metadata

dataset_info:
  - config_name: clean
    features:
      - name: audio
        dtype:
          audio:
            sampling_rate: 16000
      - name: text
        dtype: string
      - name: id
        dtype: string
      - name: session_id
        dtype: string
    splits:
      - name: train
        num_bytes: 2807175314
        num_examples: 39989
      - name: validation
        num_bytes: 256038049
        num_examples: 4161
      - name: test
        num_bytes: 253226827
        num_examples: 3813
    download_size: 3311132741
    dataset_size: 3316440190
  - config_name: default
    features:
      - name: audio
        dtype:
          audio:
            sampling_rate: 16000
      - name: text
        dtype: string
      - name: id
        dtype: string
      - name: session_id
        dtype: string
    splits:
      - name: train
        num_bytes: 3341105554.6299996
        num_examples: 46583
    download_size: 3346820592
    dataset_size: 3341105554.6299996
configs:
  - config_name: clean
    data_files:
      - split: train
        path: clean/train-*
      - split: validation
        path: clean/validation-*
      - split: test
        path: clean/test-*
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Dataset Card for code-switching yodas

This dataset is derived from espnet/yodas, more details can be found here: https://huggingface.co/datasets/espnet/yodas

This is a subset of the zh000 subset of espnet/yodas dataset, which selects videos with Mandarin-English code-switching phenomenon.

Note that code-switching is only gauranteed per video rather than per utterance. Therefore, not every utterance in the dataset contains code-switching.

Dataset Details

Dataset Usage

The default config does not modify any text of the selected samples.

from datasets import load_dataset
cs_yodas = load_dataset("georgechang8/code_switch_yodas_zh")

The clean config cleanses the text of the selected samples (as in the processing).

from datasets import load_dataset
cs_yodas_clean = load_dataset("georgechang8/code_switch_yodas_zh", "clean")

{'audio': {'path': 'GaUSbuZm5Ec-00207-00083809-00084143.wav',
  'array': array([-0.09082031,  0.01898193,  0.02850342, ...,  0.01419067,
          0.01391602,  0.01513672]),
  'sampling_rate': 16000},
 'text': '項明生，訂Agoda的項明生',
 'id': 'GaUSbuZm5Ec-00207-00083809-00084143',
 'session_id': 'GaUSbuZm5Ec'}

Dataset Description

Language(s): Chinese, English
License: CC-BY-3.0

Dataset Sources [optional]

Repository: https://huggingface.co/datasets/espnet/yodas

Dataset Creation

Data Collection and Processing

Read the text content of clips of espnet/yodas

import glob
import re
import pandas as pd
from pathlib import Path
from tqdm.auto import tqdm
from collections import defaultdict
from dataclasses import dataclass, asdict

@dataclass
class Video:
    name: str = ""
    shard: str = ""
    duration: float = 0
    content: str = ""

data = defaultdict(Video)
trange = tqdm(glob.glob("yodas/data/zh000/text/*.txt"))
for file in trange:
    shard = Path(file).stem
    with open(file, "r", encoding="utf8") as f:
        for m in re.finditer(r"(.{11})-\d{5}-\d{8}-(\d{8})\s+(.*)", f.read()):
            name = m.group(1)
            assert data[name].shard in ["", shard]
            data[name].shard = shard
            data[name].name = name
            data[name].duration = int(m.group(2)) / 100
            data[name].content += " " + m.group(3)
    trange.set_postfix(vids=len(data))

data_df = pd.DataFrame(map(asdict, data.values()))

Retain videos with chinese symbols

import re
cjk_pattern = re.compile(
    # puncs \uff00-\uffef \u3000-\u303f
    r"[\u3400-\u4db5\u4e00-\u9fa5\u9fa6-\u9fbb\uf900-\ufa2d\ufa30-\ufa6a\ufa70-\ufad9\u2e80-\u2eff\u31c0-\u31ef\u2f00-\u2fdf\u2ff0-\u2fff\u3100-\u312f\u31a0-\u31bf\ufe10-\ufe1f\ufe30-\ufe4f\u2600-\u26ff\u2700-\u27bf\u3200-\u32ff\u3300-\u33ff]"
)
chinese_df = data_df[data_df['content'].apply(lambda x: cjk_pattern.search(x) is not None)]

Filter out videos with Pingyin's

pinyin_pattern = re.compile(
    r'[üÜāáǎàōóǒòēéěèīíǐìūúǔùǖǘǚǜ]'
)
chinese_pin_df = chinese_df[chinese_df['content'].apply(lambda x: pinyin_pattern.search(x) is None)]

