Update app.py

app.py

@@ -1,491 +1,558 @@
-# Copyright (c) 2024 Alibaba Inc (authors: Xiang Lyu, Liu Yue)
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#   http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-import spaces
-import os
-import sys
-import argparse
 import gradio as gr
-import numpy as np
+import sys
+import os
 import torch
 import torchaudio
-import random
+import torchaudio.transforms as T
+import numpy as np
+import tempfile
 import librosa
-from funasr import AutoModel
-from funasr.utils.postprocess_utils import rich_transcription_postprocess
-ROOT_DIR = os.path.dirname(os.path.abspath(__file__))
-sys.path.append('{}/third_party/Matcha-TTS'.format(ROOT_DIR))
-
-from modelscope import snapshot_download, HubApi
-from huggingface_hub import snapshot_download as hf_snapshot_download
-
-hf_snapshot_download('FunAudioLLM/Fun-CosyVoice3-0.5B-2512', local_dir='pretrained_models/Fun-CosyVoice3-0.5B')
-snapshot_download('iic/SenseVoiceSmall', local_dir='pretrained_models/SenseVoiceSmall')
-hf_snapshot_download('FunAudioLLM/CosyVoice-ttsfrd', local_dir='pretrained_models/CosyVoice-ttsfrd')
-os.system(
-    "cd pretrained_models/CosyVoice-ttsfrd/ && "
-    "pip install ttsfrd_dependency-0.1-py3-none-any.whl && "
-    "pip install ttsfrd-0.4.2-cp310-cp310-linux_x86_64.whl && "
-    "apt install -y unzip && "
-    "rm -rf resource && "
-    "unzip resource.zip -d ."
-)
-
-from cosyvoice.cli.cosyvoice import AutoModel as CosyVoiceAutoModel
-from cosyvoice.utils.file_utils import logging, load_wav
-from cosyvoice.utils.common import set_all_random_seed, instruct_list
-
-# -----------------------------
-# i18n (En: British spelling)
-# -----------------------------
-LANG_EN = "En"
-LANG_ZH = "Zh"
-
-MODE_ZERO_SHOT = "zero_shot"
-MODE_INSTRUCT = "instruct"
-
-UI_TEXT = {
-    LANG_EN: {
-        "lang_label": "Language",
-        "md_links": (
-            "### Repository [CosyVoice](https://github.com/FunAudioLLM/CosyVoice)  \n"
-            "Pre-trained model [Fun-CosyVoice3-0.5B](https://huggingface.co/FunAudioLLM/Fun-CosyVoice3-0.5B-2512)  \n"
-            "[CosyVoice2-0.5B](https://www.modelscope.cn/models/iic/CosyVoice2-0.5B)  \n"
-            "[CosyVoice-300M](https://www.modelscope.cn/models/iic/CosyVoice-300M)  \n"
-            "[CosyVoice-300M-Instruct](https://www.modelscope.cn/models/iic/CosyVoice-300M-Instruct)  \n"
-            "[CosyVoice-300M-SFT](https://www.modelscope.cn/models/iic/CosyVoice-300M-SFT)"
-        ),
-        "md_hint": "#### Enter the text to synthesise, choose an inference mode, and follow the steps.",
-        "tts_label": "Text to synthesise",
-        "tts_default": "Her handwriting is very neat, which suggests she likes things tidy.",
-        "mode_label": "Inference mode",
-        "mode_zero_shot": "3s fast voice cloning",
-        "mode_instruct": "Natural language control",
-        "steps_label": "Steps",
-        "steps_zero_shot": (
-            "1. Choose a prompt audio file, or record prompt audio (≤ 30s). If both are provided, the uploaded file is used.\n"
-            "2. Enter the prompt text.\n"
-            "3. Click Generate audio."
-        ),
-        "steps_instruct": (
-            "1. Choose a prompt audio file, or record prompt audio (≤ 30s). If both are provided, the uploaded file is used.\n"
-            "2. Choose/enter the instruct text.\n"
-            "3. Click Generate audio."
