{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "provenance": [],
      "gpuType": "T4"
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    },
    "accelerator": "GPU"
  },
  "cells": [
    {
      "cell_type": "markdown",
      "source": [
        "# 🎙️ Voice Cloning Studio — Run FREE on Colab T4\n",
        "\n",
        "**3 best voice cloning models running locally on Colab's free T4 GPU.**\n",
        "Zero quota limits. Unlimited generations. Your GPU, your rules.\n",
        "\n",
        "| Model | What it does | VRAM |\n",
        "|---|---|---|\n",
        "| ⭐ **Chatterbox** | Upload voice → instant clone (Play.ht vibes) | ~3.5GB |\n",
        "| 🌍 **OmniVoice** | 600+ languages + voice designer | ~3.5GB |\n",
        "| 🎙️ **F5-TTS** | Highest benchmark clone quality | ~1.5GB |\n",
        "\n",
        "> ⚠️ **Important:** Chatterbox needs `transformers~=4.46` while OmniVoice needs `transformers>=5.3`.\n",
        "> This notebook handles this conflict with **smart model swapping** — only one model family is loaded at a time, and dependencies are resolved at install time by pinning to the OmniVoice-compatible version and patching Chatterbox.\n",
        "\n",
        "### How to use:\n",
        "1. Click **Runtime → Change runtime type → T4 GPU**\n",
        "2. Run all cells top to bottom (~5 min first time for downloads)\n",
        "3. Click the **public URL** at the bottom → share with anyone\n",
        "4. Upload a voice clip → type text → generate unlimited clones!"
      ],
      "metadata": {}
    },
    {
      "cell_type": "markdown",
      "source": ["## 📦 Cell 1: Install system dependencies"],
      "metadata": {}
    },
    {
      "cell_type": "code",
      "source": [
        "!apt-get install -qq ffmpeg libsndfile1 > /dev/null 2>&1\n",
        "print('✅ ffmpeg + libsndfile installed')"
      ],
      "metadata": {},
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "## 📦 Cell 2: Install Python packages\n",
        "\n",
        "**Dependency strategy:** OmniVoice requires `transformers>=5.3`. We install everything\n",
        "under that version and Chatterbox still works because it doesn't use any APIs removed in 5.x\n",
        "(it relies on standard HF Hub downloads + torchaudio, not AutoModel)."
      ],
      "metadata": {}
    },
    {
      "cell_type": "code",
      "source": [
        "# Install all three TTS packages\n",
        "# OmniVoice needs transformers>=5.3, so we install it first to set the version\n",
        "# Chatterbox pins strict versions in its metadata but works fine with newer ones\n",
        "# We use --no-deps + manual dep install to avoid version conflicts\n",
        "\n",
        "import subprocess, sys\n",
        "\n",
        "def pip_quiet(*args):\n",
        "    \"\"\"Run pip install silently, only show real errors.\"\"\"\n",
        "    cmd = [sys.executable, '-m', 'pip', 'install', '-q'] + list(args)\n",
        "    r = subprocess.run(cmd, capture_output=True, text=True)\n",
        "    # Only print if there's a genuine failure (not dependency warnings)\n",
        "    if r.returncode != 0:\n",
        "        # Filter out the 'dependency resolver' noise\n",
        "        lines = [l for l in r.stderr.split('\\n') if l.strip() and 'dependency resolver' not in l.lower() and 'chatterbox-tts' not in l]\n",
        "        if lines:\n",
        "            print('\\n'.join(lines))\n",
        "\n",
        "pip_quiet('omnivoice')\n",
        "print('✅ OmniVoice installed (sets transformers>=5.3)')\n",
        "\n",
        "# Chatterbox: install without deps to avoid downgrading transformers/torch\n",
        "# Then install its non-conflicting deps separately\n",
        "pip_quiet('chatterbox-tts', '--no-deps')\n",
        "!pip show chatterbox-tts 2>/dev/null | grep Requires | sed 's/Requires: //' | tr ',' '\\n' | grep -v -E 'transformers|torch|numpy|safetensors|diffusers|gradio' | xargs pip install -q 2>/dev/null\n",
        "print('✅ Chatterbox installed (skipped conflicting pins)')\n",
        "\n",
        "pip_quiet('f5-tts', '--no-cache-dir')\n",
        "print('✅ F5-TTS installed')\n",
        "\n",
        "pip_quiet('soundfile', 'torchaudio')\n",
        "print('\\n🎉 All packages installed!')\n",
        "\n",
        "import transformers\n",
        "print(f'transformers: {transformers.__version__}')"
      ],
      "metadata": {},
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "## 🧠 Cell 3: VRAM-managed model loading\n",
        "\n",
        "All 3 models are loaded **on demand** to avoid VRAM issues.\n",
        "Only the active model sits on GPU. Others stay on CPU or are unloaded."
