Update app.py

app.py

@@ -1,433 +1,495 @@
-#!/usr/bin/env python3
-# -*- coding: utf-8 -*-
-
-"""
-Vietnamese TTS - Hugging Face Spaces Demo
-Giao diện web chuyển văn bản tiếng Việt thành giọng nói.
-"""
-
-import os
-import sys
-import json
-import tempfile
-import glob
-import re
-from pathlib import Path
-
-import torch
-import numpy as np
-import soundfile as sf
-import gradio as gr
-
-# Add src to path for imports
-sys.path.insert(0, str(Path(__file__).parent))
-
-from src.vietnamese.text_processor import process_vietnamese_text
-from src.vietnamese.phonemizer import text_to_phonemes, VIPHONEME_AVAILABLE
-from src.models.synthesizer import SynthesizerTrn
-from src.text.symbols import symbols
-from src.utils import helpers as utils
-
-
-# ============== Viphoneme Check ==============
-
-def check_viphoneme():
-    """Check if viphoneme is working properly."""
-    from src.vietnamese.phonemizer import VIPHONEME_AVAILABLE, text_to_phonemes
-    
-    print(f"Viphoneme available: {VIPHONEME_AVAILABLE}")
-    
-    if VIPHONEME_AVAILABLE:
-        try:
-            # Test with a simple Vietnamese text
-            test_text = "Xin chào"
-            phones, tones, word2ph = text_to_phonemes(test_text, use_viphoneme=True)
-            print(f"✅ Viphoneme test passed: '{test_text}' -> {phones[:5]}...")
-            return True
-        except Exception as e:
-            print(f"❌ Viphoneme test failed: {e}")
-            return False
-    else:
-        print("⚠️ Viphoneme not available, using fallback phonemizer")
-        return False
-
-
-# ============== Model Loading ==============
-
-def find_latest_checkpoint(model_dir, prefix="G"):
-    """Find the latest checkpoint in model directory."""
-    pattern = os.path.join(model_dir, f"{prefix}*.pth")
-    checkpoints = glob.glob(pattern)
-    if not checkpoints:
-        return None
-    
-    def get_step(path):
-        match = re.search(rf'{prefix}(\d+)\.pth', path)
-        return int(match.group(1)) if match else 0
-    
-    checkpoints.sort(key=get_step, reverse=True)
-    return checkpoints[0]
-
-
-def download_model():
-    """Download model from Hugging Face Hub."""
-    from huggingface_hub import snapshot_download
-    
-    hf_repo = "valtecAI-team/valtec-tts-pretrained"
-    
-    # Get cache directory
-    if os.name == 'nt':  # Windows
-        cache_base = Path(os.environ.get('LOCALAPPDATA', Path.home() / 'AppData' / 'Local'))
-    else:  # Linux/Mac
-        cache_base = Path(os.environ.get('XDG_CACHE_HOME', Path.home() / '.cache'))
-    
-    model_dir = cache_base / 'valtec_tts' / 'models' / 'vits-vietnamese'
-    
-    # Check if already downloaded
-    config_path = model_dir / "config.json"
-    if config_path.exists():
-        checkpoints = list(model_dir.glob("G_*.pth"))
-        if checkpoints:
-            print(f"Using cached model from: {model_dir}")
-            return str(model_dir)
-    
-    print(f"Downloading model from {hf_repo}...")
-    model_dir.mkdir(parents=True, exist_ok=True)
-    snapshot_download(repo_id=hf_repo, local_dir=str(model_dir))
-    print("Download complete!")
-    
-    return str(model_dir)
-
-
-class VietnameseTTS:
-    """Vietnamese TTS synthesizer using trained VITS-based model."""
-    
-    def __init__(self, checkpoint_path, config_path, device="cpu"):
-        self.device = device
-        
-        # Load config
-        with open(config_path, 'r', encoding='utf-8') as f:
-            self.config = json.load(f)
-        
-        self.sampling_rate = self.config['data']['sampling_rate']
-        self.spk2id = self.config['data']['spk2id']
-        self.speakers = list(self.spk2id.keys())
-        self.add_blank = self.config['data'].get('add_blank', True)
-        
-        print(f"Available speakers: {self.speakers}")
-        
-        # Load model
-        self._load_model(checkpoint_path)
-    
-    def _load_model(self, checkpoint_path):
-        """Load the trained model."""