-        ),
-        "stream_label": "Streaming inference",
-        "stream_no": "No",
-        "dice": "🎲",
-        "seed_label": "Random inference seed",
-        "upload_label": "Choose prompt audio file (sample rate ≥ 16 kHz)",
-        "record_label": "Record prompt audio",
-        "prompt_text_label": "Prompt text",
-        "prompt_text_ph": "Enter prompt text (auto recognition supported; you can edit the result)...",
-        "instruct_label": "Choose instruct text",
-        "generate_btn": "Generate audio",
-        "output_label": "Synthesised audio",
-        "warn_too_long": "Your input text is too long; please keep it within 200 characters.",
-        "warn_instruct_empty": "You are using Natural language control; please enter instruct text.",
-        "info_instruct_need_prompt": "You are using Natural language control; please provide prompt audio.",
-        "warn_prompt_missing": "Prompt audio is empty. Did you forget to provide prompt audio?",
-        "warn_prompt_sr_low": "Prompt audio sample rate {} is below {}.",
-        "warn_prompt_too_long_10s": "Please keep the prompt audio within 10 seconds to avoid poor inference quality.",
-        "warn_prompt_text_missing": "Prompt text is empty. Did you forget to enter prompt text?",
-        "info_instruct_ignored": "You are using 3s fast voice cloning; instruct text will be ignored.",
-        "warn_invalid_mode": "Invalid mode selection.",
-    },
-    LANG_ZH: {
-        "lang_label": "语言",
-        "md_links": (
-            "### 代码库 [CosyVoice](https://github.com/FunAudioLLM/CosyVoice)  \n"
-            "预训练模型 [Fun-CosyVoice3-0.5B](https://huggingface.co/FunAudioLLM/Fun-CosyVoice3-0.5B-2512)  \n"
-            "[CosyVoice2-0.5B](https://www.modelscope.cn/models/iic/CosyVoice2-0.5B)  \n"
-            "[CosyVoice-300M](https://www.modelscope.cn/models/iic/CosyVoice-300M)  \n"
-            "[CosyVoice-300M-Instruct](https://www.modelscope.cn/models/iic/CosyVoice-300M-Instruct)  \n"
-            "[CosyVoice-300M-SFT](https://www.modelscope.cn/models/iic/CosyVoice-300M-SFT)"
-        ),
-        "md_hint": "#### 请输入需要合成的文本，选择推理模式，并按照提示步骤进行操作",
-        "tts_label": "输入合成文本",
-        "tts_default": "Her handwriting is [M][AY0][N][UW1][T]并且很整洁，说明她[h][ào]干净。",
-        "mode_label": "选择推理模式",
-        "mode_zero_shot": "3s极速复刻",
-        "mode_instruct": "自然语言控制",
-        "steps_label": "操作步骤",
-        "steps_zero_shot": (
-            "1. 选择prompt音频文件，或录入prompt音频，注意不超过30s，若同时提供，优先选择prompt音频文件\n"
-            "2. 输入prompt文本\n"
-            "3. 点击生成音频按钮"
-        ),
-        "steps_instruct": (
-            "1. 选择prompt音频文件，或录入prompt音频，注意不超过30s，若同时提供，优先选择prompt音频文件\n"
-            "2. 输入instruct文本\n"
-            "3. 点击生成音频按钮"
-        ),
-        "stream_label": "是否流式推理",
-        "stream_no": "否",
-        "dice": "🎲",
-        "seed_label": "随机推理种子",
-        "upload_label": "选择prompt音频文件，注意采样率不低于16khz",
-        "record_label": "录制prompt音频文件",
-        "prompt_text_label": "prompt文本",
-        "prompt_text_ph": "请输入prompt文本，支持自动识别，您可以自行修正识别结果...",
-        "instruct_label": "选择instruct文本",
-        "generate_btn": "生成音频",
-        "output_label": "合成音频",
-        "warn_too_long": "您输入的文字过长，请限制在200字以内",
-        "warn_instruct_empty": "您正在使用自然语言控制模式, 请输入instruct文本",
-        "info_instruct_need_prompt": "您正在使用自然语言控制模式, 请输入prompt音频",
-        "warn_prompt_missing": "prompt音频为空，您是否忘记输入prompt音频？",
-        "warn_prompt_sr_low": "prompt音频采样率{}低于{}",
-        "warn_prompt_too_long_10s": "请限制输入音频在10s内，避免推理效果过低",
-        "warn_prompt_text_missing": "prompt文本为空，您是否忘记输入prompt文本？",
-        "info_instruct_ignored": "您正在使用3s极速复刻模式，instruct文本会被忽略！",
-        "warn_invalid_mode": "无效的模式选择",
-    },
-}
-
-
-def t(lang: str, key: str) -> str:
-    lang = lang if lang in UI_TEXT else LANG_ZH
-    return UI_TEXT[lang][key]
-
-
-def mode_choices(lang: str):
-    return [
-        (t(lang, "mode_zero_shot"), MODE_ZERO_SHOT),
-        (t(lang, "mode_instruct"), MODE_INSTRUCT),
-    ]
-
-
-def steps_for(lang: str, mode_value: str) -> str:
-    if mode_value == MODE_INSTRUCT:
-        return t(lang, "steps_instruct")
-    return t(lang, "steps_zero_shot")
-
-
-# -----------------------------
-# Audio post-process
-# -----------------------------
+from pathlib import Path
+
+print("=" * 60)
+print("🎙️ Fun-CosyVoice3 TTS Initialization")
+print("=" * 60)
+
+# Step 1: Setup directories
+print("\n📁 Step 1: Setting up directories...")
+WORK_DIR = Path.cwd()
+COSYVOICE_DIR = WORK_DIR / "CosyVoice"
+MODEL_DIR = COSYVOICE_DIR / "pretrained_models" / "Fun-CosyVoice3-0.5B"
+
+print(f"Working directory: {WORK_DIR}")
+print(f"CosyVoice directory: {COSYVOICE_DIR}")
+print(f"Model directory: {MODEL_DIR}")
+
+# Step 2: Clone CosyVoice if needed
+if not COSYVOICE_DIR.exists():
+    print("\n📥 Step 2: Cloning CosyVoice repository...")
