{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Qwen3-TTS 0.6B — Indic Multilingual LoRA Inference\n",
    "\n",
    "Run Indic-language text-to-speech using **[`aguken-ai/Qwen3-TTS-0.6B-LoRA-Finetuned-Indic-Multilingual`](https://huggingface.co/aguken-ai/Qwen3-TTS-0.6B-LoRA-Finetuned-Indic-Multilingual)** (15 languages × 2 genders) on top of `Qwen/Qwen3-TTS-12Hz-0.6B-Base`.\n",
    "\n",
    "1. **Pick a language and gender** in the *Configuration* cell.\n",
    "2. Run the cells in order.\n",
    "3. Edit `TEXT` to synthesize anything you want.\n",
    "\n",
    "**Hardware:** designed for **Colab free T4** (~3-5 GB VRAM in use). Works on Kaggle T4 / P100 too.\n",
    "\n",
    "**Supported languages:** Assamese · Bengali · Bodo · Dogri · Gujarati · Kannada · Konkani · Maithili · Malayalam · Marathi · Nepali · Odia · Punjabi · Tamil · Telugu"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 1. Install dependencies\n",
    "\n",
    "Don't downgrade `torch` — use whatever Colab ships. Pinning older `torch` breaks `torchaudio` on Colab's Python 3.12."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "E: Could not open lock file /var/lib/dpkg/lock-frontend - open (13: Permission denied)\n",
      "E: Unable to acquire the dpkg frontend lock (/var/lib/dpkg/lock-frontend), are you root?\n",
      "✅ Dependencies installed\n"
     ]
    }
   ],
   "source": [
    "!apt-get -qq install -y sox libsox-dev ffmpeg libsndfile1 > /dev/null\n",
    "!pip install -q -U qwen-tts peft transformers soundfile huggingface_hub\n",
    "print('✅ Dependencies installed')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 2. Imports + environment patches"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/zeus/miniconda3/envs/cloudspace/lib/python3.12/site-packages/torchaudio/_extension/__init__.py:39: UserWarning: TorchAudio native extension could not be loaded (/system/conda/miniconda3/envs/cloudspace/lib/python3.12/site-packages/torchaudio/lib/_torchaudio.abi3.so: undefined symbol: torch_library_impl). Falling back to no-extension mode.\n",
      "  warnings.warn(f\"TorchAudio native extension could not be loaded ({_e}). \"\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "✅ Imports OK\n",
      "CUDA available: True\n",
      "Device: NVIDIA L4\n"
     ]
    }
   ],
   "source": [
    "import torch\n",
    "import soundfile as sf\n",
    "import IPython.display as ipd\n",
    "\n",
    "# PEFT patch: Colab ships torchao 0.10 but PEFT requires >=0.16.\n",
    "# We don't use torchao here; short-circuit its dispatcher check so PEFT\n",
    "# falls through to the standard nn.Linear LoRA path.\n",
    "import peft.tuners.lora.torchao as _torchao_dispatch\n",
    "_torchao_dispatch.is_torchao_available = lambda: False\n",
    "\n",
    "from qwen_tts import Qwen3TTSModel\n",
    "from peft import PeftModel\n",
    "from huggingface_hub import hf_hub_download\n",
    "\n",
    "print('✅ Imports OK')\n",
    "print('CUDA available:', torch.cuda.is_available())\n",
    "if torch.cuda.is_available():\n",
    "    print('Device:', torch.cuda.get_device_name(0))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 3. Configuration — pick your language\n",
