notebook.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "6f934064",
   "metadata": {},
   "source": [
    "# CIIL Anuvani-LID: quick start\n",
    "\n",
    "Anuvani-LID tells you **which language is spoken** in a recording, for the 22 scheduled languages of India and English.\n",
    "\n",
    "**How to use this notebook:** run the cells from top to bottom (click a cell, then press **Shift + Enter**), or choose **Runtime → Run all**. A CPU runtime is enough; a GPU is not needed.\n",
    "\n",
    "Every code cell below is copied from the [model card](https://huggingface.co/CIIL/Anuvani-LID); the short cells between them only prepare the audio files the card's examples use."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "6793c3e4",
   "metadata": {},
   "source": [
    "## 1. Install\n",
    "\n",
    "Google Colab already has PyTorch, Transformers, soundfile and librosa. This adds the two packages the long-audio and ONNX examples need. (Outside Colab: `pip install transformers torch soundfile librosa silero-vad onnxruntime`.)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "40ffaedb",
   "metadata": {},
   "outputs": [],
   "source": [
    "%pip install -q silero-vad onnxruntime"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "3568f6da",
   "metadata": {},
   "source": [
    "## 2. Get a test recording\n",
    "\n",
    "The repository ships five short clips (FLEURS, CC BY 4.0). This saves the Hindi one as `audio.wav`, the file name the card's examples use."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "295fec36",
   "metadata": {},
   "outputs": [],
   "source": [
    "import soundfile as sf\n",
    "from huggingface_hub import hf_hub_download\n",
    "\n",
    "wav, sr = sf.read(hf_hub_download(\"CIIL/Anuvani-LID\", \"samples/hi_in.wav\"))\n",
    "sf.write(\"audio.wav\", wav, sr)\n",
    "print(f\"audio.wav: {len(wav) / sr:.1f} s at {sr} Hz\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "5b07fa57",
   "metadata": {},
   "source": [
    "## 3. Identify the language with `pipeline`\n",
    "\n",
    "The first run downloads the model (about 465 MB). The result lists the most likely languages with their scores."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "fdc35307",
   "metadata": {},
   "outputs": [],
   "source": [
    "from transformers import pipeline\n",
    "repo = \"CIIL/Anuvani-LID\"\n",
    "clf = pipeline(\"audio-classification\", model=repo, feature_extractor=repo, trust_remote_code=True)\n",
    "# [{'label': 'hindi', 'score': 0.98}, ...]\n",
    "print(clf(\"audio.wav\")[:3])"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "5d34898a",
   "metadata": {},
   "source": [
    "The same on the repository's sample clips:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "84b9918f",
   "metadata": {},
   "outputs": [],
   "source": [
    "from huggingface_hub import hf_hub_download\n",
    "print(clf(hf_hub_download(repo, \"samples/hi_in.wav\"))[:3])"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "bb4489ba",
   "metadata": {},
   "source": [
    "## 4. The same with `AutoModel`\n",
    "\n",
    "Useful when you want the raw scores (logits) yourself."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "71559636",
   "metadata": {},
   "outputs": [],
   "source": [
    "import soundfile as sf, torch\n",
    "from transformers import AutoModelForAudioClassification, AutoFeatureExtractor\n",
    "repo = \"CIIL/Anuvani-LID\"\n",
    "# native Wav2Vec2FeatureExtractor\n",
    "fe = AutoFeatureExtractor.from_pretrained(repo)\n",
    "model = AutoModelForAudioClassification.from_pretrained(repo, trust_remote_code=True).eval()\n",
    "# 16 kHz mono\n",
    "wav, sr = sf.read(\"audio.wav\")\n",
    "logits = model(**fe(wav, sampling_rate=sr, return_tensors=\"pt\")).logits\n",
    "print(model.config.id2label[logits.argmax(-1).item()])"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "34e729fa",