+    import subprocess
+    try:
+        subprocess.run([
+            'git', 'clone', '--recursive',
+            'https://github.com/FunAudioLLM/CosyVoice.git',
+            str(COSYVOICE_DIR)
+        ], check=True)
+        print("✅ Repository cloned successfully")
+    except Exception as e:
+        print(f"❌ Failed to clone repository: {e}")
+        raise
+else:
+    print("\n✅ Step 2: CosyVoice repository already exists")
+
+# Step 3: Download models
+if not MODEL_DIR.exists():
+    print("\n📥 Step 3: Downloading models (this may take a few minutes)...")
+    from huggingface_hub import snapshot_download
+    try:
+        print("Downloading Fun-CosyVoice3-0.5B-2512...")
+        snapshot_download(
+            'FunAudioLLM/Fun-CosyVoice3-0.5B-2512',
+            local_dir=str(MODEL_DIR),
+            local_dir_use_symlinks=False
+        )
+        print("✅ Model downloaded successfully")
+    except Exception as e:
+        print(f"❌ Failed to download model: {e}")
+        raise
+else:
+    print("\n✅ Step 3: Models already exist")
+
+# Step 4: Download ttsfrd (optional)
+TTSFRD_DIR = COSYVOICE_DIR / "pretrained_models" / "CosyVoice-ttsfrd"
+if not TTSFRD_DIR.exists():
+    print("\n📥 Step 4: Downloading ttsfrd...")
+    from huggingface_hub import snapshot_download
+    try:
+        snapshot_download(
+            'FunAudioLLM/CosyVoice-ttsfrd',
+            local_dir=str(TTSFRD_DIR),
+            local_dir_use_symlinks=False
+        )
+        print("✅ ttsfrd downloaded successfully")
+    except Exception as e:
+        print(f"⚠️ Failed to download ttsfrd (will use WeText): {e}")
+else:
+    print("\n✅ Step 4: ttsfrd already exists")
+
+# Step 5: Add to Python path
+print("\n🔧 Step 5: Configuring Python path...")
+sys.path.insert(0, str(COSYVOICE_DIR))
+sys.path.insert(0, str(COSYVOICE_DIR / "third_party" / "Matcha-TTS"))
+print(f"Added to path: {COSYVOICE_DIR}")
+print(f"Added to path: {COSYVOICE_DIR / 'third_party' / 'Matcha-TTS'}")
+
+# Step 6: Import CosyVoice
+print("\n📦 Step 6: Importing CosyVoice...")
+try:
+    from cosyvoice.cli.cosyvoice import AutoModel as CosyVoiceAutoModel
+    from cosyvoice.utils.file_utils import load_wav
+    from cosyvoice.utils.common import set_all_random_seed
+    print("✅ CosyVoice imported successfully")
+except Exception as e:
+    print(f"❌ Failed to import CosyVoice: {e}")
+    raise
+
+print("\n" + "=" * 60)
+print("✅ Initialization completed successfully!")
+print("=" * 60 + "\n")
+
+# Global variables
+cosyvoice = None
+target_sr = 24000
+prompt_sr = 16000
 max_val = 0.8
 top_db = 60
 hop_length = 220
 win_length = 440
 
-
-def generate_seed():
-    seed = random.randint(1, 100000000)
-    return {"__type__": "update", "value": seed}
-
-
-def postprocess(wav):
-    speech = load_wav(wav, target_sr=target_sr, min_sr=16000)
-    speech, _ = librosa.effects.trim(
-        speech, top_db=top_db, frame_length=win_length, hop_length=hop_length
-    )
-    if speech.abs().max() > max_val:
-        speech = speech / speech.abs().max() * max_val
-    speech = torch.concat([speech, torch.zeros(1, int(target_sr * 0.2))], dim=1)
-    torchaudio.save(wav, speech, target_sr)
-    return wav
-
-@spaces.GPU
-def prompt_wav_recognition(prompt_wav):
-    res = asr_model.generate(
-        input=prompt_wav,
-        language="auto",  # "zn", "en", "yue", "ja", "ko", "nospeech"
-        use_itn=True,
-    )
-    text = res[0]["text"].split("|>")[-1]
-    return text
-
-
-@spaces.GPU
-def generate_audio(
-    tts_text,
-    mode_value,
-    prompt_text,
-    prompt_wav_upload,
-    prompt_wav_record,
-    instruct_text,
-    seed,
-    stream,
-    ui_lang,
-):
-    stream = False
-
-    if len(tts_text) > 200:
-        gr.Warning(t(ui_lang, "warn_too_long"))
-        return (target_sr, default_data)
-
-    sft_dropdown, speed = "", 1.0
-
-    if prompt_wav_upload is not None:
-        prompt_wav = prompt_wav_upload
-    elif prompt_wav_record is not None:
-        prompt_wav = prompt_wav_record
-    else:
-        prompt_wav = None
-
-    # instruct mode requirements
-    if mode_value == MODE_INSTRUCT:
-        if instruct_text == "":
-            gr.Warning(t(ui_lang, "warn_instruct_empty"))
-            return (target_sr, default_data)
-        if prompt_wav is None:
-            gr.Info(t(ui_lang, "info_instruct_need_prompt"))
-            return (target_sr, default_data)
-
-    # zero-shot requirements
-    if mode_value == MODE_ZERO_SHOT:
-        if prompt_wav is None:
-            gr.Warning(t(ui_lang, "warn_prompt_missing"))
-            return (target_sr, default_data)
-
-        info = torchaudio.info(prompt_wav)
+def load_model():
+    """Load the CosyVoice model"""
+    global cosyvoice
+    if cosyvoice is None:
+        print("🚀 Loading CosyVoice model...")