    "\n",
    "Edit `LANGUAGE` and `GENDER`. Available combos:\n",
    "\n",
    "| Language  | Code  | Folder name                                  |\n",
    "| --------- | ----- | -------------------------------------------- |\n",
    "| Assamese  | `as`  | `assamese_female` / `assamese_male`          |\n",
    "| Bengali   | `bn`  | `bengali_female` / `bengali_male`            |\n",
    "| Bodo      | `brx` | `bodo_female` / `bodo_male`                  |\n",
    "| Dogri     | `doi` | `dogri_female` / `dogri_male`                |\n",
    "| Gujarati  | `gu`  | `gujarati_female` / `gujarati_male`          |\n",
    "| Kannada   | `kn`  | `kannada_female` / `kannada_male`            |\n",
    "| Konkani   | `kok` | `konkani_female` / `konkani_male`            |\n",
    "| Maithili  | `mai` | `maithili_female` / `maithili_male`          |\n",
    "| Malayalam | `ml`  | `malayalam_female` / `malayalam_male`        |\n",
    "| Marathi   | `mr`  | `marathi_female` / `marathi_male`            |\n",
    "| Nepali    | `ne`  | `nepali_female` / `nepali_male`              |\n",
    "| Odia      | `or`  | `odia_female` / `odia_male`                  |\n",
    "| Punjabi   | `pa`  | `punjabi_female` / `punjabi_male`            |\n",
    "| Tamil     | `ta`  | `tamil_female` / `tamil_male`                |\n",
    "| Telugu    | `te`  | `telugu_female` / `telugu_male`              |"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "✅ Will use adapter: aguken-ai/Qwen3-TTS-0.6B-LoRA-Finetuned-Indic-Multilingual/adapters/telugu_male\n"
     ]
    }
   ],
   "source": [
    "# ─── EDIT THESE ────────────────────────────────────────────\n",
    "LANGUAGE   = 'telugu'      # one of the 15 languages above (lowercase)\n",
    "GENDER     = 'male'      # 'female' or 'male'\n",
    "LORA_SCALE = 0.1          # 0.1-0.2 = subtle, 0.3-0.4 = balanced, 0.5 = strong\n",
    "# ───────────────────────────────────────────────────────────\n",
    "\n",
    "# Adapter source — change this if you fork the repo to your own HF org.\n",
    "BASE_MODEL   = 'Qwen/Qwen3-TTS-12Hz-0.6B-Base'\n",
    "ADAPTER_REPO = 'aguken-ai/Qwen3-TTS-0.6B-LoRA-Finetuned-Indic-Multilingual'\n",
    "\n",
    "SUPPORTED_LANGUAGES = {\n",
    "    'assamese', 'bengali', 'bodo', 'dogri', 'gujarati', 'kannada', 'konkani',\n",
    "    'maithili', 'malayalam', 'marathi', 'nepali', 'odia', 'punjabi', 'tamil', 'telugu',\n",
    "}\n",
    "assert LANGUAGE in SUPPORTED_LANGUAGES, f'LANGUAGE must be one of {sorted(SUPPORTED_LANGUAGES)}'\n",
    "assert GENDER in ('female', 'male'), \"GENDER must be 'female' or 'male'\"\n",
    "\n",
    "ADAPTER_SUB = f'adapters/{LANGUAGE}_{GENDER}'\n",
    "print(f'✅ Will use adapter: {ADAPTER_REPO}/{ADAPTER_SUB}')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 4. Load base model + LoRA adapter\n",
    "\n",