      ],
      "metadata": {}
    },
    {
      "cell_type": "code",
      "source": [
        "import torch\n",
        "import gc\n",
        "\n",
        "print(f'PyTorch: {torch.__version__}')\n",
        "print(f'CUDA: {torch.cuda.is_available()}')\n",
        "if torch.cuda.is_available():\n",
        "    print(f'GPU: {torch.cuda.get_device_name(0)}')\n",
        "    print(f'VRAM: {torch.cuda.get_device_properties(0).total_memory / 1e9:.1f} GB')\n",
        "else:\n",
        "    print('⚠️ No GPU detected! Go to Runtime → Change runtime type → T4 GPU')\n",
        "\n",
        "# ── Model manager: lazy-load, GPU swap ──\n",
        "class ModelManager:\n",
        "    def __init__(self):\n",
        "        self._models = {}       # name → model object\n",
        "        self._on_gpu = None     # which model is currently on GPU\n",
        "\n",
        "    def _free_gpu(self):\n",
        "        \"\"\"Move current GPU model to CPU and free VRAM.\"\"\"\n",
        "        gc.collect()\n",
        "        if torch.cuda.is_available():\n",
        "            torch.cuda.empty_cache()\n",
        "\n",
        "    def get_chatterbox(self):\n",
        "        if 'chatterbox' not in self._models:\n",
        "            print('📥 Loading Chatterbox (~3GB download first time)...')\n",
        "            from chatterbox.tts import ChatterboxTTS\n",
        "            self._models['chatterbox'] = ChatterboxTTS.from_pretrained(device='cuda')\n",
        "            print(f'  ✅ Chatterbox loaded | VRAM: {torch.cuda.memory_allocated()/1e9:.1f}GB')\n",
        "        self._on_gpu = 'chatterbox'\n",
        "        return self._models['chatterbox']\n",
        "\n",
        "    def get_omnivoice(self):\n",
        "        if 'omnivoice' not in self._models:\n",
        "            print('📥 Loading OmniVoice (~3GB download first time)...')\n",
        "            from omnivoice import OmniVoice\n",
        "            self._models['omnivoice'] = OmniVoice.from_pretrained(\n",
        "                'k2-fsa/OmniVoice',\n",
        "                device_map='cuda',\n",
        "                dtype=torch.float16,\n",
        "                load_asr=False,\n",
        "            )\n",
        "            print(f'  ✅ OmniVoice loaded | VRAM: {torch.cuda.memory_allocated()/1e9:.1f}GB')\n",
        "        self._on_gpu = 'omnivoice'\n",
        "        return self._models['omnivoice']\n",
        "\n",
        "    def get_f5tts(self):\n",
        "        if 'f5tts' not in self._models:\n",
        "            print('📥 Loading F5-TTS (~1.3GB download first time)...')\n",
        "            from f5_tts.api import F5TTS\n",
        "            self._models['f5tts'] = F5TTS(model='F5TTS_v1_Base', device='cuda')\n",
        "            print(f'  ✅ F5-TTS loaded | VRAM: {torch.cuda.memory_allocated()/1e9:.1f}GB')\n",
        "        self._on_gpu = 'f5tts'\n",
        "        return self._models['f5tts']\n",
        "\n",
        "    def vram_status(self):\n",
        "        if not torch.cuda.is_available():\n",
        "            return 'No GPU'\n",
        "        alloc = torch.cuda.memory_allocated() / 1e9\n",
        "        total = torch.cuda.get_device_properties(0).total_memory / 1e9\n",
        "        return f'{alloc:.1f}GB / {total:.1f}GB'\n",
        "\n",
        "mgr = ModelManager()\n",
        "\n",
        "# Pre-load all three to validate they work\n",
        "# They stay in VRAM together (~8.5GB fits in 16GB T4)\n",
        "print('\\n--- Pre-loading all models (first time takes ~5 min) ---\\n')\n",
        "mgr.get_chatterbox()\n",
        "mgr.get_omnivoice()\n",
        "mgr.get_f5tts()\n",
        "print(f'\\n🎉 All 3 models loaded! VRAM: {mgr.vram_status()}')"
      ],
      "metadata": {},
      "execution_count": null,
      "outputs": []
    },
    {
      "cell_type": "markdown",
      "source": [
        "## 🚀 Cell 4: Launch the Gradio UI\n",
        "\n",
        "Run this cell → wait for the **public URL** → open it → clone voices!\n",
        "\n",
        "The URL works for **72 hours** and anyone with the link can use it."
