Upload download_datasets_in_wav_or_mp3_and_create_csv.ipynb

download_datasets_in_wav_or_mp3_and_create_csv.ipynb

@@ -7,39 +7,43 @@
    "outputs": [],
    "source": [
     "from datasets import load_dataset\n",
-    "import soundfile as sf, os, pandas as pd, re\n",
+    "import soundfile as sf, os, re, neologdn\n",
     "from tqdm import tqdm\n",
     "\n",
-    "dataset = load_dataset(\"mozilla-foundation/common_voice_17_0\", \"ja\", split=\"other\", trust_remote_code=True, streaming=True, token=\"\")\n",
+    "max = 20.0\n",
+    "min = 1.0\n",
     "\n",
-    "name = \"other\"\n",
-    "ouput_dir = \"./datasets/CV17/\"\n",
-    "output_file = 'metadata.csv'\n",
+    "dataset = load_dataset(\"Sin2pi/JA_audio_JA_text_180k_samples\", split=\"train\", trust_remote_code=True, streaming=True)\n",
+    "\n",
+    "name = \"gvs\"\n",
+    "ouput_dir = \"./datasets/\"\n",
+    "output_file = 'metadata.csv' # create metadata file with file names and transcripts\n",
     "os.makedirs(ouput_dir + name, exist_ok=True)\n",
     "folder_path = ouput_dir + name # Create a folder to store the audio and transcription files\n",
     "\n",
-    "char = '[　０１２３４５６７８９ａｂｃｄｅｆｇｈｉｊｋｌｍｎｏｐｑｒｓｔｕｖｗｘｙｚＡＢＣＤＥＦＧＨＩＪＫＬＭＮＯＰＱＲＳＴＵＶＷＸＹＺabcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ1234567890]'\n",
-    "special_characters = '[♬「」？！“%‘”~♪…?!゛＃＄％＆（）＊＋：；〈＝〉＠＾＿｛｜｝～\"█♩♫』『.;:<>_()*&^$#@`, ]'\n",
-    "\n",
-    "dsa = (dataset\n",
-    "       .filter(lambda sample: bool(sample[\"sentence\"])) # Returns only samples with transcriptions\n",
-    "       .filter(lambda sample: not re.search(char, sample[\"sentence\"])) # Returns only samples without latin/roman alpha/numerics\n",
-    "       .filter(lambda sample: sample[\"down_votes\"] == 0) # Returns only samples with 0 downvotes (commonvoice)\n",
-    "       )\n",
+    "char = '[　０１２３４５６７８９ａｂｃｄｅｆｇｈｉｊｋｌｍｎｏｐｑｒｓｔｕｖｗｘｙｚＡＢＣＤＥＦＧＨＩＪＫＬＭＮＯＰＱＲＳＴＵＶＷＸＹＺabcdefghijklmnopqrstuvwxyzABCDEFGHIJKLMNOPQRSTUVWXYZ1234567890♬♪♩♫]'\n",
+    "special_characters = '[“%‘”~゛＃＄％＆（）＊＋：；〈＝〉＠＾＿｛｜｝～\"█』『.;:<>_()*&^$#@`, ]' #「」\n",
     "\n",
-    "for i, sample in tqdm(enumerate(dsa)): # Process each sample in the filtered dataset\n",
-    "    audio_sample = name + f'_{i}.wav' # or wav\n",
+    "for i, sample in tqdm(enumerate(dataset)): # Process each sample in the filtered dataset\n",
+    "    audio_sample = name + f'_{i}.mp3' # or wav\n",
     "    audio_path = os.path.join(folder_path, audio_sample)\n",
-    "    transcription_path = os.path.join(folder_path, output_file)  # Path to save transcription file \n",
+    "    transcription_path = os.path.join(folder_path, out_file)  # Path to save transcription file \n",
     "    if not os.path.exists(audio_path):\n",
-    "        sf.write(audio_path, sample['audio']['array'], sample['audio']['sampling_rate'])\n",
+    "        patterns = [(r\"…\",'。'), (r\"!!\",'!'), (special_characters,\"\"), (r\"\\s+\", \"\")] # (r\"(.)\\1{2}\")\n",
+    "        for pattern, replace in patterns:\n",
+    "            sample[\"sentence\"] = re.sub(pattern, replace, sample[\"sentence\"])\n",
+    "        sample[\"sentence\"] = (neologdn.normalize(sample[\"sentence\"], repeat=1)) # for Japanese only, repeat number reduces repeat characters\n",
+    "        sample[\"sentence_length\"] = len(sample[\"sentence\"]) # Get sentence lengths  \n",
     "        sample[\"audio_length\"] = len(sample[\"audio\"][\"array\"]) / sample[\"audio\"][\"sampling_rate\"]  # Get audio length, remove if not needed\n",
-    "        with open(transcription_path, 'a', encoding='utf-8') as transcription_file: # Save transcription file    \n",
-    "            transcription_file.write(audio_sample+\",\") # Save transcription file name  \n",
-    "            sample[\"sentence\"] = re.sub(special_characters,'', sample[\"sentence\"])\n",
-    "            transcription_file.write(sample['sentence']) # Save transcription \n",
-    "            transcription_file.write(str(\",\"+str(sample['audio_length']))) # Save audio length, remove if not needed\n",
-    "            transcription_file.write('\\n') "
+    "        if bool(sample[\"sentence\"]) and max > sample[\"audio_length\"] > min and not re.search(char, sample[\"sentence\"]) and sample[\"sentence_length\"] > min_char:\n",
+    "            sf.write(audio_path, sample['audio']['array'], sample['audio']['sampling_rate']) # Get files      \n",
+    "            # process_directory(folder_path, (folder_path + \"/trimmed/\")) # for use with audio sample silence removal script\n",
+    "            if os.path.isfile(audio_path):\n",
+    "                os.remove(audio_path)\n",
+    "            with open(transcription_path, 'a', encoding='utf-8') as transcription_file:\n",
+    "                transcription_file.write(audio_sample+\",\") # Save transcription file name  \n",
+    "                transcription_file.write(sample['sentence']) # Save transcription \n",
+    "                transcription_file.write('\\n')"
    ]
   }
  ],