    "First run downloads ~1.8 GB base + ~45 MB adapter (~2 minutes). Subsequent runs use the cache."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "28a2677a9d3a47fe9a6e4908fa4dca14",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "Fetching 4 files:   0%|          | 0/4 [00:00<?, ?it/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "1fecf4bc9fdf4e0cbd9208eebeb5e7dd",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "adapter_config.json: 0.00B [00:00, ?B/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "5cf10d471b2e425e921697e45ea85178",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "adapters/telugu_male/adapter_model.safet(…):   0%|          | 0.00/47.6M [00:00<?, ?B/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "56becbf34922416ab08a87b0d858e63b",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "adapters/telugu_male/ref.wav:   0%|          | 0.00/2.88M [00:00<?, ?B/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "6762d423bdf04a9db52318158839b9bf",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "ref.txt: 0.00B [00:00, ?B/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "✅ Loaded telugu male (LoRA scale 0.1)\n",
      "   Reference clip:  /teamspace/studios/this_studio/.cache/huggingface/hub/models--aguken-ai--Qwen3-TTS-0.6B-LoRA-Finetuned-Indic-Multilingual/snapshots/c3dcce27a6a5a99005680107331a9e97e34be7c2/adapters/telugu_male/ref.wav\n",
      "   Reference text:  'యల్లాప్రగడ సుబ్బారావు ఆయన ఇంటర్నేషనల్ ట్రెడిషనల్ కరాటే ఫెడరేషన్, ఐటికేఎఫ్ను స్థాపించారు. అతను పంతొమ్మిది వందల యాభై ఒకటిలో ప్రారంభమైన ది టైమ్స్ ఆఫ్ ఇండియాలో అతని సృష్టి, ది కామన్ మ్యాన్ మరియు అతని రోజువారీ కార్టూన్ స్ట్రిప్ యు సేడ్ ఇట్కు ప్రసిద్ధి చెందాడు. అంటే ఒమన్ దేశం వైశాల్యం ఆంధ్ర ప్రదేశ్ కంటే ఎక్కువ. ప్రయాణీకులకు ఉచితంగా సహాయం చేయాలనుకున్న ఆ నిజాయితీగల ఆటో డ్రైవర్ని వాళ్ళే దోచుకున్నారంటే, నాకు బాధేసింది. మీ కొత్త నెట్వర్కింగ్ కాంటాక్ట్ను మీరు నెట్వర్కింగ్ ఈవెంట్లో జరిగిన చర్చను కొనసాగించడానికి వారు కాఫీ కోసం కలుస్తారేమో అడగవచ్చు. మీ బాధ్యతలను వదులుకుని ఇక్కడనుంచి వెళ్లిపోవడానికి సిద్ధంగా ఉంటే, ఇక్కడ మీకున్న పని అనుభవానికి అర్ధం లేనట్టే. ఈ అద్దె ఇల్లు ఎంత ఆధునికంగా ఉంది, ఎంత చక్కగా చూస్కున్నారో, భలే బాగుంది! నాకు ఇది చూపించిన ఇళ్ళ బ్రోకర్కి కృతజ్ఞతలు చెప్పాలి. అరణ్యక్ ఎక్స్ప్రెస్'\n"
     ]
    }
   ],
   "source": [
    "device = 'cuda:0' if torch.cuda.is_available() else 'cpu'\n",
    "dtype = torch.bfloat16 if torch.cuda.is_available() else torch.float32\n",
    "\n",
    "# 1) Base model wrapper (Qwen3TTSModel is a wrapper class, not an nn.Module).\n",
    "wrapper = Qwen3TTSModel.from_pretrained(\n",
    "    BASE_MODEL,\n",
    "    device_map=device,\n",
    "    dtype=dtype,\n",
    ")\n",
    "\n",
    "# 2) Attach LoRA to the INNER HF nn.Module (wrapper.model), then put it back.\n",
    "wrapper.model = PeftModel.from_pretrained(\n",
    "    wrapper.model,\n",
    "    ADAPTER_REPO,\n",
    "    subfolder=ADAPTER_SUB,\n",
    ")\n",
    "\n",
    "# 3) Apply LoRA scale (handles both modern + older PEFT versions).\n",
    "def _apply_scale(peft_model, scale):\n",
    "    try:\n",
    "        peft_model.set_adapter_scale(scale)\n",
    "    except AttributeError:\n",
    "        for m in peft_model.modules():\n",
    "            if hasattr(m, 'scaling') and isinstance(m.scaling, dict):\n",
    "                for k in m.scaling:\n",
    "                    m.scaling[k] = scale\n",
    "\n",
    "_apply_scale(wrapper.model, LORA_SCALE)\n",
    "wrapper.model.eval()\n",
    "\n",
    "# 4) Fetch reference audio + transcript from the same adapter folder on HF.\n",