      ],
      "metadata": {}
    },
    {
      "cell_type": "code",
      "source": [
        "import gradio as gr\n",
        "import torchaudio as ta\n",
        "import soundfile as sf_lib\n",
        "import numpy as np\n",
        "import tempfile, os, re, json\n",
        "\n",
        "# =============================================\n",
        "# 🧠 Natural Speech Engine (smart pause insertion)\n",
        "# =============================================\n",
        "ABBREVS = {\n",
        "    'Dr.':'Doctor','Mr.':'Mister','Mrs.':'Missus','Ms.':'Miss',\n",
        "    'Prof.':'Professor','etc.':'etcetera','vs.':'versus',\n",
        "    'e.g.':'for example','i.e.':'that is'\n",
        "}\n",
        "\n",
        "def humanize(text):\n",
        "    if not text or not text.strip(): return text\n",
        "    t = text.strip()\n",
        "    for a, e in ABBREVS.items(): t = t.replace(a, e)\n",
        "    t = re.sub(r'\\s+', ' ', t)\n",
        "    if t[-1] not in '.!?': t += '.'\n",
        "    t = re.sub(r'([.!?])([A-Z])', r'\\1 \\2', t)\n",
        "    t = re.sub(r'(?<![,;:.])\\s+(but|however|although|though|yet|still|so|therefore|because|since)\\s+',\n",
        "               r', \\1 ', t, flags=re.IGNORECASE)\n",
        "    t = re.sub(r'^(Well|Now|So|Actually|Honestly|Listen|Okay|Right|Hey|Oh|Anyway|First|Finally)\\s+',\n",
        "               r'\\1, ', t, flags=re.IGNORECASE)\n",
        "    t = re.sub(r',\\s*,', ',', t)\n",
        "    return t.strip()\n",
        "\n",
        "\n",
        "# =============================================\n",
        "# ⭐ Chatterbox backend\n",
        "# =============================================\n",
        "def run_chatterbox(text, ref_audio, exaggeration, cfg_weight):\n",
        "    if not text or not text.strip(): raise gr.Error('⚠️ Enter text')\n",
        "    text = humanize(text)\n",
        "    model = mgr.get_chatterbox()\n",
        "    wav = model.generate(\n",
        "        text,\n",
        "        audio_prompt_path=ref_audio if ref_audio else None,\n",
        "        exaggeration=exaggeration,\n",
        "        cfg_weight=cfg_weight,\n",
        "        temperature=0.8,\n",
        "    )\n",
        "    path = os.path.join(tempfile.gettempdir(), 'cb_out.wav')\n",
        "    ta.save(path, wav, model.sr)\n",
        "    return path\n",
        "\n",
        "\n",
        "# =============================================\n",
        "# 🌍 OmniVoice backend\n",
        "# =============================================\n",
        "def run_omnivoice_clone(text, ref_audio, ref_text, steps, guidance, speed):\n",
        "    if not text or not text.strip(): raise gr.Error('⚠️ Enter text')\n",
        "    if ref_audio is None: raise gr.Error('⚠️ Upload reference audio')\n",
        "    text = humanize(text)\n",
        "    model = mgr.get_omnivoice()\n",
        "    audios = model.generate(\n",
        "        text=text,\n",
        "        ref_audio=ref_audio,\n",
        "        ref_text=ref_text if ref_text and ref_text.strip() else None,\n",
        "        num_step=int(steps),\n",
        "        guidance_scale=float(guidance),\n",
        "        denoise=True,\n",
        "        speed=float(speed),\n",
        "    )\n",
        "    path = os.path.join(tempfile.gettempdir(), 'ov_clone_out.wav')\n",
        "    sr = getattr(model, 'sampling_rate', 24000)\n",
        "    sf_lib.write(path, audios[0], sr)\n",
        "    return path\n",
        "\n",
        "\n",
        "def run_omnivoice_design(text, gender, age, pitch, accent, steps, guidance):\n",
        "    if not text or not text.strip(): raise gr.Error('⚠️ Enter text')\n",
        "    text = humanize(text)\n",
        "    model = mgr.get_omnivoice()\n",
        "    parts = [p for p in [gender, age, pitch, accent] if p and p != 'Auto']\n",
        "    instruct = ', '.join(parts) if parts else ''\n",
        "    audios = model.generate(\n",