+        try:
+            cosyvoice = CosyVoiceAutoModel(
+                model_dir=str(MODEL_DIR),
+                load_trt=False,
+                fp16=False
+            )
+            print("✅ Model loaded successfully!")
+        except Exception as e:
+            print(f"❌ Error loading model: {e}")
+            import traceback
+            traceback.print_exc()
+            raise gr.Error(f"Failed to load model: {e}")
+    return cosyvoice
+
+def postprocess(wav_path):
+    """Post-process audio - trim silence and normalize (from official code)"""
+    try:
+        speech = load_wav(wav_path, target_sr=target_sr, min_sr=16000)
+        
+        # Trim silence from beginning and end
+        speech, _ = librosa.effects.trim(
+            speech, top_db=top_db, 
+            frame_length=win_length, 
+            hop_length=hop_length
+        )
+        
+        # Normalize if too loud
+        if speech.abs().max() > max_val:
+            speech = speech / speech.abs().max() * max_val
+        
+        # Add silence at the end
+        speech = torch.concat([speech, torch.zeros(1, int(target_sr * 0.2))], dim=1)
+        
+        # Save back
+        torchaudio.save(wav_path, speech, target_sr)
+        return wav_path
+    except Exception as e:
+        print(f"⚠️ Postprocess warning: {e}")
+        return wav_path
+
+def process_audio(audio_input):
+    """
+    Convert audio input to proper format for CosyVoice
+    Handles: stereo->mono, different dtypes, resampling
+    """
+    if audio_input is None:
+        return None
+    
+    try:
+        sr, audio_data = audio_input
+        
+        print(f"📊 Input audio - shape: {audio_data.shape}, dtype: {audio_data.dtype}, sr: {sr}Hz")
+        
+        # Step 1: Normalize data type to float32
+        if audio_data.dtype == np.int16:
+            audio_data = audio_data.astype(np.float32) / 32768.0
+        elif audio_data.dtype == np.int32:
+            audio_data = audio_data.astype(np.float32) / 2147483648.0
+        elif audio_data.dtype == np.float64:
+            audio_data = audio_data.astype(np.float32)
+        elif audio_data.dtype != np.float32:
+            audio_data = audio_data.astype(np.float32)
+        
+        # Step 2: Convert stereo to mono if needed
+        if len(audio_data.shape) == 2:
+            print(f"   Converting stereo ({audio_data.shape[1]} channels) to mono...")
+            if audio_data.shape[1] == 2:
+                audio_data = audio_data.mean(axis=1)
+            elif audio_data.shape[1] == 1:
+                audio_data = audio_data.squeeze()
+            else:
+                audio_data = audio_data[:, 0]
+        
+        # Step 3: Ensure 1D array
+        audio_data = audio_data.flatten()
+        
+        # Step 4: Check and adjust duration
+        duration = len(audio_data) / sr
+        print(f"   Duration: {duration:.2f}s")
+        
+        if duration < 1:
+            return None, "❌ Audio too short (minimum 1 second)"
+        
+        if duration > 30:
+            print(f"   ⚠️ Truncating audio from {duration:.2f}s to 30s")
+            audio_data = audio_data[:sr * 30]
+        
+        # Step 5: Convert to torch tensor
+        audio_tensor = torch.from_numpy(audio_data).float()
+        
+        # Step 6: Add channel dimension (1, samples)
+        if audio_tensor.dim() == 1:
+            audio_tensor = audio_tensor.unsqueeze(0)
+        
+        print(f"   Tensor shape: {audio_tensor.shape}")
+        
+        # Step 7: Resample if needed
+        if sr != target_sr:
+            print(f"   🔄 Resampling from {sr}Hz to {target_sr}Hz...")