    "def fetch_ref(repo, subfolder):\n",
    "    wav = hf_hub_download(repo_id=repo, filename=f'{subfolder}/ref.wav')\n",
    "    txt_path, last_err = None, None\n",
    "    for name in ('ref.text', 'ref.txt'):\n",
    "        try:\n",
    "            txt_path = hf_hub_download(repo_id=repo, filename=f'{subfolder}/{name}')\n",
    "            break\n",
    "        except Exception as e:\n",
    "            last_err = e\n",
    "    if txt_path is None:\n",
    "        raise FileNotFoundError(f'No ref.text/ref.txt under {subfolder}: {last_err}')\n",
    "    with open(txt_path, 'r', encoding='utf-8') as f:\n",
    "        return wav, f.read().strip()\n",
    "\n",
    "ref_audio, ref_text = fetch_ref(ADAPTER_REPO, ADAPTER_SUB)\n",
    "print(f'✅ Loaded {LANGUAGE} {GENDER} (LoRA scale {LORA_SCALE})')\n",
    "print(f'   Reference clip:  {ref_audio}')\n",
    "print(f'   Reference text:  {ref_text!r}')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 5. Generate speech from text\n",
    "\n",
    "`SAMPLE_TEXTS` has a starter phrase per language so you can test without typing in scripts you don't read. Replace `TEXT` with anything you want — just make sure it's in the appropriate script for the chosen language (Devanagari for Marathi/Nepali/Maithili/Dogri/Bodo/Konkani, Bengali script for Bengali/Assamese, Telugu script for Telugu, etc.)."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import time\n",
    "SAMPLE_TEXTS = {\n",
    "    'assamese':  'নমস্কাৰ! মই Qwen3-TTS।',\n",
    "    'bengali':   'নমস্কার! আমি Qwen3-TTS।',\n",
    "    'bodo':      'खौसे! आं Qwen3-TTS।',\n",
    "    'dogri':     'नमस्ते! मैं Qwen3-TTS आं।',\n",
    "    'gujarati':  'નમસ્કાર! હું Qwen3-TTS છું.',\n",
    "    'kannada':   'ನಮಸ್ಕಾರ! ನಾನು Qwen3-TTS.',\n",
    "    'konkani':   'नमस्कार! हांव Qwen3-TTS.',\n",
    "    'maithili':  'नमस्कार! हम Qwen3-TTS छी।',\n",
    "    'malayalam': 'നമസ്കാരം! ഞാൻ Qwen3-TTS ആണ്.',\n",
    "    'marathi':   'नमस्कार! मी Qwen3-TTS आहे.',\n",
    "    'nepali':    'नमस्ते! म Qwen3-TTS हुँ।',\n",
    "    'odia':      'ନମସ୍କାର! ମୁଁ Qwen3-TTS।',\n",
    "    'punjabi':   'ਸਤ ਸ੍ਰੀ ਅਕਾਲ! ਮੈਂ Qwen3-TTS ਹਾਂ।',\n",
    "    'tamil':     'வணக்கம்! நான் Qwen3-TTS.',\n",
    "    'telugu':    'నమస్కారం! నేను Qwen3-TTS.',\n",
    "}\n",
    "\n",
    "# ─── EDIT THIS ─────────────────────────────────────────────\n",
    "TEXT = SAMPLE_TEXTS[LANGUAGE]              # or replace with your own string\n",
    "OUTPUT_PATH = f'output_{LANGUAGE}_{GENDER}.wav'\n",
    "# ───────────────────────────────────────────────────────────\n",
    "ref_text =\"నమస్కారం అందరికీ. నా పేరు **సాయి కృష్ణ**. నేను **హైదరాబాద్** నుండి వచ్చాను. ప్రస్తుతం **ఆగుకెన్ ఏఐ**లో పని చేస్తున్నాను. నాకు కొత్త విషయాలు నేర్చుకోవడం, టెక్నాలజీ గురించి తెలుసుకోవడం, అలాగే నా నైపుణ్యాలను అభివృద్ధి చేసుకోవడం చాలా ఇష్టం. ఈ అవకాశాన్ని ఇచ్చినందుకు ధన్యవాదాలు. మీ అందరితో కలిసి నేర్చుకోవడానికి మరియు నా అనుభవాన్ని పంచుకోవడానికి ఎదురుచూస్తున్నాను. ధన్యవాదాలు!\"\n",
    "print(\"starting generation...\", flush=True)\n",
    "t0 = time.time()\n",
    "with torch.no_grad():\n",
    "    wavs, sr = wrapper.generate_voice_clone(\n",
    "        text=TEXT,\n",