-        self.model = SynthesizerTrn(
-            len(symbols),
-            self.config['data']['filter_length'] // 2 + 1,
-            self.config['train']['segment_size'] // self.config['data']['hop_length'],
-            n_speakers=self.config['data']['n_speakers'],
-            **self.config['model'],
-        ).to(self.device)
-        
-        # Load checkpoint
-        checkpoint = torch.load(checkpoint_path, map_location=self.device)
-        
-        # Handle DDP checkpoint
-        state_dict = checkpoint['model']
-        new_state_dict = {}
-        for k, v in state_dict.items():
-            if k.startswith('module.'):
-                new_state_dict[k[7:]] = v
-            else:
-                new_state_dict[k] = v
-        
-        self.model.load_state_dict(new_state_dict, strict=False)
-        self.model.eval()
-        
-        print(f"Model loaded from {checkpoint_path}")
-    
-    def text_to_sequence(self, text, speaker):
-        """Convert text to model input tensors."""
-        from src.text import cleaned_text_to_sequence
-        from src.nn import commons
-        
-        # Normalize text
-        normalized_text = process_vietnamese_text(text)
-        
-        # Convert to phonemes
-        phones, tones, word2ph = text_to_phonemes(normalized_text, use_viphoneme=VIPHONEME_AVAILABLE)
-        
-        # Convert to sequence
-        phone_ids, tone_ids, lang_ids = cleaned_text_to_sequence(phones, tones, "VI")
-        
-        # Add blanks if needed
-        if self.add_blank:
-            phone_ids = commons.intersperse(phone_ids, 0)
-            tone_ids = commons.intersperse(tone_ids, 0)
-            lang_ids = commons.intersperse(lang_ids, 0)
-        
-        # Get speaker ID
-        if speaker not in self.spk2id:
-            print(f"Warning: Speaker '{speaker}' not found, using first speaker: {self.speakers[0]}")
-            speaker = self.speakers[0]
-        speaker_id = self.spk2id[speaker]
-        
-        # Create tensors
-        x = torch.LongTensor(phone_ids).unsqueeze(0).to(self.device)
-        x_lengths = torch.LongTensor([len(phone_ids)]).to(self.device)
-        tone = torch.LongTensor(tone_ids).unsqueeze(0).to(self.device)
-        language = torch.LongTensor(lang_ids).unsqueeze(0).to(self.device)
-        sid = torch.LongTensor([speaker_id]).to(self.device)
-        
-        # Create dummy BERT features
-        bert = torch.zeros(1024, len(phone_ids)).unsqueeze(0).to(self.device)
-        ja_bert = torch.zeros(768, len(phone_ids)).unsqueeze(0).to(self.device)
-        
-        return x, x_lengths, tone, language, sid, bert, ja_bert
-    
-    @torch.no_grad()
-    def synthesize(self, text, speaker, sdp_ratio=0.0, noise_scale=0.667, 
-                   noise_scale_w=0.8, length_scale=1.0):
-        """Synthesize speech from text."""
-        x, x_lengths, tone, language, sid, bert, ja_bert = self.text_to_sequence(text, speaker)
-        
-        audio, attn, *_ = self.model.infer(
-            x, x_lengths, sid, tone, language, bert, ja_bert,
-            sdp_ratio=sdp_ratio,
-            noise_scale=noise_scale,
-            noise_scale_w=noise_scale_w,
-            length_scale=length_scale,
-        )
-        
-        audio = audio[0, 0].cpu().numpy()
-        return audio, self.sampling_rate
-
-
-# ============== Gradio Interface ==============
-
-class TTSInterface:
-    """Wrapper for TTS model with Gradio interface."""
-    
-    def __init__(self):
-        print("Initializing TTS...")
-        
-        # Detect device
-        self.device = "cuda" if torch.cuda.is_available() else "cpu"
-        print(f"Using device: {self.device}")
-        
-        # Download/load model
-        model_dir = download_model()
-        checkpoint_path = find_latest_checkpoint(model_dir, "G")
-        config_path = os.path.join(model_dir, "config.json")
-        
-        self.tts = VietnameseTTS(checkpoint_path, config_path, self.device)
-        self.temp_dir = Path(tempfile.gettempdir()) / "valtec_tts_demo"
-        self.temp_dir.mkdir(exist_ok=True)
-        
-        print("TTS initialized successfully!")
-    
-    def synthesize(self, text, speaker, speed, noise_scale, noise_scale_w, sdp_ratio):
-        """Synthesize speech from text."""
-        try:
-            if not text or not text.strip():
-                return None, "⚠️ Vui lòng nhập văn bản"
-            
-            audio, sr = self.tts.synthesize(
-                text=text.strip(),
-                speaker=speaker,
-                length_scale=speed,
-                noise_scale=noise_scale,
-                noise_scale_w=noise_scale_w,
-                sdp_ratio=sdp_ratio,
-            )
-            
-            # Save to temp file
-            output_path = self.temp_dir / f"output_{hash(text)}.wav"
-            sf.write(str(output_path), audio, sr)
-            
-            return str(output_path), f"✅ Tạo giọng nói thành công! ({len(audio)/sr:.2f}s)"
-            
-        except Exception as e:
-            return None, f"❌ Lỗi: {str(e)}"
-
-
-def create_demo(tts_interface):
-    """Create Gradio interface."""
