georgechang8/ASCEND_CLEAN

Dataset Card for Dataset Name

This dataset is derived from CAiRE/ASCEND. More information is available at https://huggingface.co/datasets/CAiRE/ASCEND.

• Removed 嗯 呃 um uh
• Resolved [UNK]'s using whisper-medium

Usage

• Default utterances with cleaned transcripts

from datasets import load_dataset
data = load_dataset("georgechang8/ASCEND_CLEAN")  # add split="train" for train set, etc.

• Concatenated 30s utterances with cleaned transcripts
  • https://github.com/George0828Zhang/distil-whisper/blob/main/training/run_concatenate.py

data = load_dataset("georgechang8/ASCEND_CLEAN", "30s")  # add split="train" for train set, etc.

Dataset Details

Dataset Description

• Language(s): English, Simplified Chinese, Mixed
• License: Creative Common Attribution Share-Alike 4.0 International (CC-BY-SA 4.0)

Dataset Creation

Source Data

https://huggingface.co/datasets/CAiRE/ASCEND

Data Collection and Processing

1. Load from source

from datasets import load_dataset, Audio as DSAudio
data_raw = load_dataset("CAiRE/ASCEND")
data_raw = data_raw.cast_column("audio", DSAudio(sampling_rate=16000))

2. Clean stop words

import re

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 = 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) # replace any uh surrounded by cjk or space
    x = x.replace('嗯', ' ')
    x = x.replace('呃', ' ')
    x = re.sub(r"\s+", " ", x)
    return x.strip()

data = data_raw.map(lambda x: {"transcription": clean_transcripts(x['transcription'])})
data = data.filter(lambda x: x["transcription"] != "")

3. Isolate samples with UNKs

unks = data.filter(lambda x: "[UNK]" in x["transcription"])
unks.shape

{'train': (402, 9), 'test': (36, 9), 'validation': (63, 9)}

4. Load whisper model. For Chinese, medium performs best.

from stable_whisper import load_faster_whisper
model = load_faster_whisper(
    "medium",
    device="cuda",
    compute_type="float16",
)

5. Resolve UNKs with whisper-medium

from sacrebleu.tokenizers.tokenizer_zh import TokenizerZh
from whisper_normalizer.basic import BasicTextNormalizer
import cn2an
import json
import jiwer
from tqdm.auto import tqdm

sacretok = TokenizerZh()
whisper_norm = BasicTextNormalizer()
def compute_mer(hyp, ref):
    def norm(x):
        return sacretok(cn2an.transform(whisper_norm(x), "an2cn"))
    return jiwer.process_words(norm(hyp), norm(ref)).wer * 100

adjusted = {split:dict() for split in data}
double_check = {split:dict() for split in data}

UNK = "[UNK]"

for split in data:
    trange = tqdm(unks[split], desc=split)
    for i,sample in enumerate(trange):
        transcription = sample['transcription']
        texts = transcription.split(UNK)
        words = []
        for sent in texts[1:]:
            for w in sacretok(sent).split():
                if w not in words:
                    words += [w]
        keyword = "关键词"
        header = "字幕"
        prompt = f"{keyword} \"{'/'.join(words)}\" {header} "
        result = model.transcribe_stable(
            audio=sample['audio']['array'],
            initial_prompt=prompt,  # encourage reuse of words
            prefix=texts[0],        # forcing start to follow real start
            language=sample['language'].replace('mixed', 'zh'),
            regroup=False,
            verbose=None,
            no_speech_threshold=1.0,
            suppress_silence=False,
            word_timestamps=True  # though unused, timestamps reduce hallucination
        ).merge_all_segments()
        adjustment = clean_transcripts(
            result.text
            .replace(keyword, " ")
            .replace(header, " ")
        )
        mer=compute_mer(transcription, adjustment)
        adjusted[split][sample['id']] = adjustment
        trange.set_postfix(mer=f"{mer:.2f}", dc=len(double_check[split]))
        if mer > 30:
            double_check[split][sample['id']] = mer
            print(transcription, "||", adjustment)
        if i % 5 == 0 or i == len(unks[split]) - 1:
            with open(f"checkpoint_{split}.json", "w") as f:
                json.dump(adjusted[split], f)

6. Replace UNK utterances with resolved ones

from datasets import DatasetDict
import json

adjusted_transcripts = {}
for split in data_raw:
    with open(f"checkpoint_{split}.json", "r", encoding="utf8") as f:
        adjusted_transcripts[split] = json.load(f)

UNK = "[UNK]"

def fix_unk(sample, adjusted_dict):
    def bad(orig, new):
        return sacretok(new) in sacretok(orig)

    transcription = clean_transcripts(sample['transcription'].replace(UNK, ""))
    sid = sample['id']
    adjustment = adjusted_dict.get(sid, transcription)
    if bad(transcription, adjustment):
        # adjustment worse than just removing UNK
        # print("skipped:", transcription, "||", adjustment)
        adjustment = transcription
    return {"transcription": adjustment}

data = DatasetDict({
    split: data_raw[split].map(lambda x: fix_unk(x, adjusted_transcripts[split]), load_from_cache_file=False)
    for split in data_raw
})
data = data.sort(["session_id","id"], load_from_cache_file=False)

for split in data:
    for line in data[split]['transcription']:
        assert UNK not in line

train adjusted 402 samples, 75 of which just removes UNKs.

test adjusted 36 samples, 9 of which just removes UNKs.

validation adjusted 63 samples, 7 of which just removes UNKs.