    "        language='auto',           # wrapper rejects Indic names; 'auto' works\n",
    "        ref_audio=\"telugu_male/ref.wav\",\n",
    "        ref_text=ref_text,\n",
    "    )\n",
    "print(f\"generation done in {time.time() - t0:.1f}s\", flush=True)\n",
    "sf.write(OUTPUT_PATH, wavs[0], sr)\n",
    "print(f'✅ Wrote {OUTPUT_PATH} ({len(wavs[0])/sr:.1f}s @ {sr} Hz)')\n",
    "ipd.display(ipd.Audio(OUTPUT_PATH))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 6. (Optional) Switch language without reloading the base model\n",
    "\n",
    "The slow part is the 1.8 GB base download. Once that's in memory you can hot-swap LoRA adapters in seconds. Call `swap_adapter(language, gender)` and then re-run the generate cell."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def swap_adapter(language, gender, lora_scale=0.35):\n",
    "    '''Swap LoRA adapter on the already-loaded wrapper. No base re-download.'''\n",
    "    global wrapper, ref_audio, ref_text, LANGUAGE, GENDER, LORA_SCALE, ADAPTER_SUB\n",
    "\n",
    "    assert language in SUPPORTED_LANGUAGES, f'Bad language: {language}'\n",
    "    assert gender in ('female', 'male'), 'Gender must be female/male'\n",
    "\n",
    "    LANGUAGE, GENDER, LORA_SCALE = language, gender, lora_scale\n",
    "    ADAPTER_SUB = f'adapters/{language}_{gender}'\n",
    "\n",
    "    # Strip the current PEFT layers, return to the plain HF model.\n",
    "    if hasattr(wrapper.model, 'unload'):\n",
    "        wrapper.model = wrapper.model.unload()\n",
    "    elif hasattr(wrapper.model, 'base_model'):\n",
    "        wrapper.model = wrapper.model.base_model.model\n",
    "\n",
    "    # Attach the new adapter.\n",
    "    wrapper.model = PeftModel.from_pretrained(\n",
    "        wrapper.model, ADAPTER_REPO, subfolder=ADAPTER_SUB,\n",
    "    )\n",
    "    _apply_scale(wrapper.model, lora_scale)\n",
    "    wrapper.model.eval()\n",
    "\n",
    "    ref_audio, ref_text = fetch_ref(ADAPTER_REPO, ADAPTER_SUB)\n",
    "    print(f'✅ Switched to {language} {gender} (LoRA scale {lora_scale})')\n",
    "\n",
    "# Example — uncomment to try:\n",
    "# swap_adapter('tamil', 'male', lora_scale=0.35)\n",
    "# with torch.no_grad():\n",
    "#     wavs, sr = wrapper.generate_voice_clone(\n",
    "#         text=SAMPLE_TEXTS[LANGUAGE], language='auto',\n",
    "#         ref_audio=ref_audio, ref_text=ref_text,\n",
    "#     )\n",
    "# sf.write(f'output_{LANGUAGE}_{GENDER}.wav', wavs[0], sr)\n",
    "# ipd.display(ipd.Audio(f'output_{LANGUAGE}_{GENDER}.wav'))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "ename": "NameError",
     "evalue": "name 'wrapper' is not defined",
     "output_type": "error",
     "traceback": [
      "\u001b[0;31m---------------------------------------------------------------------------\u001b[0m",
      "\u001b[0;31mNameError\u001b[0m                                 Traceback (most recent call last)",