-    
-    examples = [
-        ["Xin chào, chúc bạn một ngày tốt lành", "male", 1.0, 0.667, 0.8, 0.0],
-        ["Buổi sáng hôm nay trời trong xanh và gió thổi rất nhẹ", "male", 1.0, 0.667, 0.8, 0.0],
-        ["Tôi pha một tách cà phê nóng và ngồi nhìn ánh nắng chiếu qua cửa sổ", "female", 1.0, 0.667, 0.8, 0.0],
-        ["Việt Nam là một đất nước xinh đẹp với văn hóa phong phú", "male", 0.9, 0.667, 0.8, 0.0],
-        ["Con đường làng quê rợp bóng tre xanh mát rượi", "female", 1.1, 0.667, 0.8, 0.0],
-    ]
-    
-    with gr.Blocks(
-        theme=gr.themes.Soft(primary_hue="blue", secondary_hue="cyan"),
-        title="Chuyển Văn Bản Thành Giọng Nói Tiếng Việt",
-        css="""
-        /* Base styles - Mobile first */
-        .gradio-container { 
-            max-width: 100% !important; 
-            margin: 0 auto !important;
-            padding: 10px !important;
-        }
-        .main {
-            margin: 0 auto !important;
-            padding: 0 10px !important;
-        }
-        #title {
-            text-align: center;
-            background: linear-gradient(90deg, #667eea 0%, #764ba2 100%);
-            -webkit-background-clip: text;
-            -webkit-text-fill-color: transparent;
-            font-weight: bold;
-            font-size: 1.5rem;
-        }
-        .contain {
-            max-width: 100% !important;
-            margin: 0 auto !important;
-        }
-        
-        /* Tablet (768px+) */
-        @media (min-width: 768px) {
-            .gradio-container { 
-                max-width: 750px !important; 
-                padding: 20px !important;
-            }
-            #title {
-                font-size: 2rem;
-            }
-        }
-        
-        /* Desktop (1024px+) */
-        @media (min-width: 1024px) {
-            .gradio-container { 
-                max-width: 900px !important; 
-            }
-            #title {
-                font-size: 2.5rem;
-            }
-        }
-        
-        /* Large screens (1200px+) */
-        @media (min-width: 1200px) {
-            .gradio-container { 
-                max-width: 1000px !important; 
-            }
-        }
-        """
-    ) as demo:
-        
-        gr.Markdown("""
-            # <span id="title">🎙️ Chuyển Văn Bản Thành Giọng Nói</span>
-            
-            ### Hệ thống tổng hợp giọng nói tiếng Việt
-            
-            Nhập văn bản tiếng Việt và chọn giọng đọc để tạo audio.
-        """)
-        
-        with gr.Row():
-            with gr.Column(scale=2):
-                text_input = gr.Textbox(
-                    label="📝 Văn bản đầu vào",
-                    placeholder="Nhập văn bản tiếng Việt ở đây...",
-                    lines=5,
-                    max_lines=10,
-                )
-                
-                speaker_dropdown = gr.Dropdown(
-                    choices=tts_interface.tts.speakers,
-                    value="female",
-                    label="🎤 Chọn giọng đọc",
-                    info="Chọn người đọc từ danh sách"
-                )
-                
-                synthesize_btn = gr.Button(
-                    "🔊 Tạo giọng nói",
-                    variant="primary",
-                    size="lg"
-                )
-            
-            with gr.Column(scale=1):
-                with gr.Accordion("⚙️ Cài đặt nâng cao", open=False):
-                    speed_slider = gr.Slider(
-                        minimum=0.5, maximum=2.0, value=1.0, step=0.1,
-                        label="Tốc độ",
-                        info="< 1.0: Nhanh hơn | > 1.0: Chậm hơn"
-                    )
-                    
-                    noise_scale_slider = gr.Slider(
-                        minimum=0.1, maximum=1.5, value=0.667, step=0.01,
-                        label="Noise Scale",
-                        info="Điều khiển độ biến thiên giọng nói"
-                    )
-                    
-                    noise_scale_w_slider = gr.Slider(
-                        minimum=0.1, maximum=1.5, value=0.8, step=0.01,
-                        label="Duration Noise",
-                        info="Điều khiển độ biến thiên thời lượng"
-                    )
-                    
-                    sdp_ratio_slider = gr.Slider(
-                        minimum=0.0, maximum=1.0, value=0.0, step=0.1,
-                        label="SDP Ratio",