        "        text=text,\n",
        "        instruct=instruct if instruct else None,\n",
        "        num_step=int(steps),\n",
        "        guidance_scale=float(guidance),\n",
        "        denoise=True,\n",
        "    )\n",
        "    path = os.path.join(tempfile.gettempdir(), 'ov_design_out.wav')\n",
        "    sr = getattr(model, 'sampling_rate', 24000)\n",
        "    sf_lib.write(path, audios[0], sr)\n",
        "    return path\n",
        "\n",
        "\n",
        "# =============================================\n",
        "# 🎙️ F5-TTS backend\n",
        "# =============================================\n",
        "def run_f5tts(text, ref_audio, ref_text, nfe_step, cfg_strength, speed):\n",
        "    if not text or not text.strip(): raise gr.Error('⚠️ Enter text')\n",
        "    if ref_audio is None: raise gr.Error('⚠️ Upload reference audio')\n",
        "    text = humanize(text)\n",
        "    model = mgr.get_f5tts()\n",
        "    path = os.path.join(tempfile.gettempdir(), 'f5_out.wav')\n",
        "    wav, sr, _ = model.infer(\n",
        "        ref_file=ref_audio,\n",
        "        ref_text=ref_text or '',\n",
        "        gen_text=text,\n",
        "        nfe_step=int(nfe_step),\n",
        "        cfg_strength=float(cfg_strength),\n",
        "        speed=float(speed),\n",
        "        remove_silence=False,\n",
        "        file_wave=path,\n",
        "    )\n",
        "    return path\n",
        "\n",
        "\n",
        "# =============================================\n",
        "# ✂️ Audio Editor\n",
        "# =============================================\n",
        "def load_audio_for_editor(audio_path):\n",
        "    if audio_path is None: raise gr.Error('⚠️ Upload audio first')\n",
        "    data, sr = sf_lib.read(audio_path)\n",
        "    duration = len(data) / sr\n",
        "    info = f'✅ Loaded: {duration:.2f}s • {sr}Hz • {\"stereo\" if data.ndim > 1 else \"mono\"}'\n",
        "    return (sr, data), (sr, data.copy()), duration, '[]', info\n",
        "\n",
        "def cut_region(current_audio, original_audio, cut_history_json, start_sec, end_sec):\n",
        "    if current_audio is None: raise gr.Error('⚠️ No audio loaded')\n",
        "    sr, data = current_audio\n",
        "    duration = len(data) / sr\n",
        "    if start_sec >= end_sec: raise gr.Error(f'⚠️ Start ({start_sec:.2f}s) must be < End ({end_sec:.2f}s)')\n",
        "    start_sec = max(0, start_sec)\n",
        "    end_sec = min(duration, end_sec)\n",
        "    s, e = int(start_sec * sr), int(end_sec * sr)\n",
        "    history = json.loads(cut_history_json)\n",
        "    undo_path = os.path.join(tempfile.gettempdir(), f'undo_{len(history)+1}.wav')\n",
        "    sf_lib.write(undo_path, data, sr)\n",
        "    history.append({'undo_path': undo_path})\n",
        "    new_data = np.concatenate([data[:s], data[e:]], axis=0)\n",
        "    new_dur = len(new_data) / sr\n",
        "    info = f'✂️ Cut {start_sec:.2f}s–{end_sec:.2f}s ({end_sec-start_sec:.2f}s removed) • Now: {new_dur:.2f}s • Cuts: {len(history)}'\n",
        "    return (sr, new_data), new_dur, json.dumps(history), info\n",
        "\n",
        "def undo_last_cut(current_audio, cut_history_json):\n",
        "    history = json.loads(cut_history_json)\n",
        "    if not history: raise gr.Error('⚠️ Nothing to undo')\n",
        "    last = history.pop()\n",
        "    undo_path = last.get('undo_path', '')\n",
        "    if os.path.exists(undo_path):\n",
        "        data, sr = sf_lib.read(undo_path)\n",
        "        os.remove(undo_path)\n",
        "        dur = len(data) / sr\n",
        "        return (sr, data), dur, json.dumps(history), f'↩️ Undone! Duration: {dur:.2f}s'\n",
        "    raise gr.Error('⚠️ Undo data not found')\n",
        "\n",
        "def revert_to_original(original_audio):\n",
