{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "be1c7379",
   "metadata": {},
   "outputs": [],
   "source": [
    "!pip install --quiet --root-user-action=ignore --upgrade pip\n",
    "!pip install --quiet --root-user-action=ignore \"datasets>=1.18.3\" \"transformers==4.11.3\" librosa jiwer huggingface_hub  \n",
    "!pip install --quiet --root-user-action=ignore https://github.com/kpu/kenlm/archive/master.zip pyctcdecode\n",
    "!pip install --quiet --root-user-action=ignore --upgrade transformers\n",
    "!pip install --quiet --root-user-action=ignore torch_audiomentations audiomentations  "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "8892305a",
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Reusing dataset common_voice (/ai_data/cache/common_voice/de/6.1.0/a1dc74461f6c839bfe1e8cf1262fd4cf24297e3fbd4087a711bd090779023a5e)\n",
      "Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.\n"
     ]
    },
    {
     "data": {
      "application/vnd.jupyter.widget-view+json": {
       "model_id": "efc316d6eedb4dcab341dfd0fe8cc926",
       "version_major": 2,
       "version_minor": 0
      },
      "text/plain": [
       "  0%|          | 0/15588 [00:00<?, ?ex/s]"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "from datasets import load_dataset, Audio, load_metric\n",
    "from transformers import AutoModelForCTC, Wav2Vec2ProcessorWithLM\n",
    "import torchaudio.transforms as T\n",
    "import torch\n",
    "import unicodedata\n",
    "import numpy as np\n",
    "import re\n",
    "\n",
    "# load testing dataset \n",
    "testing_dataset = load_dataset(\"common_voice\", \"de\", split=\"test\")\n",
    "\n",
    "# replace invisible characters with space\n",
    "allchars = list(set([c for t in testing_dataset['sentence'] for c in list(t)]))\n",
    "map_to_space = [c for c in allchars if unicodedata.category(c)[0] in 'PSZ' and c not in 'ʻ-']\n",
    "replacements = ''.maketrans(''.join(map_to_space), ''.join(' ' for i in range(len(map_to_space))), '\\'ʻ')\n",
    "\n",
    "def text_fix(text):\n",
    "    # change ß to ss\n",
    "    text = text.replace('ß','ss')\n",
    "    # convert dash to space and remove double-space\n",
    "    text = text.replace('-',' ').replace('  ',' ').replace('  ',' ')\n",
    "    # make lowercase\n",
    "    text = text.lower()\n",
    "    # remap all invisible characters to space\n",
    "    text = text.translate(replacements).strip()\n",
    "    # for easier comparison to Zimmermeister, replace unrepresentable characters with ?\n",
    "    text = re.sub(\"[âşěýňעảנźțãòàǔł̇æồאắîשðșęūāñë生בøúıśžçćńřğ]+\",\"?\",text)\n",
    "    # remove multiple spaces (again)\n",
    "    text = ' '.join([w for w in text.split(' ') if w != ''])\n",
    "    return text\n",
    "\n",
    "# load model\n",
    "model = AutoModelForCTC.from_pretrained(\"fxtentacle/wav2vec2-xls-r-1b-tevr\")\n",
    "model.to('cuda')\n",
    "# load processor\n",
    "class HajoProcessor(Wav2Vec2ProcessorWithLM):\n",
    "    @staticmethod\n",
    "    def get_missing_alphabet_tokens(decoder, tokenizer):\n",
    "        return []\n",
    "processor = HajoProcessor.from_pretrained(\"fxtentacle/wav2vec2-xls-r-1b-tevr\")\n",
    "\n",
    "# this function will be called for each WAV file\n",
    "def predict_single_audio(batch, image=False):    \n",
    "    audio = batch['audio']['array']\n",
    "    # resample, if needed\n",
    "    if batch['audio']['sampling_rate'] != 16000:\n",
    "        audio = T.Resample(orig_freq=batch['audio']['sampling_rate'], new_freq=16000)(torch.from_numpy(audio)).numpy()\n",
    "    # normalize\n",
    "    audio = (audio - audio.mean()) / np.sqrt(audio.var() + 1e-7)\n",
    "    # ask HF processor to prepare audio for GPU eval\n",
    "    input_values = processor(audio, return_tensors=\"pt\", sampling_rate=16_000).input_values\n",
    "    # call model on GPU\n",
    "    with torch.no_grad():\n",
    "        logits = model(input_values.to('cuda')).logits.cpu().numpy()[0]\n",
    "    # ask HF processor to decode logits\n",
    "    decoded = processor.decode(logits, beam_width=500)\n",
    "    # return as dictionary\n",
    "    return { 'groundtruth': text_fix(batch['sentence']), 'prediction': decoded.text }\n",
    "\n",
    "# process all audio files\n",
    "all_predictions = testing_dataset.map(predict_single_audio, remove_columns=testing_dataset.column_names)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "id": "38f5481f",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "0 PRED:  mückenstiche sollte man nicht aufkratzen\n",
      "0   GT:  mückenstiche sollte man nicht aufkratzen\n",
      "1 PRED:  ist diese leitung sicher\n",
      "1   GT:  ist diese leitung sicher\n",
      "2 PRED:  die ratten verlassen das sinkende schiff\n",
      "2   GT:  die ratten verlassen das sinkende schiff\n",
      "3 PRED:  ich habe eine neue arbeit\n",
      "3   GT:  ich habe eine neue arbeit\n",
      "4 PRED:  was sieht kamera eins gerade\n",
      "4   GT:  was sieht kamera eins gerade\n",
      "5 PRED:  was für ein angeber dachte horst im stillen\n",
      "5   GT:  was für ein angeber dachte horst im stillen\n",
      "6 PRED:  rückgängig machen\n",
      "6   GT:  rückgängig machen\n",
      "7 PRED:  war die integration erfolgreich\n",
      "7   GT:  war die integration erfolgreich\n"
     ]
    }
   ],
   "source": [
    "# log example results\n",
    "for i in range(8):\n",
    "    print(i,'PRED: ',all_predictions[i]['prediction'])\n",
    "    print(i,'  GT: ',all_predictions[i]['groundtruth'])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "cbabe801",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "WER 3.6433399042523233 %\n",
      "CER 1.5398893560981173 %\n"
     ]
    }
   ],
   "source": [
    "# print results\n",
    "print('WER', load_metric(\"wer\").compute(predictions=all_predictions['prediction'], references=all_predictions['groundtruth'])*100.0, '%')\n",
    "print('CER', load_metric(\"cer\").compute(predictions=all_predictions['prediction'], references=all_predictions['groundtruth'])*100.0, '%')"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "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": "3.7.5"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}