-                        info="0: Xác định | 1: Ngẫu nhiên"
-                    )
-        
-        with gr.Row():
-            with gr.Column():
-                audio_output = gr.Audio(
-                    label="🔊 Audio đầu ra",
-                    type="filepath",
-                    interactive=False
-                )
-                status_output = gr.Textbox(
-                    label="📊 Trạng thái",
-                    interactive=False,
-                    show_label=False
-                )
-        
-        gr.Markdown("### 📚 Ví dụ")
-        gr.Examples(
-            examples=examples,
-            inputs=[text_input, speaker_dropdown, speed_slider, 
-                    noise_scale_slider, noise_scale_w_slider, sdp_ratio_slider],
-            outputs=[audio_output, status_output],
-            fn=tts_interface.synthesize,
-            cache_examples=False,
-        )
-        
-        synthesize_btn.click(
-            fn=tts_interface.synthesize,
-            inputs=[text_input, speaker_dropdown, speed_slider,
-                    noise_scale_slider, noise_scale_w_slider, sdp_ratio_slider],
-            outputs=[audio_output, status_output],
-        )
-        
-        gr.Markdown("""
-            ---
-            <div style="text-align: center; color: #666; font-size: 0.9em;">
-                Hệ thống tổng hợp giọng nói tiếng Việt | Powered by <b>Valtec AI Team</b>
-            </div>
-        """)
-    
-    return demo
-
-
-# ============== Main ==============
-
-if __name__ == "__main__":
-    print("Đang khởi động hệ thống tổng hợp giọng nói tiếng Việt...")
-    
-    # Check viphoneme
-    check_viphoneme()
-    
-    tts_interface = TTSInterface()
-    demo = create_demo(tts_interface)
-    
-    demo.launch(
-        server_name="0.0.0.0",
-        server_port=7860,
-        show_error=True,
-    )
+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+"""
+CVNSS4.0 Vietnamese TTS Studio
+- Architecture: Modular CSS & Component Separation
+- UX: High Contrast Input Fields 
+- Core: Optimized Logic Flow
+"""
+
+import os
+import sys
+import json
+import time
+import glob
+import re
+import hashlib
+import tempfile
+from pathlib import Path
+
+import torch
+import numpy as np
+import soundfile as sf
+import gradio as gr
+
+# Add src to path for imports
+sys.path.insert(0, str(Path(__file__).parent))
+
+# Import core modules
+try:
+    from src.vietnamese.text_processor import process_vietnamese_text
+    from src.vietnamese.phonemizer import text_to_phonemes, VIPHONEME_AVAILABLE
+    from src.models.synthesizer import SynthesizerTrn
+    from src.text.symbols import symbols
+except ImportError:
+    # Fallback for environment setup if src is missing during init
+    print("⚠️ Core modules not found. Ensure 'src' directory exists.")
+    VIPHONEME_AVAILABLE = False
+    symbols = []
+
+
+# =========================================================
+# 1) SYSTEM CONFIGURATION & CSS (The Expert Layer)
+# =========================================================
+
+# Expert CSS: Definitive Z-Index Management & Neon Theme
+NEON_CSS = r"""
+:root {
+    --bg-dark: #0f172a;
+    --bg-panel: rgba(30, 41, 59, 0.7);
+    --line: rgba(148, 163, 184, 0.1);
+    --text-primary: #e2e8f0;
+    --neon-cyan: #06b6d4;
+    --neon-accent: #38bdf8;
+    --radius-lg: 16px;
+    --radius-sm: 8px;
+    
+    /* UX Color Palette for Inputs */
+    --input-bg: #f1f5f9;       /* Light Blue-Grey for readability */
+    --input-text: #0f4c81;     /* Classic Blue (Dark Blue) for high contrast */
+    --input-placeholder: #64748b;
+}
+
+body, .gradio-container, .app {
+    background: radial-gradient(circle at 50% 0%, #1e293b 0%, #0f172a 100%) !important;
+    color: var(--text-primary) !important;
+    font-family: 'Inter', 'Segoe UI', sans-serif;
+}
+
+/* --- ISOLATION FULL: CVNSS4.0 Vietnamese TTS Studio --- */
+.panelNeon {
+    border: 1px solid rgba(255,255,255,0.08);
+    border-radius: var(--radius-lg);
+    background: var(--bg-panel);
+    backdrop-filter: blur(12px);
+    box-shadow: 0 4px 6px -1px rgba(0, 0, 0, 0.1), 0 2px 4px -1px rgba(0, 0, 0, 0.06);
+    padding: 20px;
+    position: relative;
+    isolation: isolate; 
+    z-index: 1; 
+    margin-bottom: 20px;
+}
+
+/* UX IMPROVEMENT: High Contrast Input Styling */
+.panelNeon textarea, .panelNeon input[type="text"] {