   "metadata": {},
   "source": [
    "## 5. Long audio: a language timeline\n",
    "\n",
    "For recordings longer than about two minutes, `classify_long_audio` finds the speech with Silero VAD, classifies each phrase and reports a timeline. To show it, this joins the five sample clips (Hindi, Bengali, Marathi, Tamil, Telugu) into one recording, `lecture_1hour.mp3`, the file name the card's example uses."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "eb8dad25",
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np, soundfile as sf\n",
    "from huggingface_hub import hf_hub_download\n",
    "\n",
    "pause = np.zeros(8000)                                              # 0.5 s of silence between clips\n",
    "parts = [sf.read(hf_hub_download(\"CIIL/Anuvani-LID\", f\"samples/{c}.wav\"))[0]\n",
    "         for c in [\"hi_in\", \"bn_in\", \"mr_in\", \"ta_in\", \"te_in\"]]\n",
    "sf.write(\"lecture_1hour.mp3\", np.concatenate([np.concatenate([p, pause]) for p in parts]), 16000)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ec7f8eb0",
   "metadata": {},
   "outputs": [],
   "source": [
    "import librosa\n",
    "# any format/length\n",
    "wav, _ = librosa.load(\"lecture_1hour.mp3\", sr=16000, mono=True)\n",
    "# needs: pip install silero-vad\n",
    "out = model.classify_long_audio(wav)\n",
    "# overall (duration-weighted over the speech pieces)\n",
    "print(out[\"label\"], out[\"score\"])\n",
    "# share of speech time per language\n",
    "print(out[\"dominant_by_time\"], out[\"language_time_share\"])\n",
    "# per phrase piece — surfaces code-switching\n",
    "for p in out[\"timeline\"]:\n",
    "    print(f\"{p['start']:7.1f}-{p['end']:7.1f}s  {p['label']}  {p['score']:.3f}\")"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "13d597cb",
   "metadata": {},
   "source": [
    "Each line of the timeline is one stretch of speech and its language, so a recording that switches language shows every language in turn."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "dde8ce79",
   "metadata": {},
   "source": [
    "## 6. ONNX (optional)\n",
    "\n",
    "The most portable way to run the model, without PyTorch or Transformers. It downloads a further 464 MB."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "bbe6efca",
   "metadata": {},
   "outputs": [],
   "source": [
    "import onnxruntime as ort, soundfile as sf, numpy as np\n",
    "from huggingface_hub import hf_hub_download\n",
    "repo = \"CIIL/Anuvani-LID\"\n",
    "sess = ort.InferenceSession(hf_hub_download(repo, \"triton_inference/assets/model/anuvani_lid.onnx\"))\n",
    "labels = open(hf_hub_download(repo, \"labels.txt\")).read().split()\n",
    "# 16 kHz mono\n",
    "wav, _ = sf.read(\"audio.wav\", dtype=\"float32\")\n",
    "logits = sess.run([\"logits\"], {\"audio\": wav[None], \"audio_length\": np.array([len(wav)], np.int64)})[0]\n",
    "print(labels[int(logits.argmax())])"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "1faaac1c",
   "metadata": {},
   "source": [
    "## 7. Try your own recording\n",
    "\n",
    "Run the cell below and choose a file (WAV, MP3, M4A, OGG, ...). It runs last so that **Run all** finishes the examples above first."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "b520a373",
   "metadata": {},
   "outputs": [],
   "source": [
    "try:\n",
    "    from google.colab import files\n",
    "    for name in files.upload():\n",
    "        print(name, clf(name)[:3])\n",
    "except ImportError:\n",
    "    print(\"Not running in Google Colab: call clf(\\\"path/to/your_audio.wav\\\") instead.\")"
   ]
  }
 ],
 "metadata": {
  "colab": {
   "name": "CIIL Anuvani-LID quick start",
   "provenance": []
  },
  "kernelspec": {
   "display_name": "Python 3",
   "name": "python3"
  },
  "language_info": {
   "name": "python"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}