbox.update(visible=is_tts, lines=4 if is_tts else 2)
    )

def use_microphone(microphone):
    return gr.Audio.update(source="microphone" if microphone else "upload")

# CSS tetap sama
css = """
@import url('https://fonts.googleapis.com/css2?family=Inter:wght@300;400;500;600;700&family=Quicksand:wght@400;600;700&display=swap');
body, .gradio-container { background-color: #ffffff !important; font-family: 'Inter', sans-serif !important; }
footer { display: none !important; }
.arona-loading-container { display: flex; align-items: center; justify-content: center; gap: 15px; margin-top: 15px; padding: 10px; }
.loading-text-blue { font-family: 'Quicksand', sans-serif; font-size: 20px; font-weight: 700; color: #00b0ff; letter-spacing: 1px; }
.loading-gif-small { width: 100px; height: auto; border-radius: 8px; }
.header-img-container { text-align: center; padding: 10px 0; background: #ffffff !important; }
.header-img { width: 100%; max-width: 500px; border-radius: 15px; margin: 0 auto; display: block; }
.status-card { background: #ffffff; border: 1px solid #e1f0ff; border-radius: 14px; padding: 15px 10px; margin: 0 auto 15px auto; max-width: 400px; display: flex; flex-direction: column; align-items: center; }
.status-online-box { display: flex; align-items: center; gap: 8px; margin-bottom: 12px; }
.status-details-container { display: flex; width: 100%; justify-content: center; align-items: center; border-top: 1px solid #f0f7ff; padding-top: 10px; }
.status-detail-item { flex: 1; display: flex; flex-direction: column; align-items: center; text-align: center; }
.status-detail-item:first-child { border-right: 1px solid #e1f0ff; }
.status-text-main { font-size: 13px !important; font-weight: 600; color: #546e7a; }
.status-text-sub { font-size: 11px !important; color: #90a4ae; }
.dot-online { height: 8px; width: 8px; background-color: #2ecc71; border-radius: 50%; display: inline-block; animation: blink-green 1.5s infinite; }
@keyframes blink-green { 0% { opacity: 1; } 50% { opacity: 0.4; } 100% { opacity: 1; } }
.gr-form .gr-block label span, .gr-box label span, .gr-panel label span { background: linear-gradient(135deg, #4fc3f7 0%, #00b0ff 100%) !important; color: white !important; padding: 4px 12px !important; border-radius: 8px !important; font-weight: 600 !important; box-shadow: 0 0 15px rgba(79, 195, 247, 0.4) !important; }
input[type="range"] { accent-color: #00b0ff !important; }
.char-scroll-box { display: grid !important; grid-template-columns: repeat(2, 1fr) !important; gap: 12px !important; max-height: 280px; overflow-y: auto; padding: 15px; background: #ffffff; border: 2px solid #eef5ff; border-radius: 14px; }
.char-card { background: white; padding: 12px; border-radius: 12px; cursor: pointer; border: 1px solid #e1f5fe; border-left: 5px solid #4fc3f7; transition: all 0.2s ease; display: flex; flex-direction: column; height: 65px; }
.char-name-jp { font-weight: 700; font-size: 11px !important; color: #455a64; }
.char-name-en { font-size: 8.5px !important; color: #90a4ae; text-transform: uppercase; }
.speed-section { margin-top: 20px; padding: 18px; border-radius: 20px; background: linear-gradient(135deg, #f0f7ff 0%, #ffffff 100%); border: 2px solid #e1f0ff; }
.speed-title { font-family: 'Quicksand', sans-serif; font-weight: 700; color: #4ea8de; text-align: center; margin-bottom: 12px; font-size: 14px; }
.generate-btn { font-family: 'Quicksand', sans-serif; font-weight: 700 !important; background: linear-gradient(135deg, #64b5f6 0%, #2196f3 100%) !important; color: white !important; border-radius: 12px !important; }
.footer-text { text-align: center; padding: 20px; border-top: 1px solid #f0f4f8; color: #b0bec5; font-size: 11px; }
.speed-notes-box { font-family: 'Arial'; border: 1px solid #ffd8b2; border-radius: 8px; padding: 12px; background: #fff7ed; border-left: 4px solid #fb923c; margin-top: 10px; }
.speed-notes-title { color: #c2410c; font-size: 12px; margin: 0 0 5px 0; font-weight: bold; }
.speed-notes-content { color: #9a3412; font-size: 11px; margin: 0; }
"""

if __name__ == '__main__':
    # Preload hubert model
    load_hubert()
    
    # Load models dengan cache
    categories = load_model()
    total_models = sum(len(models) for _, _, _, models in categories)
    