+    background: var(--input-bg) !important;
+    color: var(--input-text) !important; /* DARK BLUE TEXT requested */
+    border: 2px solid transparent !important;
+    border-radius: var(--radius-sm) !important;
+    font-weight: 500 !important;
+    font-size: 1rem !important;
+    line-height: 1.5 !important;
+    padding: 12px !important;
+    transition: all 0.2s ease;
+    z-index: 10 !important;
+    position: relative !important;
+}
+
+.panelNeon textarea::placeholder {
+    color: var(--input-placeholder) !important;
+}
+
+.panelNeon textarea:focus, .panelNeon input:focus {
+    background: #ffffff !important;
+    border-color: var(--neon-cyan) !important;
+    box-shadow: 0 0 0 4px rgba(6, 182, 212, 0.15) !important;
+    color: #000000 !important; /* Even darker on focus */
+}
+
+/* Label Styling */
+.panelNeon label span {
+    color: var(--neon-accent) !important;
+    font-weight: 600;
+    font-size: 0.85rem;
+    text-transform: uppercase;
+    letter-spacing: 0.05em;
+    margin-bottom: 8px;
+    display: block;
+}
+
+/* Dropdown & Slider fixes */
+.panelNeon .wrap, .panelNeon .range-compact {
+    z-index: 10 !important;
+}
+
+/* Button Upgrades */
+button.primary, .gr-button-primary {
+    background: linear-gradient(135deg, #06b6d4 0%, #3b82f6 100%) !important;
+    border: none !important;
+    color: white !important;
+    font-weight: 700 !important;
+    transition: transform 0.1s ease, box-shadow 0.2s ease;
+}
+button.primary:hover, .gr-button-primary:hover {
+    box-shadow: 0 10px 15px -3px rgba(6, 182, 212, 0.3) !important;
+    transform: translateY(-1px);
+}
+button.primary:active {
+    transform: translateY(0px);
+}
+
+/* Status Panel */
+.statusCard {
+    background: rgba(15, 23, 42, 0.6);
+    border-radius: var(--radius-sm);
+    padding: 16px;
+    border: 1px solid rgba(255,255,255,0.05);
+}
+.pill {
+    display: inline-flex;
+    align-items: center;
+    padding: 4px 12px;
+    border-radius: 99px;
+    background: rgba(56, 189, 248, 0.1);
+    color: #38bdf8;
+    border: 1px solid rgba(56, 189, 248, 0.2);
+    font-size: 0.8rem;
+    font-weight: 600;
+    margin-right: 6px;
+    margin-bottom: 6px;
+}
+.alert { padding: 12px; border-radius: 8px; margin-top: 12px; font-size: 0.9rem; font-weight: 500; display: flex; align-items: center; gap: 8px;}
+.alertOk { background: rgba(34, 197, 94, 0.1); color: #4ade80; border: 1px solid rgba(34, 197, 94, 0.2); }
+.alertWarn { background: rgba(234, 179, 8, 0.1); color: #facc15; border: 1px solid rgba(234, 179, 8, 0.2); }
+"""
+
+# =========================================================
+# 2) UTILITIES & HELPERS
+# =========================================================
+def check_viphoneme():
+    if not VIPHONEME_AVAILABLE:
+        print("⚠️ Viphoneme not available.")
+        return False
+    try:
+        phones, _, _ = text_to_phonemes("Test", use_viphoneme=True)
+        print("✅ Viphoneme active.")
+        return True
+    except Exception as e:
+        print(f"❌ Viphoneme error: {e}")
+        return False
+
+def md5_key(*parts: str) -> str:
+    return hashlib.md5("|".join(parts).encode("utf-8")).hexdigest()
+
+def split_sentences_vi(text: str, max_chars: int):
+    # Improved splitting logic
+    if not text: return []
+    text = re.sub(r'\s+', ' ', text).strip()
+    # Split by delimiters keeping delimiters
+    parts = re.split(r'([.?!;:])', text)
+    
+    chunks = []
+    current_chunk = ""
+    
+    for i in range(0, len(parts) - 1, 2):
+        sentence = parts[i] + parts[i+1]
+        if len(current_chunk) + len(sentence) <= max_chars:
+            current_chunk += sentence
+        else:
+            if current_chunk: chunks.append(current_chunk.strip())
+            current_chunk = sentence
+            
+    if len(parts) % 2 != 0 and parts[-1]:
+        sentence = parts[-1]
+        if len(current_chunk) + len(sentence) <= max_chars:
+            current_chunk += sentence
+        else:
+            if current_chunk: chunks.append(current_chunk.strip())
+            current_chunk = sentence
+            
+    if current_chunk: chunks.append(current_chunk.strip())
+    return chunks
+
+# =========================================================
+# 3) CORE ENGINE WRAPPER
+# =========================================================
+class TTSManager:
+    """Singleton-like manager for TTS operations."""