+            resampler = T.Resample(sr, target_sr)
+            audio_tensor = resampler(audio_tensor)
+            sr = target_sr
+        
+        # Step 8: Save to temporary file
+        temp_path = tempfile.mktemp(suffix='.wav')
+        torchaudio.save(temp_path, audio_tensor, sr)
+        
+        # Step 9: Post-process (trim silence, normalize)
+        temp_path = postprocess(temp_path)
+        
+        print(f"   ✅ Audio processed and saved: {os.path.basename(temp_path)}")
+        return temp_path
+        
+    except Exception as e:
+        print(f"❌ Error processing audio: {e}")
+        import traceback
+        traceback.print_exc()
+        return None
+
+def zero_shot_tts(tts_text, prompt_text, prompt_audio, seed, speed):
+    """Zero-shot TTS synthesis - following official code structure"""
+    try:
+        # Validation
+        if not tts_text or not tts_text.strip():
+            return None, "❌ Please provide text to synthesize"
+        
+        if len(tts_text) > 200:
+            return None, "❌ Text too long, please keep within 200 characters"
+        
+        if not prompt_audio:
+            return None, "❌ Please upload reference audio"
+        
+        if not prompt_text or not prompt_text.strip():
+            return None, "❌ Please provide prompt text"
+        
+        # Load model
+        model = load_model()
+        
+        # Process audio
+        prompt_audio_path = process_audio(prompt_audio)
+        if prompt_audio_path is None:
+            return None, "❌ Failed to process audio"
+        
+        # Check sample rate
+        info = torchaudio.info(prompt_audio_path)
         if info.sample_rate < prompt_sr:
-            gr.Warning(t(ui_lang, "warn_prompt_sr_low").format(info.sample_rate, prompt_sr))
-            return (target_sr, default_data)
-
-        if info.num_frames / info.sample_rate > 10:
-            gr.Warning(t(ui_lang, "warn_prompt_too_long_10s"))
-            return (target_sr, default_data)
-
-        if prompt_text == "":
-            gr.Warning(t(ui_lang, "warn_prompt_text_missing"))
-            return (target_sr, default_data)
-
-        if instruct_text != "":
-            gr.Info(t(ui_lang, "info_instruct_ignored"))
-
-    if mode_value == MODE_ZERO_SHOT:
-        logging.info("get zero_shot inference request")
+            return None, f"❌ Audio sample rate {info.sample_rate} is below {prompt_sr}Hz"
+        
+        # Check duration
+        duration = info.num_frames / info.sample_rate
+        if duration > 10:
+            return None, "❌ Please keep prompt audio within 10 seconds"
+        
+        # Clean inputs
+        tts_text = tts_text.strip()
+        prompt_text = prompt_text.strip()
+        
+        # Build prompt following official format
+        # IMPORTANT: This is the official format from the code
+        full_prompt = f"You are a helpful assistant.<|endofprompt|>{prompt_text}"
+        
+        print(f"\n🎵 Generating speech...")
+        print(f"   TTS text: '{tts_text[:100]}{'...' if len(tts_text) > 100 else ''}'")
+        print(f"   Prompt text: '{prompt_text[:50]}{'...' if len(prompt_text) > 50 else ''}'")
+        print(f"   Full prompt: '{full_prompt[:80]}{'...' if len(full_prompt) > 80 else ''}'")
+        print(f"   Seed: {seed}, Speed: {speed}")
+        
+        # Set random seed
         set_all_random_seed(seed)
+        
+        # Generate - following official code exactly
         speech_list = []
-        for i in cosyvoice.inference_zero_shot(
-            tts_text,
-            "You are a helpful assistant.<|endofprompt|>" + prompt_text,
-            postprocess(prompt_wav),
-            stream=stream,
-            speed=speed,
+        for i in model.inference_zero_shot(
+            tts_text,              # Text to synthesize
+            full_prompt,           # Prompt with special format
+            prompt_audio_path,     # Processed prompt audio
+            stream=False,
+            speed=speed
         ):
             speech_list.append(i["tts_speech"])
-        return (target_sr, torch.concat(speech_list, dim=1).numpy().flatten())
-
-    if mode_value == MODE_INSTRUCT:
-        logging.info("get instruct inference request")
+        
+        # Concatenate all speech segments
+        output_speech = torch.concat(speech_list, dim=1)
+        
+        # Clean up
+        if os.path.exists(prompt_audio_path):
+            os.remove(prompt_audio_path)
+        
+        print(f"   ✅ Generated audio shape: {output_speech.shape}")
+        print("✅ Speech generated successfully!\n")
+        
+        # Return as numpy array for Gradio
+        return (target_sr, output_speech.numpy().flatten()), "✅ Success!"
+    
+    except Exception as e:
+        print(f"❌ Error in zero_shot_tts: {e}")
+        import traceback
+        traceback.print_exc()
+        
+        # Clean up on error
+        try:
+            if prompt_audio_path and os.path.exists(prompt_audio_path):
+                os.remove(prompt_audio_path)
+        except:
+            pass
+        
+        return None, f"❌ Error: {str(e)}"
+
+def instruct_tts(tts_text, instruct_text, prompt_audio, seed, speed):
+    """Instruction-based TTS - following official code structure"""
+    try:
+        # Validation
+        if not tts_text or not tts_text.strip():
+            return None, "❌ Please provide text to synthesize"
+        
+        if len(tts_text) > 200:
+            return None, "❌ Text too long, please keep within 200 characters"
+        
+        if not prompt_audio:
+            return None, "❌ Please upload reference audio"
+        
+        if not instruct_text or not instruct_text.strip():
+            return None, "❌ Please provide instruction text"
+        
+        # Load model
+        model = load_model()
+        
+        # Process audio
+        prompt_audio_path = process_audio(prompt_audio)
+        if prompt_audio_path is None:
+            return None, "❌ Failed to process audio"
+        
+        # Clean inputs
+        tts_text = tts_text.strip()
+        instruct_text = instruct_text.strip()
+        
+        print(f"\n📝 Generating speech with instruction...")