Retain videos with latin scripts

az_pattern = re.compile(
    r"[a-zA-Z]+"
)
mixed_df = chinese_pin_df[chinese_pin_df['content'].apply(lambda x: az_pattern.search(x) is not None)]

Retain videos with punctuations

punc_pattern = re.compile(
    r'[！？。，、·.,?!]'
)
mixed_punc_df = mixed_df[mixed_df['content'].apply(lambda x: punc_pattern.search(x) is not None)]

Sort by increasing proportion of chinese characters

def func(x):
    return x.apply(lambda z: len(cjk_pattern.findall(z)) / len(z))
mixed_punc_df = mixed_punc_df.sort_values(by='content', key=func)

This gives around 1000 videos left.

Save to csv to for manual inspection

mixed_punc_df.to_csv('sanity.csv')

Manually inspect 0-500

NwRTR8mY-7A: mostly english
ASL3yEYC1IE, etc.: contains English translation for each line
Recurring creators whose content is not good code-switching: "天天開心","日向蓝子","笑花兒","关于麻将的职人","大濕：","朋友sisi","please my hero","金玲老師"
Manually pick exceptions to previous rule to add to accepted list
Recurring creators whose content is good code-switching: "我是小夫","久德電子","GL_TECH"
Most videos about: "U.S. stock market", "tech reviews" are accepted.

Quickly skim through 501-1000 (only 10 were picked)

A total of 176 videos were picked in step 8 & 9

Extract selected video clips' audio

from tqdm.auto import tqdm
from pathlib import path
import tarfile

with open("codeswitch.txt", "r") as f: # list of 176 picked video_ids
    codeswitch = set(map(str.strip, f.readlines()))
code_switch_data = data_df[data_df['name'].apply(lambda x: x in codeswitch)]

shard_names = {}
for name, shard in zip(
    code_switch_data['name'].tolist(),
    code_switch_data['shard'].tolist()
):
    if shard not in shard_names:
        shard_names[shard] = set()
    shard_names[shard].add(name)

def extract_wav_files(shard, output_dir):
    # Create the output directory if it doesn't exist
    tar_file_path = f"yodas/data/zh000/audio/{shard}.tar.gz"
    names = shard_names[shard]

    # Open the tar.gz file
    with tarfile.open(tar_file_path, 'r:gz') as tar:
        # Iterate through the contents of the tar file
        for member in tar.getmembers():
            # Check if the member is a WAV file
            video_id = re.search(r"(.{11})-\d{5}-\d{8}-\d{8}", member.name)
            if video_id and video_id.group(1) in names:
                # Extract the WAV file contents into the output directory
                output_path = Path(output_dir, Path(member.name).name)
                with open(output_path, 'wb') as output_file:
                    output_file.write(tar.extractfile(member).read())

output_dir = "./code_switch_yodas"
Path(output_dir).mkdir(exist_ok=True, parents=True)
for shard in tqdm(shard_names):
    extract_wav_files(shard, output_dir)

Publish the subset

import datasets
from datasets import Dataset

audio_dataset = Dataset.from_dict({
    "audio": [
        f"{output_dir}/{clip_id}.wav"
        for clip_id in clip_ids
    ],
    "text": texts,
    "id": clip_ids,
    "session_id": [x[:11] for x in clip_ids]
})
audio_dataset = audio_dataset.cast_column("audio", datasets.features.Audio(sampling_rate=16000))
audio_dataset = audio_dataset.sort("id")
audio_dataset.push_to_hub(
    "georgechang8/code_switch_yodas_zh",
    commit_message="Initial commit",
    embed_external_files=True
)

Extra (without punctuations)

Doing step 1-10, but reverse step 5 to look for ones without punctuations, this yields a small extra set:

extra_set = {
    "37s5xmYYSM8",
    "3ZVVBEugui4",
    "-zHxyIuEw-8",
    "Dngt6Ca8-3u",
    "zJcle9SO98Q",
    "murJVhx5dd0",
    "6hCLoOVtM5Y", # test
    "U-1tallz0hM",
    "wfCUHCYJgIU",
    "GrKoml8qb78",
    "YMTMTFpV7_M",
    "GJV0ZRzAARy",
    "BtMii9364Fg",
    "apK8JYOq6gI",
    "IF-GnMzu7y8",
    "0qJ61eujIVo",
    "Okq02I_jTcA",
    "hCnZlSbTht8",
    "rMk21JBTisE", # validation
    "s9qzwyIM3JI",
    "NBf6Z9R1r7I",
    "jIbc2Jzfa0g",
}

train:
20 videos
validation:
1 video
test:
1 video
DatasetDict({
    train: Dataset({
        features: ['audio', 'text', 'id', 'session_id'],
        num_rows: 5990
    })
    validation: Dataset({
        features: ['audio', 'text', 'id', 'session_id'],
        num_rows: 397
    })
    test: Dataset({
        features: ['audio', 'text', 'id', 'session_id'],
        num_rows: 282
    })
})

