{
 "nbformat": 4,
 "nbformat_minor": 5,
 "metadata": {
  "colab": {
   "provenance": [],
   "gpuType": "T4"
  },
  "kernelspec": {
   "name": "python3",
   "display_name": "Python 3"
  },
  "language_info": {
   "name": "python"
  },
  "accelerator": "GPU"
 },
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# PakUrdu-Conversational-ASR \u2014 full-quality inference\n",
    "\n",
    "Run the **full r14 model** (merged adapter, fp16) on a **free GPU** \u2014 nothing installs on your machine.\n",
    "\n",
    "**Setup (one click):** enable a GPU accelerator before running:\n",
    "- **Colab:** `Runtime \u2192 Change runtime type \u2192 T4 GPU`\n",
    "- **Kaggle:** `Settings \u2192 Accelerator \u2192 GPU T4 x2` (or P100)\n",
    "\n",
    "Then upload an Urdu audio clip (wav/mp3/m4a\u2026) and run the cells top to bottom.\n",
    "\n",
    "More: [model card](https://huggingface.co/KhiredNetworks/PakUrdu-Conversational-ASR) \u00b7 ",
    "[browser demo](https://huggingface.co/spaces/KhiredNetworks/PakUrdu-ASR-Demo) (runs on your device, no GPU needed) \u00b7 ",
    "[proof notebook](https://www.kaggle.com/code/adeelejaz8611/e1-sr-proof-r14-sr133) (re-derives the certified 7.96% WER end-to-end)\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 1 \u2014 Load the model (base + fine-tuned adapter, merged)\n",
    "\n",
    "Downloads ~1.6 GB once per session; takes a couple of minutes."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# One-time environment fix \u2014 some cloud images ship torchao 0.10, which peft's\n",
    "# import check rejects. torchao is only needed for torchao-quantized adapters\n",
    "# (not this one), so removing it is safe and instant.\n",
    "%pip uninstall -y -q torchao\n",
    "print(\"environment ok \u2014 continue below\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import torch, torchaudio\n",
    "from peft import PeftModel\n",
    "from transformers import WhisperForConditionalGeneration, WhisperProcessor\n",
    "\n",
    "BASE = \"openai/whisper-large-v3-turbo\"\n",
    "ADAPTER = \"KhiredNetworks/PakUrdu-Conversational-ASR\"\n",
    "\n",
    "processor = WhisperProcessor.from_pretrained(BASE)\n",
    "model = WhisperForConditionalGeneration.from_pretrained(BASE, torch_dtype=torch.float16)\n",
    "model = PeftModel.from_pretrained(model, ADAPTER).merge_and_unload().to(\"cuda\").eval()\n",
    "print(\"model ready \u2014 full fp16 r14 (merged adapter)\")"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 2 \u2014 Point it at your audio\n",
    "\n",
    "How to get your clip in:\n",
    "- **Colab:** folder icon (left) \u2192 Upload, or run `from google.colab import files; files.upload()`\n",
    "- **Kaggle:** `+ Add Input \u2192 Upload` (attach the clip as a dataset); the path looks like `/kaggle/input/<your-dataset>/clip.wav`\n",
    "\n",
    "Then set the path below."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "AUDIO_PATH = \"your_audio.wav\"   # <- set this to your clip's path"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 3 \u2014 Transcribe\n",
    "Greedy decode, `language=ur` (the intended configuration). Clips longer than 30 s are\n",
    "windowed (25 s windows, 1.5 s overlap) and stitched, so long audio works too."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "SR = 16000\n",
    "\n",
    "def merge_texts(parts):\n",
    "    # drop the duplicated overlap words at each window seam\n",
    "    out = []\n",
    "    for p in parts:\n",
    "        words = (p or \"\").strip().split()\n",
    "        drop = 0\n",
    "        for k in range(min(len(out), len(words), 12), 0, -1):\n",
    "            if out[len(out) - k:] == words[:k]:\n",
    "                drop = k\n",
    "                break\n",
    "        out.extend(words[drop:])\n",
    "    return \" \".join(out)\n",
    "\n",
    "@torch.no_grad()\n",
    "def transcribe(path):\n",
    "    speech, sr = torchaudio.load(path)          # (channels, samples)\n",
    "    speech = speech.mean(dim=0)\n",
    "    if sr != SR:\n",
    "        speech = torchaudio.functional.resample(speech, sr, SR)\n",
    "    texts, start, CHUNK, OVERLAP = [], 0, 25 * SR, int(1.5 * SR)\n",
    "    while start < len(speech):\n",
    "        end = min(start + CHUNK, len(speech))\n",
    "        feats = processor(speech[start:end].numpy(), sampling_rate=SR,\n",
    "                          return_tensors=\"pt\").input_features.half().to(\"cuda\")\n",
    "        ids = model.generate(feats, language=\"ur\", task=\"transcribe\", do_sample=False)\n",
    "        texts.append(processor.batch_decode(ids, skip_special_tokens=True)[0].strip())\n",
    "        if end == len(speech):\n",
    "            break\n",
    "        start = end - OVERLAP\n",
    "    return merge_texts(texts)\n",
    "\n",
    "print(transcribe(AUDIO_PATH))"
   ]
  }
 ]
}