+    def __init__(self):
+        self.device = "cuda" if torch.cuda.is_available() else "cpu"
+        print(f"🔧 Initializing TTS on {self.device}...")
+        
+        self.model_dir = self._get_model_dir()
+        self.ckpt_path = find_latest_checkpoint(self.model_dir, "G")
+        self.cfg_path = os.path.join(self.model_dir, "config.json")
+        
+        if not self.ckpt_path:
+            raise FileNotFoundError(f"No checkpoint found in {self.model_dir}")
+
+        self.tts = VietnameseTTS(self.ckpt_path, self.cfg_path, self.device)
+        self.temp_dir = Path(tempfile.gettempdir()) / "neon_tts_cache"
+        self.temp_dir.mkdir(parents=True, exist_ok=True)
+
+    def _get_model_dir(self):
+        return download_model()
+
+    def synthesize(self, text, speaker, speed, noise_scale, noise_scale_w, sdp_ratio):
+        try:
+            if not text or not text.strip():
+                return None, "⚠️ Empty input"
+                
+            key = md5_key(speaker, f"{speed:.2f}", text[:20], str(len(text)))
+            out_path = self.temp_dir / f"{key}.wav"
+            
+            if out_path.exists():
+                return str(out_path), "✅ Cached (From history)"
+                
+            audio, sr = self.tts.synthesize(
+                text=text, speaker=speaker, length_scale=speed,
+                noise_scale=noise_scale, noise_scale_w=noise_scale_w, sdp_ratio=sdp_ratio
+            )
+            sf.write(str(out_path), audio, sr)
+            return str(out_path), "✅ Generated successfully"
+        except Exception as e:
+            # Capture full traceback if needed, but return clean msg
+            return None, f"❌ Error: {str(e)}"
+
+# =========================================================
+# 4) MODEL LOGIC (PRESERVED & FIXED)
+# =========================================================
+def find_latest_checkpoint(model_dir, prefix="G"):
+    pattern = os.path.join(model_dir, f"{prefix}*.pth")
+    checkpoints = glob.glob(pattern)
+    if not checkpoints: return None
+    checkpoints.sort(key=lambda x: int(re.search(rf"{prefix}(\d+)\.pth", x).group(1)) if re.search(rf"{prefix}(\d+)\.pth", x) else 0, reverse=True)
+    return checkpoints[0]
+
+def download_model():
+    from huggingface_hub import snapshot_download
+    hf_repo = "valtecAI-team/valtec-tts-pretrained"
+    cache_base = Path(os.environ.get("XDG_CACHE_HOME", Path.home() / ".cache"))
+    if os.name == "nt": cache_base = Path(os.environ.get("LOCALAPPDATA", Path.home() / "AppData" / "Local"))
+    
+    model_dir = cache_base / "valtec_tts" / "models" / "vits-vietnamese"
+    if (model_dir / "config.json").exists() and list(model_dir.glob("G_*.pth")):
+        return str(model_dir)
+        
+    print(f"⬇️ Downloading {hf_repo}...")