    # Optimasi Gradio dengan queue yang lebih efisien
    with gr.Blocks(css=css, theme=gr.themes.Soft()) as app:
        gr.HTML('<div class="header-img-container"><img src="https://huggingface.co/spaces/Rosmontis-Chan/Blue-Archive-TTS-v2.0/resolve/main/Blue-Archive-TTS-v2.0.PNG" class="header-img"></div>')
        
        gr.HTML(f'''
            <div class="status-card">
                <div class="status-online-box"><span class="dot-online"></span><b style="color: #4fc3f7; font-size: 14px;">System Online</b></div>
                <div class="status-details-container">
                    <div class="status-detail-item"><span class="status-text-main">👥 {total_models} Students</span><span class="status-text-sub">Ready</span></div>
                    <div class="status-detail-item"><span class="status-text-main">📊 Total</span><span class="status-text-sub">Database: {total_models}</span></div>
                </div>
            </div>
        ''')
        
        for cat_idx, (folder_title, folder, description, models) in enumerate(categories):
            with gr.TabItem(folder_title):
                with gr.Accordion("📑 Select Student", open=True):
                    char_html = "".join([f'<div class="char-card" onclick="selectModel(\'{folder_title}\', \'{name}\')"><span class="char-name-jp">{clean_title(title)}</span><span class="char-name-en">{name}</span></div>' for name, title, author, cover, version, vc_fn in models])
                    gr.HTML(f'<div class="char-scroll-box">{char_html}</div>')
                
                with gr.Tabs():
                    for model_idx, (name, title, author, cover, model_version, vc_fn) in enumerate(models):
                        with gr.TabItem(name, id=f"model_{cat_idx}_{model_idx}"):
                            with gr.Row():
                                with gr.Column(scale=1):
                                    gr.HTML(f'<div style="display:flex;flex-direction:column;align-items:center;padding:20px;background:white;border-radius:20px;border:1px solid #eef5ff;"><img style="width:200px;height:260px;object-fit:cover;border-radius:15px;" src="file/{cover}"><div style="font-family:\'Quicksand\',sans-serif;font-weight:700;font-size:18px;color:#039be5;margin-top:15px;">{clean_title(title)}</div><div style="font-size:11px;color:#b0bec5;margin-top:5px;">{model_version} • {author}</div></div>')
                                
                                with gr.Column(scale=2):
                                    with gr.Group():
                                        vc_audio_mode = gr.Dropdown(label="Input Mode", choices=audio_mode, value="TTS Audio")
                                        vc_input = gr.Textbox(visible=False)
                                        vc_microphone_mode = gr.Checkbox(label="Use Microphone", value=False)
                                        vc_upload = gr.Audio(label="Upload Audio Source", source="upload", visible=False)
                                        tts_text = gr.Textbox(label="TTS Text", visible=True, placeholder="Type message here...", lines=3)
                                    
                                    with gr.Row():
                                        with gr.Column():
                                            vc_transform0 = gr.Slider(minimum=-12, maximum=12, label="Pitch (Nada)", value=12, step=1)
                                            f0method0 = gr.Radio(label="Conversion Algorithm", choices=f0method_mode, value="rmvpe")
                                        with gr.Column():
                                            with gr.Accordion("⚙️ Advanced Tuning", open=True):
                                                index_rate1 = gr.Slider(0, 1, label="Index Rate", value=0.75)
                                                filter_radius0 = gr.Slider(0, 7, label="Filter", value=7, step=1)
                                                resample_sr0 = gr.Slider(0, 48000, label="Resample", value=0)
                                                rms_mix_rate0 = gr.Slider(0, 1, label="Volume Mix", value=0.76)
                                                protect0 = gr.Slider(0, 0.5, label="Voice Protect", value=0.33)
                                    
                                    # BOX NOTES & SARAN - TAMPILAN LENGKAP
                                    with gr.Row():
                                        with gr.Column():
                                            gr.HTML("""<div style="font-family: 'Arial'; border: 1px solid #bae6fd; border-radius: 8px; padding: 12px; background: #f0f9ff; border-left: 4px solid #0ea5e9; margin-bottom: 8px;">
                                                <h4 style="color: #0369a1; font-size: 13px; margin: 0 0 5px 0;">📝 Notes & Panduan Fitur</h4>
                                                <p style="color: #075985; font-size: 11px; margin: 0 0 3px 0;"><b>Pitch:</b> Mengatur nada suara (naik/turun)</p>
                                                <p style="color: #075985; font-size: 11px; margin: 0 0 3px 0;"><b>Algoritma:</b> Metode ekstraksi nada (RMVPE paling akurat)</p>
                                                <p style="color: #075985; font-size: 11px; margin: 0 0 3px 0;"><b>Retrieval:</b> Kemiripan karakter suara (0-1)</p>
                                                <p style="color: #075985; font-size: 11px; margin: 0 0 3px 0;"><b>Filter:</b> Smoothing untuk mengurangi noise</p>
                                                <p style="color: #075985; font-size: 11px; margin: 0 0 3px 0;"><b>Volume:</b> Stabilitas volume output</p>
                                                <p style="color: #075985; font-size: 11px; margin: 0;"><b>Protect:</b> Proteksi suara agar tetap natural</p>
                                            </div>""")
                                        with gr.Column():
                                            gr.HTML("""<div style="font-family: 'Arial'; border: 1px solid #dcfce7; border-radius: 8px; padding: 12px; background: #f0fdf4; border-left: 4px solid #22c55e;">
                                                <h4 style="color: #166534; font-size: 13px; margin: 0 0 5px 0;">📑 DI SARANKAN 📑</h4>
                                                <p style="color: #166534; font-size: 11px; margin: 0 0 3px 0;"><b>Pitch:</b> <span style="color: #15803d; font-weight: bold;">+12</span> (Optimal untuk karakter perempuan)</p>
                                                <p style="color: #166534; font-size: 11px; margin: 0 0 3px 0;"><b>Algoritma:</b> <span style="color: #15803d; font-weight: bold;">RMVPE</span> (Akurasi tinggi)</p>
                                                <p style="color: #166534; font-size: 11px; margin: 0 0 3px 0;"><b>Retrieval:</b> <span style="color: #15803d; font-weight: bold;">0.75</span> (Keseimbangan)</p>
                                                <p style="color: #166534; font-size: 11px; margin: 0 0 3px 0;"><b>Filter:</b> <span style="color: #15803d; font-weight: bold;">7</span> (Noise reduction optimal)</p>
                                                <p style="color: #166534; font-size: 11px; margin: 0 0 3px 0;"><b>Volume:</b> <span style="color: #15803d; font-weight: bold;">0.76</span> (Stabil)</p>
                                                <p style="color: #166534; font-size: 11px; margin: 0;"><b>Protect:</b> <span style="color: #15803d; font-weight: bold;">0.33</span> (Natural)</p>
                                            </div>""")