Data Cleaning

The video Pew9CK74axu is manually cleaned

def filter_fn(batch):
    return (z == 'Pew9CK74axu' for z in batch['session_id'])

special_care = audio_dataset.filter(filter_fn, num_proc=8, batched=True)
with open("manual_edit.txt", "w", encoding="utf8") as f:
    for l in special_care['text']:
        f.write(l + "\n")
# manual cleaning ...
with open("manual_edit_finish.txt", "r", encoding="utf8") as f:
    lines = list(map(str.strip, f.readlines()))
replace_dict = {
    a: b 
    for a, b in zip(special_care['id'], lines)
}
def manual_edit(batch):
    texts = []
    for sid, orig in zip(batch['id'], batch['text']):
        texts += [replace_dict.get(sid, orig)]
    return {'text': texts}

audio_dataset_manual = audio_dataset.map(manual_edit, batched=True, num_proc=8)

Low log-prob filtering Using whisper-medium to compute the logprob, then filter by a handpicked threshold -3.5

# Get rid of low-prob videos
low_prob_set = {
    '9lQs7INyYBQ',
    'HezOD6XPr_M',
    'HfeLdctBVGY',
    'IzfrgOUd2Uc',
    'UFklIGGKWN0',
    '_x8LwaPRtCE',
    'eK9m6uCNN6Q',
    'erbZNpDMHN0',
    'l9BjfWr1_Pg',
    'nSStWkJtbR4',
    'wrEY_EzQEsy',
    '3Zed0NHrmxo',
    'r29FW7K4iok',
    'MgdQuY0-abI',
    'yHh4rM2KX5Q'
}
audio_dataset_manual = audio_dataset_manual.filter(lambda batch: [s not in low_prob_set for s in batch['session_id']], num_proc=2, batched=True)
# 176 - 14 = 161 videos

train/dev/test split

from datasets import DatasetDict

validation_set = {
    "AyPua3Mi9FU",
    "r29FW7K4iok", # low prob
    "GaUSbuZm5Ec",
    "AKW9vmSy8lQ",
    "3Zed0NHrmxo", # low prob
    "ZHPFLOuT48u",
    "RiCN24FLVLk",
    "zrV_ZNWo8PQ",
    # "rMk21JBTisE", # new (no punc) ==> not in 'default' config
}
test_set = {
    "lH7bZ-8hF1o",
    "WF4ovtdi6wu",
    "MgdQuY0-abI", # low prob
    "yHh4rM2KX5Q", # low prob
    "e_cxHBDSqsM",
    "NO6985Bf_Ro",
    # "6hCLoOVtM5Y", # new (no punc) ==> not in 'default' config
}

def train_fn(batch):
    return (z not in (validation_set|test_set) for z in batch['session_id'])
def validation_fn(batch):
    return (z in validation_set for z in batch['session_id'])
def test_fn(batch):
    return (z in test_set for z in batch['session_id'])

audio_dataset_manual = DatasetDict(
    train=audio_dataset_manual.filter(train_fn, num_proc=2, batched=True),
    validation=audio_dataset_manual.filter(validation_fn, num_proc=2, batched=True),
    test=audio_dataset_manual.filter(test_fn, num_proc=2, batched=True)
)

Don't forget to merge with extra set

from datasets import concatenate_datasets
ds_extra = load_dataset("georgechang8/code_switch_yodas_zh", "clean_extra") # no longer available
audio_dataset_manual = DatasetDict({
    split: concatenate_datasets([audio_dataset_manual[split], ds_extra[split]])
    for split in audio_dataset_manual
})

Do sanity check

ds_full = audio_dataset_manual
for split in ds_full:
    print(split, len(set(ds_full[split]['id'])))
assert len(set(ds_full['train']['id']) & set(ds_full['validation']['id'])) == 0
assert len(set(ds_full['train']['id']) & set(ds_full['test']['id'])) == 0
assert len(set(ds_full['test']['id']) & set(ds_full['validation']['id'])) == 0