+    snapshot_download(repo_id=hf_repo, local_dir=str(model_dir))
+    return str(model_dir)
+
+class VietnameseTTS:
+    def __init__(self, ckpt, cfg, device="cpu"):
+        self.device = device
+        with open(cfg, "r", encoding="utf-8") as f: self.config = json.load(f)
+        self.spk2id = self.config["data"]["spk2id"]
+        self.speakers = list(self.spk2id.keys())
+        self._load(ckpt)
+        
+    def _load(self, ckpt):
+        self.model = SynthesizerTrn(
+            len(symbols),
+            self.config["data"]["filter_length"] // 2 + 1,
+            self.config["train"]["segment_size"] // self.config["data"]["hop_length"],
+            n_speakers=self.config["data"]["n_speakers"],
+            **self.config["model"]
+        ).to(self.device)
+        state = torch.load(ckpt, map_location=self.device)["model"]
+        self.model.load_state_dict({k.replace("module.", ""): v for k,v in state.items()}, strict=False)
+        self.model.eval()
+
+    def synthesize(self, text, speaker, **kwargs):
+        from src.text import cleaned_text_to_sequence
+        from src.nn import commons
+        
+        # 1. Text Processing
+        norm_text = process_vietnamese_text(text)
+        phones, tones, _ = text_to_phonemes(norm_text, use_viphoneme=VIPHONEME_AVAILABLE)
+        phone_ids, tone_ids, lang_ids = cleaned_text_to_sequence(phones, tones, "VI")
+        
+        phone_ids = commons.intersperse(phone_ids, 0)
+        tone_ids = commons.intersperse(tone_ids, 0)
+        lang_ids = commons.intersperse(lang_ids, 0)
+        
+        # 2. Prepare Tensors
+        x = torch.LongTensor(phone_ids).unsqueeze(0).to(self.device)
+        x_len = torch.LongTensor([len(phone_ids)]).to(self.device)
+        tone = torch.LongTensor(tone_ids).unsqueeze(0).to(self.device)
+        lang = torch.LongTensor(lang_ids).unsqueeze(0).to(self.device)
+        sid = torch.LongTensor([self.spk2id.get(speaker, 0)]).to(self.device)
+        
+        # 3. Inference with Gradient Safety (FIX IS HERE)
+        with torch.no_grad():
+            bert = torch.zeros(1024, len(phone_ids)).unsqueeze(0).to(self.device)
+            ja_bert = torch.zeros(768, len(phone_ids)).unsqueeze(0).to(self.device)
+            
+            # Run inference
+            # The error "Can't call numpy() on Tensor that requires grad" means output has grad_fn.
+            # We use .detach() before .cpu() to ensure the graph is cut.
+            outputs = self.model.infer(
+                x, x_len, sid, tone, lang, 
+                bert, ja_bert, 
+                **kwargs
+            )
+            
+            audio = outputs[0][0,0].detach().cpu().numpy()
+            
+        return audio, self.config["data"]["sampling_rate"]
+
+# =========================================================
+# 5) UI CONSTRUCTION (REFACTORED)
+# =========================================================
+def create_ui(manager: TTSManager):
+    
+    def ui_header():
+        return gr.HTML("""
+        <div style="border-bottom: 1px solid rgba(255,255,255,0.08); padding-bottom: 20px; margin-bottom: 25px;">
+            <h1 style="color: #38bdf8; margin:0; font-weight:800; font-size: 2rem; letter-spacing: -0.02em;">
+                🎛️ CVNSS4.0 Vietnamese TTS Studio
+            </h1>
+            <div style="color: #94a3b8; font-size: 1rem; margin-top: 5px; font-weight: 400;">
+                Công nghệ tích hợp giọng nói AI tiên tiến • Phiên bản 1.0.0 Demo • Dự án mã nguồn mở
+            </div>
+        </div>
+        """)
+
+    def ui_status_render(text, speaker, speed, chunks, dur, msg):
+        return f"""
+        <div class="statusCard">
+            <div style="margin-bottom:12px; font-weight:700; color:#38bdf8; font-size: 0.9rem; text-transform: uppercase;">
+                📟 Trạng thái hoạt động
+            </div>
+            <div style="display:flex; flex-wrap:wrap; gap:8px;">
+                <span class="pill">🎤 {speaker}</span>
+                <span class="pill">⚡ {speed}x</span>
+                <span class="pill">📄 {len(text)} ký tự</span>
+                <span class="pill">🧩 {chunks} đoạn</span>
+            </div>
+            <div class="alert {'alertOk' if '✅' in msg else 'alertWarn'}">
+                {msg}
+            </div>
+        </div>
+        """
+
+    with gr.Blocks(theme=gr.themes.Base(), css=NEON_CSS, title="Neon TTS Expert") as app:
+        ui_header()
+        
+        with gr.Tabs():
+            # --- TAB BASIC ---
+            with gr.Tab("⚡ Chế độ Nhanh"):
+                with gr.Row():
+                    # INPUT COLUMN
+                    with gr.Column(scale=2):
+                        # REFACTOR: Using a specific ID for the container to target with CSS isolation
+                        with gr.Group(elem_classes=["panelNeon"], elem_id="input-panel-basic"):
+                            gr.HTML('<div class="panelTitle">📝 Văn bản đầu vào</div>')
+                            
+                            # THE FIX: Pure Textbox with updated styling (Dark Blue text)
+                            txt_basic = gr.Textbox(
+                                label="",
+                                show_label=False,