                                    with gr.Column(elem_classes="speed-section"):
                                        gr.HTML('<div class="speed-title">⚡ KECEPATAN SUARA ⚡</div>')
                                        speed_slider = gr.Slider(0.5, 2.0, value=1.0, step=0.1, label=None)
                                        
                                        # NOTES KHUSUS UNTUK SLIDER KECEPATAN - DIPERBAIKI
                                        gr.HTML("""<div class="speed-notes-box">
                                            <div class="speed-notes-title">ℹ️ Petunjuk Penggunaan Kecepatan</div>
                                            <div class="speed-notes-content">
                                                • <b>Kiri (0.5):</b> Memperlambat suara karakter hingga 50%<br>
                                                • <b>Tengah (1.0):</b> Kecepatan normal (disarankan)<br>
                                                • <b>Kanan (2.0):</b> Mempercepat suara karakter hingga 200%<br><br>
                                                <b>Tips:</b> Atur ke kiri untuk suara lebih lambat dan dramatis, atur ke kanan untuk suara lebih cepat dan bersemangat. Disarankan tetap di 1.0 untuk hasil natural.
                                            </div>
                                        </div>""")
                                        
                                        gr.HTML('<div class="arona-loading-container"><div class="loading-text-blue">Ready!</div><img class="loading-gif-small" src="https://huggingface.co/spaces/Rosmontis-Chan/Blue-Archive-TTS-v2.0/resolve/main/Arona.gif"></div>')
                                
                                with gr.Column(scale=1):
                                    vc_log = gr.Textbox(label="Process Logs", interactive=False)
                                    vc_output = gr.Audio(label="Result Audio", interactive=False)
                                    vc_convert = gr.Button("🎐 GENERATE VOICE 🎐", variant="primary", elem_classes="generate-btn")

                            vc_convert.click(
                                fn=vc_fn, 
                                inputs=[vc_audio_mode, vc_input, vc_upload, tts_text, vc_transform0, f0method0, 
                                       index_rate1, filter_radius0, resample_sr0, rms_mix_rate0, protect0, speed_slider], 
                                outputs=[vc_log, vc_output]
                            )
                            vc_audio_mode.change(fn=change_audio_mode, inputs=[vc_audio_mode], outputs=[vc_input, vc_microphone_mode, vc_upload, tts_text])
                            vc_microphone_mode.change(fn=use_microphone, inputs=vc_microphone_mode, outputs=vc_upload)

        gr.HTML('<div class="footer-text"><div>DESIGNED BY PLANA-CHAN</div><div style="font-weight:700; color:#90a4ae;">Blue Archive Voice Conversion v2.0 • 2024</div></div>')
        app.load(None, None, None, js="""() => { window.selectModel = (cat, mod) => { const tabs = document.querySelectorAll('.tabs .tab-nav button'); for (let t of tabs) { if (t.textContent.trim() === cat) { t.click(); setTimeout(() => { const mTabs = document.querySelectorAll('.tabs .tab-nav button'); for (let mt of mTabs) { if (mt.textContent.trim() === mod) mt.click(); } }, 50); break; } } } }""")

    # DIPERBAIKI: Sesuaikan dengan Gradio 3.50.2
    app.queue(
        max_size=3  # Kurangi queue size untuk respons lebih cepat
    ).launch(
        share=False,
        server_name="0.0.0.0" if os.getenv('SPACE_ID') else "127.0.0.1",
        server_port=7860,
        quiet=True,  # Kurangi logging
        show_error=True
    )