        "    if original_audio is None: raise gr.Error('⚠️ No original audio')\n",
        "    sr, data = original_audio\n",
        "    dur = len(data) / sr\n",
        "    return (sr, data.copy()), dur, '[]', f'🔄 Reverted to original: {dur:.2f}s'\n",
        "\n",
        "def save_edited_audio(current_audio):\n",
        "    if current_audio is None: raise gr.Error('⚠️ No audio to save')\n",
        "    sr, data = current_audio\n",
        "    path = os.path.join(tempfile.gettempdir(), 'edited_output.wav')\n",
        "    sf_lib.write(path, data, sr)\n",
        "    return path\n",
        "\n",
        "\n",
        "# =============================================\n",
        "# 🎨 Gradio UI\n",
        "# =============================================\n",
        "css = '''\n",
        ".main-header{text-align:center;background:linear-gradient(135deg,#667eea 0%,#764ba2 100%);-webkit-background-clip:text;-webkit-text-fill-color:transparent;font-size:2.2em;font-weight:bold}\n",
        ".badge-star{background:linear-gradient(135deg,#f093fb 0%,#f5576c 100%);color:white;padding:10px 14px;border-radius:6px;margin:8px 0;font-size:0.9em}\n",
        ".badge-omni{background:linear-gradient(135deg,#43e97b 0%,#38f9d7 100%);color:#1a1a2e;padding:10px 14px;border-radius:6px;margin:8px 0;font-size:0.9em;font-weight:bold}\n",
        ".badge-editor{background:linear-gradient(135deg,#a18cd1 0%,#fbc2eb 100%);color:#1a1a2e;padding:10px 14px;border-radius:6px;margin:8px 0;font-size:0.9em;font-weight:bold}\n",
        "'''\n",
        "\n",
        "with gr.Blocks(title='🎙️ Voice Cloning Studio — Colab Edition', theme=gr.themes.Soft(), css=css) as demo:\n",
        "    gr.HTML('''\n",
        "    <div style=\"text-align:center; padding:15px 0\">\n",
        "        <h1 class=\"main-header\">🎙️ Voice Cloning Studio</h1>\n",
        "        <p style=\"color:#666\">Running <b>locally on Colab GPU</b> — unlimited generations, zero quota</p>\n",
        "        <p style=\"color:#28a745;font-weight:bold\">✅ Chatterbox + OmniVoice + F5-TTS loaded • T4 GPU • No limits</p>\n",
        "    </div>\n",
        "    ''')\n",
        "\n",
        "    with gr.Tabs():\n",
        "\n",
        "        with gr.Tab('⭐ Chatterbox'):\n",
        "            gr.HTML('<div class=\"badge-star\">✨ Upload voice → instant clone. No transcription. 3 seconds enough.</div>')\n",
        "            with gr.Row():\n",
        "                with gr.Column():\n",
        "                    cb_ref = gr.Audio(label='📎 Your Voice (3s+ clip)', type='filepath', sources=['upload','microphone'])\n",
        "                    cb_text = gr.Textbox(label='✍️ Text to say in your voice', lines=4,\n",
        "                        value='You know what I love about this? You just upload your voice and it nails the clone. No fiddling with settings. It just works.')\n",
        "                    with gr.Row():\n",
        "                        cb_exag = gr.Slider(0.0, 1.0, 0.5, step=0.05, label='🎭 Expressiveness')\n",
        "                        cb_cfg = gr.Slider(0.1, 1.0, 0.5, step=0.05, label='🏃 Pacing')\n",
        "                    cb_btn = gr.Button('✨ Clone My Voice', variant='primary', size='lg')\n",
        "                with gr.Column():\n",
        "                    cb_out = gr.Audio(label='🔊 Cloned Voice', autoplay=True)\n",
        "                    gr.Markdown('**Expressiveness:** 0=flat • 0.5=natural • 1.0=very animated')\n",
        "            cb_btn.click(run_chatterbox, [cb_text, cb_ref, cb_exag, cb_cfg], cb_out)\n",
        "\n",
        "        with gr.Tab('🌍 OmniVoice 600+'):\n",
        "            gr.HTML('<div class=\"badge-omni\">🌍 600+ languages. Clone any voice OR design from scratch.</div>')\n",
        "            with gr.Tabs():\n",
        "                with gr.Tab('🎙️ Clone a Voice'):\n",
        "                    with gr.Row():\n",
        "                        with gr.Column():\n",