+        print(f"   TTS text: '{tts_text[:100]}{'...' if len(tts_text) > 100 else ''}'")
+        print(f"   Instruction: '{instruct_text}'")
+        print(f"   Seed: {seed}, Speed: {speed}")
+        
+        # Set random seed
         set_all_random_seed(seed)
+        
+        # Generate - following official code
         speech_list = []
-        for i in cosyvoice.inference_instruct2(
-            tts_text,
-            instruct_text,
-            postprocess(prompt_wav),
-            stream=stream,
-            speed=speed,
+        for i in model.inference_instruct2(
+            tts_text,              # Text to synthesize
+            instruct_text,         # Instruction
+            prompt_audio_path,     # Processed prompt audio
+            stream=False,
+            speed=speed
         ):
             speech_list.append(i["tts_speech"])
-        return (target_sr, torch.concat(speech_list, dim=1).numpy().flatten())
-
-    gr.Warning(t(ui_lang, "warn_invalid_mode"))
-    return (target_sr, default_data)
-
-
-def on_mode_change(mode_value, ui_lang):
-    return steps_for(ui_lang, mode_value)
-
-
-def on_language_change(ui_lang, current_mode_value):
-    lang = ui_lang if ui_lang in (LANG_EN, LANG_ZH) else LANG_ZH
-    return (
-        gr.update(value=UI_TEXT[lang]["md_links"]),  # md_links
-        gr.update(value=UI_TEXT[lang]["md_hint"]),  # md_hint
-        gr.update(label=t(lang, "lang_label")),  # lang_radio label
-        gr.update(choices=mode_choices(lang), label=t(lang, "mode_label")),  # mode radio
-        gr.update(value=steps_for(lang, current_mode_value), label=t(lang, "steps_label")),  # steps box
-        gr.update(
-            choices=[(t(lang, "stream_no"), False)],
-            label=t(lang, "stream_label"),
-            value=False,
-        ),  # stream radio
-        gr.update(value=t(lang, "dice")),  # seed button text
-        gr.update(label=t(lang, "seed_label")),  # seed label
-        gr.update(label=t(lang, "tts_label"), value=t(lang, "tts_default")),  # tts textbox
-        gr.update(label=t(lang, "upload_label")),  # upload label
-        gr.update(label=t(lang, "record_label")),  # record label
-        gr.update(label=t(lang, "prompt_text_label"), placeholder=t(lang, "prompt_text_ph")),  # prompt text
-        gr.update(label=t(lang, "instruct_label")),  # instruct dropdown
-        gr.update(value=t(lang, "generate_btn")),  # generate button
-        gr.update(label=t(lang, "output_label")),  # output label
-    )
-
-
-def main():
-    with gr.Blocks() as demo:
-        md_links = gr.Markdown(UI_TEXT[LANG_ZH]["md_links"])
-        md_hint = gr.Markdown(UI_TEXT[LANG_ZH]["md_hint"])
-
-        lang_radio = gr.Radio(
-            choices=[LANG_EN, LANG_ZH],
-            value=LANG_ZH,
-            label=t(LANG_ZH, "lang_label"),
-        )
-
-        tts_text = gr.Textbox(
-            label=t(LANG_ZH, "tts_label"),
-            lines=1,
-            value=t(LANG_ZH, "tts_default"),
-        )
-
-        with gr.Row():
-            mode_radio = gr.Radio(
-                choices=mode_choices(LANG_ZH),
-                label=t(LANG_ZH, "mode_label"),
-                value=MODE_ZERO_SHOT,
-            )
-            steps_box = gr.Textbox(
-                label=t(LANG_ZH, "steps_label"),
-                value=steps_for(LANG_ZH, MODE_ZERO_SHOT),
-                lines=4,
-                interactive=False,
-                scale=0.5,
-            )
-            stream = gr.Radio(
-                choices=[(t(LANG_ZH, "stream_no"), False)],
-                label=t(LANG_ZH, "stream_label"),
-                value=False,
+        
+        # Concatenate all speech segments
+        output_speech = torch.concat(speech_list, dim=1)
+        
+        # Clean up
+        if os.path.exists(prompt_audio_path):
+            os.remove(prompt_audio_path)
+        
+        print(f"   ✅ Generated audio shape: {output_speech.shape}")
+        print("✅ Speech generated successfully!\n")
+        
+        # Return as numpy array for Gradio
+        return (target_sr, output_speech.numpy().flatten()), "✅ Success!"
