Update README.md

README.md

@@ -42,4 +42,156 @@ configs:
     path: data/test-*
   - split: validation
     path: data/validation-*
+license: cc-by-sa-4.0
+language:
+- en
+- zh
+size_categories:
+- 10K<n<100K
 ---
+# 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
+
+## Dataset Details
+
+### Dataset Description
+
+<!-- Provide a longer summary of what this dataset is. -->
+
+- **Language(s) (NLP):** English, Simplified Chinese, Mixed
+- **License:** Creative Common Attribution Share-Alike 4.0 International (CC-BY-SA 4.0)
+
+## Dataset Creation
+
+### Source Data
+
+<!-- This section describes the source data (e.g. news text and headlines, social media posts, translated sentences, ...). -->
+https://huggingface.co/datasets/CAiRE/ASCEND
+
+#### Data Collection and Processing
+
+<!-- This section describes the data collection and processing process such as data selection criteria, filtering and normalization methods, tools and libraries used, etc. -->
+1. Load from source
+```python
+from datasets import load_dataset
+data_raw = load_dataset("CAiRE/ASCEND")
+```
+2. Clean stop words
+```python
+import re
+
+def clean_transcripts(x):
+    x = re.sub(r"[嗯呃]|\s+(um|uh)", " ", " " + x)
+    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
+```python
+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.
+```python
+from stable_whisper import load_faster_whisper
+model = load_faster_whisper(
+    "medium",
+    device="cuda",
+    compute_type="float16",
+)
+```
+5. Resolve UNKs with whisper-medium
+```python
+from stable_whisper.audio.utils import SAMPLE_RATE, resample
+from sacrebleu.tokenizers.tokenizer_zh import TokenizerZh
+import json
+from tqdm.auto import tqdm
+from dataclasses import dataclass, asdict
+
+sacretok = TokenizerZh()
+
+@dataclass
+class Entry:
+    id: str
+    transcription: str
+    adjustment: str
+
+adjusted = {'validation':[], 'test':[], 'train':[]}
+
+for split in unk:
+    trange = tqdm(unks[split], desc=split)
+    for i,sample in enumerate(trange):
+        if sample['audio']['sampling_rate'] != SAMPLE_RATE:
+            sample['audio'].update(
+                array=resample(sample['audio']['array'], sample['audio']['sampling_rate'], SAMPLE_RATE),
+                sampling_rate=SAMPLE_RATE
+            )
+        texts = sample['transcription'].split("[UNK]")
+        words = sacretok(" ".join(texts[1:]))
+        keyword = "关键词："
+        header = "文本："
+        prompt = f"{keyword}\"{words}\"\n{header}"  # encourages these words
+        result = model.transcribe_stable(
+            audio=sample['audio']['array'],
+            initial_prompt=prompt,
+            prefix=texts[0],                        # forced decoding the same prefix
+            language=sample['language'].replace('mixed', 'zh'),
+            regroup=False,
+            verbose=None,
+            no_speech_threshold=1.0,
+            suppress_silence=False,
+            word_timestamps=True
+        ).merge_all_segments()
+        adjustment = re.sub(
+            r"\s",
+            " ",
+            result.text.replace(keyword, " ").replace(header, " ")
+        ).strip()
+        entry = Entry(
+            id=sample['id'],
+            transcription=sample['transcription'],
+            adjustment=adjustment,
+        )
+        adjusted[split].append(entry)
+        if i % 5 == 0:
+            with open(f"checkpoint_{split}.json", "w") as f:
+                json.dump(list(map(asdict, adjusted[split])), f)
+```
+6. Replace UNK utterances with resolved ones
+```python
+adjusted = {}
+for split in data_raw:
+    with open(f"checkpoint_{split}.json", "r", encoding="utf8") as f:
+        adjusted[split] = json.load(f)
+
+def fix_unk(data):
+    for split in data:
+        n_adjusted = 0
+        for sample in adjusted[split]:
+            sid = sample['id']
+            adjust = sample['adjustment']
+            index = data[split]['id'].index(sid)
+            data[split][index]['transcription'] = adjust
+            n_adjusted += 1
+        print(split, "adjusted", n_adjusted, "samples.")
+
+    return data
+
+data = fix_unk(data)
+data = data.filter(lambda x: x["transcription"] != "")
+data = data.sort("id")
+data.save_to_disk('./ASCEND_CLEAN')
+```
+> train adjusted 401 samples.
+
+> test adjusted 36 samples.
+
+> validation adjusted 61 samples.