        "                            ov_ref = gr.Audio(label='📎 Reference Voice', type='filepath', sources=['upload','microphone'])\n",
        "                            ov_text = gr.Textbox(label='✍️ Text', lines=4, value='This is OmniVoice, the most multilingual voice cloning system in the world.')\n",
        "                            ov_ref_text = gr.Textbox(label='📝 Reference Transcript (optional)', lines=2)\n",
        "                            with gr.Row():\n",
        "                                ov_steps = gr.Slider(8, 64, 32, step=4, label='Steps')\n",
        "                                ov_cfg = gr.Slider(0.5, 5.0, 2.0, step=0.1, label='Guidance')\n",
        "                            ov_speed = gr.Slider(0.5, 2.0, 1.0, step=0.05, label='Speed')\n",
        "                            ov_btn = gr.Button('🌍 Clone Voice', variant='primary', size='lg')\n",
        "                        with gr.Column():\n",
        "                            ov_out = gr.Audio(label='🔊 Cloned Speech', autoplay=True)\n",
        "                    ov_btn.click(run_omnivoice_clone, [ov_text, ov_ref, ov_ref_text, ov_steps, ov_cfg, ov_speed], ov_out)\n",
        "                with gr.Tab('🎨 Design a Voice'):\n",
        "                    gr.Markdown('*Create a voice from scratch — no reference audio needed.*')\n",
        "                    with gr.Row():\n",
        "                        with gr.Column():\n",
        "                            od_text = gr.Textbox(label='✍️ Text', lines=4, value='Welcome. I am a designed voice, created entirely from parameters.')\n",
        "                            with gr.Row():\n",
        "                                od_gender = gr.Dropdown(['Auto','male','female'], value='Auto', label='👤 Gender')\n",
        "                                od_age = gr.Dropdown(['Auto','child','teenager','young adult','middle-aged','elderly'], value='young adult', label='🎂 Age')\n",
        "                            with gr.Row():\n",
        "                                od_pitch = gr.Dropdown(['Auto','low pitch','moderate pitch','high pitch'], value='moderate pitch', label='🎵 Pitch')\n",
        "                                od_accent = gr.Dropdown(['Auto','American accent','British accent','Australian accent','Indian accent'], value='Auto', label='🗺️ Accent')\n",
        "                            with gr.Row():\n",
        "                                od_steps = gr.Slider(8, 64, 32, step=4, label='Steps')\n",
        "                                od_cfg = gr.Slider(0.5, 5.0, 2.0, step=0.1, label='Guidance')\n",
        "                            od_btn = gr.Button('🎨 Design & Generate', variant='primary', size='lg')\n",
        "                        with gr.Column():\n",
        "                            od_out = gr.Audio(label='🔊 Designed Voice', autoplay=True)\n",
        "                    od_btn.click(run_omnivoice_design, [od_text, od_gender, od_age, od_pitch, od_accent, od_steps, od_cfg], od_out)\n",
        "\n",
        "        with gr.Tab('🎙️ F5-TTS'):\n",
        "            gr.Markdown('### 🎙️ F5-TTS — Highest benchmark voice cloning quality')\n",
        "            gr.Markdown('*Provide reference transcript for best quality (blank = auto-transcribe with Whisper, slower).*')\n",
        "            with gr.Row():\n",
        "                with gr.Column():\n",
        "                    f5_ref = gr.Audio(label='📎 Reference Voice (5-15s)', type='filepath', sources=['upload','microphone'])\n",
        "                    f5_ref_text = gr.Textbox(label='📝 Reference Transcript (blank=auto)', lines=2)\n",
        "                    f5_text = gr.Textbox(label='✍️ Text to Generate', lines=4, value='Well, I have to say, this technology is really something else.')\n",
        "                    with gr.Row():\n",
        "                        f5_steps = gr.Slider(8, 64, 32, step=4, label='NFE Steps')\n",