+    
+    except Exception as e:
+        print(f"❌ Error: {e}")
+        import traceback
+        traceback.print_exc()
+        
+        # Clean up on error
+        try:
+            if prompt_audio_path and os.path.exists(prompt_audio_path):
+                os.remove(prompt_audio_path)
+        except:
+            pass
+        
+        return None, f"❌ Error: {str(e)}"
+
+# Instruction options (from official code)
+instruct_options = [
+    "You are a helpful assistant. 请用广东话表达。<|endofprompt|>",
+    "You are a helpful assistant. 请用尽可能快地语速说一句话。<|endofprompt|>",
+    "You are a helpful assistant. 请用正常的语速说一句话。<|endofprompt|>",
+    "You are a helpful assistant. 请用慢一点的语速说一句话。<|endofprompt|>",
+    "You are a helpful assistant. Please speak in a professional tone.<|endofprompt|>",
+    "You are a helpful assistant. Please speak in a friendly tone.<|endofprompt|>",
+]
+
+# Create Gradio interface
+with gr.Blocks(title="Fun-CosyVoice3 TTS") as demo:
+    gr.Markdown("""
+    # 🎙️ Fun-CosyVoice3-0.5B Text-to-Speech
+    
+    Advanced multilingual zero-shot TTS system supporting **9 languages** and **18+ Chinese dialects**.
+    
+    Based on the official [CosyVoice](https://github.com/FunAudioLLM/CosyVoice) implementation.
+    """)
+    
+    with gr.Tabs():
+        # Tab 1: Zero-Shot TTS
+        with gr.Tab("🎯 Zero-Shot Voice Cloning (3s Fast Cloning)"):
+            gr.Markdown("""
+            ### Clone any voice with 3-10 seconds of reference audio
+            
+            **Steps:**
+            1. Upload or record reference audio (≤30s, ≥16kHz)
+            2. Enter the **prompt text** (transcription of the reference audio)
+            3. Enter the **text to synthesize** (what you want the voice to say)
+            4. Click Generate
+            """)
+            
+            with gr.Row():
+                with gr.Column():
+                    zs_tts_text = gr.Textbox(
+                        label="Text to synthesize (what will be spoken)",
+                        placeholder="Enter the text you want to synthesize...",
+                        lines=2,
+                        value="Her handwriting is very neat, which suggests she likes things tidy."
+                    )
+                    
+                    zs_prompt_audio = gr.Audio(
+                        label="Reference audio (upload or record)",
+                        type="numpy"
+                    )
+                    
+                    zs_prompt_text = gr.Textbox(
+                        label="Prompt text (transcription of reference audio)",
+                        placeholder="Enter what is said in the reference audio...",
+                        lines=2,
+                        value=""
+                    )
+                    
+                    with gr.Row():
+                        zs_seed = gr.Number(label="Random seed", value=0, precision=0)
+                        zs_speed = gr.Slider(label="Speed", minimum=0.5, maximum=2.0, value=1.0, step=0.1)
+                    
+                    zs_btn = gr.Button("🎵 Generate Speech", variant="primary", size="lg")
+                
+                with gr.Column():
+                    zs_output = gr.Audio(label="Generated speech")
+                    zs_status = gr.Textbox(label="Status", interactive=False)
+            
+            zs_btn.click(
+                fn=zero_shot_tts,
+                inputs=[zs_tts_text, zs_prompt_text, zs_prompt_audio, zs_seed, zs_speed],
+                outputs=[zs_output, zs_status]
             )
-            with gr.Column(scale=0.25):
-                seed_button = gr.Button(value=t(LANG_ZH, "dice"))
-                seed = gr.Number(value=0, label=t(LANG_ZH, "seed_label"))
-
-        with gr.Row():
-            prompt_wav_upload = gr.Audio(
-                sources="upload",
-                type="filepath",
-                label=t(LANG_ZH, "upload_label"),
-            )
-            prompt_wav_record = gr.Audio(
-                sources="microphone",
-                type="filepath",
-                label=t(LANG_ZH, "record_label"),
+            
+            gr.Markdown("""
+            **Important:**
+            - **Text to synthesize**: The new text you want to hear in the cloned voice
+            - **Prompt text**: Transcription of what is said in your reference audio
+            - **Reference audio**: 3-10 seconds of clear speech
+            
+            **Example:**
+            - Reference audio: Someone saying "Hello, how are you?"
+            - Prompt text: "Hello, how are you?"
+            - Text to synthesize: "This is a test of voice cloning"
+            - Result: "This is a test of voice cloning" in the cloned voice
+            """)
+        
+        # Tab 2: Instruction-Based TTS
+        with gr.Tab("📝 Instruction-Based Control (Natural Language)"):
+            gr.Markdown("""
+            ### Control voice characteristics with natural language instructions
+            
+            **Steps:**
+            1. Upload or record reference audio
+            2. Select or enter instruction (speed, dialect, emotion)
+            3. Enter text to synthesize
+            4. Click Generate
+            """)
+            
+            with gr.Row():
+                with gr.Column():
+                    inst_tts_text = gr.Textbox(
+                        label="Text to synthesize",
+                        placeholder="Enter your text...",
+                        lines=2,
+                        value="Welcome to the natural language control demo."