General cleansing pipeline

import re
import html

def remove_emojies(text):
  # Ref: https://gist.github.com/Alex-Just/e86110836f3f93fe7932290526529cd1#gistcomment-3208085
  # Ref: https://en.wikipedia.org/wiki/Unicode_block
  EMOJI_PATTERN = re.compile(
    "["
    "\U0001F1E0-\U0001F1FF"  # flags (iOS)
    "\U0001F300-\U0001F5FF"  # symbols & pictographs
    "\U0001F600-\U0001F64F"  # emoticons
    "\U0001F680-\U0001F6FF"  # transport & map symbols
    "\U0001F700-\U0001F77F"  # alchemical symbols
    "\U0001F780-\U0001F7FF"  # Geometric Shapes Extended
    "\U0001F800-\U0001F8FF"  # Supplemental Arrows-C
    "\U0001F900-\U0001F9FF"  # Supplemental Symbols and Pictographs
    "\U0001FA00-\U0001FA6F"  # Chess Symbols
    "\U0001FA70-\U0001FAFF"  # Symbols and Pictographs Extended-A
    "\U00002702-\U000027B0"  # Dingbats
    "]"
  )
  text = re.sub(EMOJI_PATTERN, r' ', text)
  return text

def clean_transcripts(x):
    cjk = "[\u3400-\u4db5\u4e00-\u9fa5\u9fa6-\u9fbb\uf900-\ufa2d\ufa30-\ufa6a\ufa70-\ufad9\uff00-\uffef\u2e80-\u2eff\u3000-\u303f\u31c0-\u31ef\u2f00-\u2fdf\u2ff0-\u2fff\u3100-\u312f\u31a0-\u31bf\ufe10-\ufe1f\ufe30-\ufe4f\u2600-\u26ff\u2700-\u27bf\u3200-\u32ff\u3300-\u33ff]"
    x = html.unescape(x)
    x = remove_emojies(x)
    x = re.sub(r'\.{3,}', ' ', x)
    x = re.sub(r'…+', ' ', x)
    x = re.sub(r'\s+|^|$', '  ', x) # expanding space allows matching " uh uh" case
    x = re.sub(rf"({cjk}|\s)([Uu][mh]|U[MH])({cjk}|\s)", r"\1 \3", x) # uh/um surrounded by cjk or space
    x = re.sub(r"([HhEe]mm+|[HE]MM+)", " ", x) # hmm emm
    x = re.sub(fr"\*+({cjk}+|[A-Za-z]+)\*+", " ", x)  # *叹气* 
    x = re.sub(r'[呃嗯]+', ' ', x)  # 呃嗯
    def replace_except(pattern, repl, z, excs):
        for e, t in excs:
            z = z.replace(e, t)
        z = re.sub(pattern, repl, z)
        for e, t in excs:
            z = z.replace(t, e)
        return z
    # remove 恩 except for 恩桥 感恩 恩怨
    x = replace_except("恩", ' ', x, excs=[("感恩", "呃"),("恩桥", "嗯"),("恩怨", "emm")])
    x = re.sub(r'（[^（）]*）', ' ', x)  # remove （...）
    x = re.sub(r'[（）]+', ' ', x) # remove isolated（）
    x = re.sub(r"\s+", " ", x)
    # remove (...) except for 'Program Files (x86)'
    x = replace_except(r'\([^()]*\)', ' ', x, excs=[("Program Files (x86)", "呃")])
    x = re.sub(r'[()]+', ' ', x) # remove isolated ()
    puncs = r'[，？！。：；～?!,.:;~]'
    x = re.sub(rf'({puncs})(?:\s*\1)+', r'\1', x) # ？？？ -> ？
    x = re.sub(rf"\s+({puncs})", r'\1', x) # text , -> text,
    sp_puncs = r'[?!,.;]' # puncs with spaces
    x = re.sub(rf"({puncs}*{sp_puncs})([^\d])", r'\1 \2', x) # text!?cont -> text!? cont
    x = re.sub(rf"^[\s]*{puncs}+", "", x) # leading puncs
    x = re.sub(r"\s+", " ", x) # excess spaces
    return x.strip()

def clean_batch(batch):
    return {'text': [clean_transcripts(x) for x in batch['text']]}

audio_dataset_manual_clean = audio_dataset_manual.map(clean_batch, batched=True, num_proc=8)

Publish

audio_dataset_manual_clean.push_to_hub(
    "georgechang8/code_switch_yodas_zh",
    config_name="clean",
    set_default=False,
    commit_message="Clean transcript",
    max_shard_size="1GB",
    embed_external_files=True,
)

Limitations

The filtering & hand-picking process might left out useful videos.
The transcriptions is not processed in any way, so might need further cleansing.

Dataset Card Contact

Original dataset: https://huggingface.co/datasets/espnet/yodas CS processing: Chih-Chiang Chang (cc.chang0828@gmail.com)