+                                placeholder="Nhập nội dung tiếng Việt vào... (Ví dụ: Xin chào, bạn đã học qua CVNSS4.0 chưa?)",
+                                lines=6,
+                                elem_id="main-input-basic"
+                            )
+                            
+                            with gr.Row():
+                                spk_basic = gr.Dropdown(choices=manager.tts.speakers, value=manager.tts.speakers[0], label="Giọng đọc")
+                                spd_basic = gr.Slider(0.5, 2.0, value=1.0, step=0.1, label="Tốc độ đọc")
+                            
+                            btn_basic = gr.Button("🔊 Đọc ngay", variant="primary")
+
+                    # OUTPUT COLUMN
+                    with gr.Column(scale=1):
+                        with gr.Group(elem_classes=["panelNeon"]):
+                            gr.HTML('<div class="panelTitle">🔊 Kết quả</div>')
+                            out_audio_basic = gr.Audio(label="Audio Output", type="filepath", interactive=False)
+                            out_status_basic = gr.HTML()
+
+            # --- TAB ADVANCED ---
+            with gr.Tab("🧠 Chế độ Chuyên sâu"):
+                with gr.Row():
+                    with gr.Column(scale=2):
+                        with gr.Group(elem_classes=["panelNeon"], elem_id="input-panel-adv"):
+                            gr.HTML('<div class="panelTitle">📝 Xử lý văn bản dài</div>')
+                            txt_adv = gr.Textbox(
+                                label="",
+                                show_label=False,
+                                lines=8, 
+                                placeholder="Nhập văn bản dài cần ngắt câu tự động...", 
+                                elem_id="main-input-adv"
+                            )
+                            
+                            gr.HTML('<div style="height:15px"></div>')
+                            gr.HTML('<div class="panelTitle">🎚️ Tham số âm thanh</div>')
+                            
+                            with gr.Row():
+                                ns = gr.Slider(0.1, 1.5, 0.667, step=0.01, label="Noise Scale (Độ biến thiên)")
+                                nsw = gr.Slider(0.1, 1.5, 0.8, step=0.01, label="Duration Noise (Độ động)")
+                            
+                            with gr.Row():
+                                sdp = gr.Slider(0.0, 1.0, 0.2, step=0.1, label="SDP Ratio (Ngẫu nhiên)")
+                                max_chars = gr.Slider(50, 500, 300, step=10, label="Ngắt đoạn (ký tự)")
+                                pause = gr.Slider(0, 1000, 250, step=50, label="Nghỉ câu (ms)")
+
+                            btn_adv = gr.Button("🧠 Xử lý & Ghép nối", variant="primary")
+
+                    with gr.Column(scale=1):
+                        with gr.Group(elem_classes=["panelNeon"]):
+                            out_audio_adv = gr.Audio(label="Merged Audio", type="filepath", interactive=False)
+                            out_status_adv = gr.HTML()
+
+        # --- LOGIC BINDING ---
+        def run_basic(text, spk, spd):
+            path, msg = manager.synthesize(text, spk, spd, 0.667, 0.8, 0.2)
+            html = ui_status_render(text, spk, spd, 1, 0, msg)
+            return path, html
+
+        def run_adv(text, spk, spd, ns, nsw, sdp, mc, p, progress=gr.Progress()):
+            chunks = split_sentences_vi(text, int(mc))
+            audios = []
+            sr = 44100
+            
+            for i, chunk in enumerate(chunks):
+                progress((i)/len(chunks), desc=f"Đang xử lý đoạn {i+1}/{len(chunks)}")
+                path, _ = manager.synthesize(chunk, spk, spd, ns, nsw, sdp)
+                if path:
+                    data, rate = sf.read(path)
+                    audios.append(data)
+                    sr = rate
+                    if p > 0:
+                        audios.append(np.zeros(int(rate * p/1000)))
+            
+            if not audios: return None, "❌ Không tạo được âm thanh"
+            
+            full_audio = np.concatenate(audios)
+            out_path = manager.temp_dir / f"merged_{int(time.time())}.wav"
+            sf.write(str(out_path), full_audio, sr)
+            
+            html = ui_status_render(text, spk, spd, len(chunks), len(full_audio)/sr, "✅ Đã ghép nối thành công")
+            return str(out_path), html
+
+        btn_basic.click(run_basic, [txt_basic, spk_basic, spd_basic], [out_audio_basic, out_status_basic])
+        btn_adv.click(run_adv, [txt_adv, spk_basic, spd_basic, ns, nsw, sdp, max_chars, pause], [out_audio_adv, out_status_adv])
+
+    return app
+
+# =========================================================
+# 6) ENTRY POINT
+# =========================================================
+if __name__ == "__main__":
+    print("🚀 Starting Expert Neon TTS...")
+    
+    # Check dependencies
+    check_viphoneme()
+    
+    # Init Manager & UI
+    try:
+        tts_manager = TTSManager()
+        app = create_ui(tts_manager)
+        
+        port = int(os.environ.get("PORT", "7860"))
+        app.queue(max_size=10).launch(server_name="0.0.0.0", server_port=port)
+    except Exception as e:
+        print(f"❌ Fatal Error: {e}")