+                    )
+                    
+                    inst_prompt_audio = gr.Audio(
+                        label="Reference audio",
+                        type="numpy"
+                    )
+                    
+                    inst_text = gr.Dropdown(
+                        label="Instruction",
+                        choices=instruct_options,
+                        value=instruct_options[0]
+                    )
+                    
+                    with gr.Row():
+                        inst_seed = gr.Number(label="Random seed", value=0, precision=0)
+                        inst_speed = gr.Slider(label="Speed", minimum=0.5, maximum=2.0, value=1.0, step=0.1)
+                    
+                    inst_btn = gr.Button("🎵 Generate Speech", variant="primary", size="lg")
+                
+                with gr.Column():
+                    inst_output = gr.Audio(label="Generated speech")
+                    inst_status = gr.Textbox(label="Status", interactive=False)
+            
+            inst_btn.click(
+                fn=instruct_tts,
+                inputs=[inst_tts_text, inst_text, inst_prompt_audio, inst_seed, inst_speed],
+                outputs=[inst_output, inst_status]
             )
-
-        prompt_text = gr.Textbox(
-            label=t(LANG_ZH, "prompt_text_label"),
-            lines=1,
-            placeholder=t(LANG_ZH, "prompt_text_ph"),
-            value="",
-        )
-        instruct_text = gr.Dropdown(
-            choices=instruct_list,
-            label=t(LANG_ZH, "instruct_label"),
-            value=instruct_list[0],
-        )
-
-        generate_button = gr.Button(t(LANG_ZH, "generate_btn"))
-        audio_output = gr.Audio(
-            label=t(LANG_ZH, "output_label"),
-            autoplay=True,
-            streaming=False,
-        )
-
-        seed_button.click(generate_seed, inputs=[], outputs=seed)
-
-        generate_button.click(
-            generate_audio,
-            inputs=[
-                tts_text,
-                mode_radio,
-                prompt_text,
-                prompt_wav_upload,
-                prompt_wav_record,
-                instruct_text,
-                seed,
-                stream,
-                lang_radio,  # ui_lang
-            ],
-            outputs=[audio_output],
-        )
-
-        mode_radio.change(
-            fn=on_mode_change,
-            inputs=[mode_radio, lang_radio],
-            outputs=[steps_box],
-        )
-
-        prompt_wav_upload.change(
-            fn=prompt_wav_recognition,
-            inputs=[prompt_wav_upload],
-            outputs=[prompt_text],
-        )
-        prompt_wav_record.change(
-            fn=prompt_wav_recognition,
-            inputs=[prompt_wav_record],
-            outputs=[prompt_text],
-        )
-
-        lang_radio.change(
-            fn=on_language_change,
-            inputs=[lang_radio, mode_radio],
-            outputs=[
-                md_links,
-                md_hint,
-                lang_radio,
-                mode_radio,
-                steps_box,
-                stream,
-                seed_button,
-                seed,
-                tts_text,
-                prompt_wav_upload,
-                prompt_wav_record,
-                prompt_text,
-                instruct_text,
-                generate_button,
-                audio_output,
-            ],
-        )
-
-    demo.queue(default_concurrency_limit=4).launch()
-
+            
+            gr.Markdown("""
+            **Example instructions:**
+            - "请用广东话表达" (Speak in Cantonese)
+            - "请用尽可能快地语速说" (Speak as fast as possible)
+            - "Please speak in a professional tone"
+            """)
+    
+    gr.Markdown("""
+    ---
+    ### 📋 Supported Languages & Dialects
+    
+    **Languages:** Chinese, English, Japanese, Korean, German, Spanish, French, Italian, Russian
+    
+    **Chinese Dialects:** Guangdong, Minnan, Sichuan, Dongbei, Shanxi, Shanghai, Tianjin, Shandong, and more
+    
+    ### ⚡ Performance
+    - Model: Fun-CosyVoice3-0.5B (500M parameters)
+    - Sample Rate: 24kHz
+    - Latency: ~5-10s on CPU, ~2-3s on GPU
+    
+    ### 📚 Resources
+    [Paper](https://arxiv.org/abs/2505.17589) • [GitHub](https://github.com/FunAudioLLM/CosyVoice) • [Model](https://huggingface.co/FunAudioLLM/Fun-CosyVoice3-0.5B-2512)
+    """)
 
 if __name__ == "__main__":
-    cosyvoice = CosyVoiceAutoModel(
-        model_dir="pretrained_models/Fun-CosyVoice3-0.5B",
-        load_trt=False,
-        fp16=False,
-    )
-    sft_spk = cosyvoice.list_available_spks()
-
-    for stream in [False]:
-        for i, j in enumerate(
-            cosyvoice.inference_zero_shot(
-                "收到好友从远方寄来的生日礼物，那份意外的惊喜与深深的祝福让我心中充满了甜蜜的快乐，笑容如花儿般绽放。",
-                "You are a helpful assistant.<|endofprompt|>希望你以后能够做的比我还好呦。",
-                "zero_shot_prompt.wav",
-                stream=stream,
-            )
-        ):
-            continue
-
-    prompt_sr = 16000
-    target_sr = 24000
-    default_data = np.zeros(target_sr)
-
-    model_dir = "pretrained_models/SenseVoiceSmall"
-    asr_model = AutoModel(
-        model=model_dir,
-        disable_update=True,
-        log_level="DEBUG",
-        device="cuda:0",
-    )
-
-    main()
+    print("\n🚀 Launching Gradio interface...")
+    demo.queue(max_size=10, default_concurrency_limit=2)
+    demo.launch(
+        server_name="0.0.0.0",
+        server_port=7860,
+        show_error=True
+    )