      "Cell \u001b[0;32mIn[5], line 11\u001b[0m\n\u001b[1;32m      8\u001b[0m OUTPUT_PATH \u001b[38;5;241m=\u001b[39m \u001b[38;5;124mf\u001b[39m\u001b[38;5;124m\"\u001b[39m\u001b[38;5;124moutput_\u001b[39m\u001b[38;5;132;01m{\u001b[39;00mLANGUAGE\u001b[38;5;132;01m}\u001b[39;00m\u001b[38;5;124m_\u001b[39m\u001b[38;5;132;01m{\u001b[39;00mGENDER\u001b[38;5;132;01m}\u001b[39;00m\u001b[38;5;124m.wav\u001b[39m\u001b[38;5;124m\"\u001b[39m\n\u001b[1;32m     10\u001b[0m \u001b[38;5;28;01mwith\u001b[39;00m torch\u001b[38;5;241m.\u001b[39mno_grad():\n\u001b[0;32m---> 11\u001b[0m     wavs, sr \u001b[38;5;241m=\u001b[39m \u001b[43mwrapper\u001b[49m\u001b[38;5;241m.\u001b[39mgenerate_voice_clone(\n\u001b[1;32m     12\u001b[0m         text\u001b[38;5;241m=\u001b[39mTEXT,\n\u001b[1;32m     13\u001b[0m         language\u001b[38;5;241m=\u001b[39m\u001b[38;5;124m\"\u001b[39m\u001b[38;5;124mauto\u001b[39m\u001b[38;5;124m\"\u001b[39m,\n\u001b[1;32m     14\u001b[0m         ref_audio\u001b[38;5;241m=\u001b[39mref_audio,\n\u001b[1;32m     15\u001b[0m         ref_text\u001b[38;5;241m=\u001b[39mref_text,\n\u001b[1;32m     16\u001b[0m     )\n\u001b[1;32m     18\u001b[0m sf\u001b[38;5;241m.\u001b[39mwrite(OUTPUT_PATH, wavs[\u001b[38;5;241m0\u001b[39m], sr)\n\u001b[1;32m     19\u001b[0m \u001b[38;5;28mprint\u001b[39m(\u001b[38;5;124mf\u001b[39m\u001b[38;5;124m\"\u001b[39m\u001b[38;5;124m✅ Wrote \u001b[39m\u001b[38;5;132;01m{\u001b[39;00mOUTPUT_PATH\u001b[38;5;132;01m}\u001b[39;00m\u001b[38;5;124m (\u001b[39m\u001b[38;5;132;01m{\u001b[39;00m\u001b[38;5;28mlen\u001b[39m(wavs[\u001b[38;5;241m0\u001b[39m])\u001b[38;5;250m \u001b[39m\u001b[38;5;241m/\u001b[39m\u001b[38;5;250m \u001b[39msr\u001b[38;5;132;01m:\u001b[39;00m\u001b[38;5;124m.1f\u001b[39m\u001b[38;5;132;01m}\u001b[39;00m\u001b[38;5;124ms @ \u001b[39m\u001b[38;5;132;01m{\u001b[39;00msr\u001b[38;5;132;01m}\u001b[39;00m\u001b[38;5;124m Hz)\u001b[39m\u001b[38;5;124m\"\u001b[39m)\n",
      "\u001b[0;31mNameError\u001b[0m: name 'wrapper' is not defined"
     ]
    }
   ],
   "source": [
    "REF_DIR = \"telugu_male\"\n",
    "ref_audio = f\"{REF_DIR}/ref.wav\"\n",
    "\n",
    "with open(f\"{REF_DIR}/ref.txt\", \"r\", encoding=\"utf-8\") as f:\n",
    "    ref_text = f.read().strip()\n",
    "\n",
    "TEXT = \"నమస్కారం అందరికీ. నా పేరు **సాయి కృష్ణ**. నేను **హైదరాబాద్** నుండి వచ్చాను. ప్రస్తుతం **ఆగుకెన్ ఏఐ**లో పని చేస్తున్నాను. నాకు కొత్త విషయాలు నేర్చుకోవడం, టెక్నాలజీ గురించి తెలుసుకోవడం, అలాగే నా నైపుణ్యాలను అభివృద్ధి చేసుకోవడం చాలా ఇష్టం. ఈ అవకాశాన్ని ఇచ్చినందుకు ధన్యవాదాలు. మీ అందరితో కలిసి నేర్చుకోవడానికి మరియు నా అనుభవాన్ని పంచుకోవడానికి ఎదురుచూస్తున్నాను. ధన్యవాదాలు!\"\n",
    "OUTPUT_PATH = f\"output_{LANGUAGE}_{GENDER}.wav\"\n",
    "\n",
    "with torch.no_grad():\n",
    "    wavs, sr = wrapper.generate_voice_clone(\n",
    "        text=TEXT,\n",
    "        language=\"auto\",\n",
    "        ref_audio=ref_audio,\n",
    "        ref_text=ref_text,\n",
    "    )\n",
    "\n",
    "sf.write(OUTPUT_PATH, wavs[0], sr)\n",
    "print(f\"✅ Wrote {OUTPUT_PATH} ({len(wavs[0]) / sr:.1f}s @ {sr} Hz)\")\n",
    "ipd.display(ipd.Audio(OUTPUT_PATH))"
   ]
  }
 ],
 "metadata": {
  "accelerator": "GPU",
  "colab": {
   "gpuType": "T4",
   "machine_shape": "hm",
   "provenance": []
  },
  "kernelspec": {
   "display_name": "cloudspace",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.12.11"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 0
}