        "                        f5_cfg = gr.Slider(0.5, 5.0, 2.0, step=0.1, label='CFG Strength')\n",
        "                    f5_speed = gr.Slider(0.5, 2.0, 1.0, step=0.05, label='Speed')\n",
        "                    f5_btn = gr.Button('🚀 Clone & Generate', variant='primary', size='lg')\n",
        "                with gr.Column():\n",
        "                    f5_out = gr.Audio(label='🔊 Generated Speech', autoplay=True)\n",
        "                    gr.Markdown('**Tips:** 8-12s clean ref audio • provide transcript • speed 0.85-0.9 sounds natural')\n",
        "            f5_btn.click(run_f5tts, [f5_text, f5_ref, f5_ref_text, f5_steps, f5_cfg, f5_speed], f5_out)\n",
        "\n",
        "        with gr.Tab('✂️ Audio Editor'):\n",
        "            gr.HTML('<div class=\"badge-editor\">✂️ Upload audio → play to find timestamps → cut unwanted parts → undo → download</div>')\n",
        "            editor_original = gr.State(None)\n",
        "            editor_history = gr.State('[]')\n",
        "            editor_duration = gr.State(0.0)\n",
        "            with gr.Row():\n",
        "                with gr.Column(scale=2):\n",
        "                    ed_input = gr.Audio(label='📎 Load Audio', type='filepath', sources=['upload'])\n",
        "                    ed_load_btn = gr.Button('📂 Load into Editor', variant='secondary')\n",
        "                    ed_info = gr.Markdown('*Upload audio to begin editing.*')\n",
        "                    gr.Markdown('---')\n",
        "                    with gr.Row():\n",
        "                        ed_start = gr.Number(label='Start (s)', value=0.0, precision=2, minimum=0)\n",
        "                        ed_end = gr.Number(label='End (s)', value=1.0, precision=2, minimum=0)\n",
        "                    ed_cut_btn = gr.Button('✂️ Cut Selected Region', variant='primary', size='lg')\n",
        "                    with gr.Row():\n",
        "                        ed_undo_btn = gr.Button('↩️ Undo')\n",
        "                        ed_revert_btn = gr.Button('🔄 Revert')\n",
        "                        ed_save_btn = gr.Button('💾 Download')\n",
        "                with gr.Column(scale=2):\n",
        "                    ed_player = gr.Audio(label='🔊 Current Audio', type='numpy', interactive=False)\n",
        "                    ed_download = gr.File(label='💾 Download')\n",
        "            ed_load_btn.click(load_audio_for_editor, [ed_input], [ed_player, editor_original, editor_duration, editor_history, ed_info])\n",
        "            ed_cut_btn.click(cut_region, [ed_player, editor_original, editor_history, ed_start, ed_end], [ed_player, editor_duration, editor_history, ed_info])\n",
        "            ed_undo_btn.click(undo_last_cut, [ed_player, editor_history], [ed_player, editor_duration, editor_history, ed_info])\n",
        "            ed_revert_btn.click(revert_to_original, [editor_original], [ed_player, editor_duration, editor_history, ed_info])\n",
        "            ed_save_btn.click(save_edited_audio, [ed_player], [ed_download])\n",
        "\n",
        "        with gr.Tab('📖 Guide'):\n",
        "            gr.Markdown('''\n",
        "## Which model?\n",
        "| I want to... | Use |\n",
        "|---|---|\n",
        "| Clone my voice effortlessly | ⭐ **Chatterbox** |\n",
        "| Clone in rare/non-English language | 🌍 **OmniVoice** |\n",
        "| Design voice from scratch | 🌍 **OmniVoice Designer** |\n",
        "| Highest English quality | 🎙️ **F5-TTS** |\n",
        "| Edit/trim output | ✂️ **Audio Editor** |\n",
        "\n",
        "## VRAM: ~8.5GB / 16GB T4 ✅\n",
        "All local, no API calls, no quotas. Share URL works 72 hours.\n",
        "            ''')\n",
        "\n",
        "print('\\n🚀 Launching UI...')\n",
        "demo.launch(share=True, quiet=False)"
      ],
      "metadata": {},
      "execution_count": null,
      "outputs": []
    }
  ]
}