Update README.md

Updated clean procedure

README.md

@@ -54,7 +54,7 @@ configs:
 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
+- Resolved [UNK]'s using whisper-medium
 
 ## Dataset Details
 
@@ -62,7 +62,7 @@ This dataset is derived from CAiRE/ASCEND. More information is available at http
 
 <!-- Provide a longer summary of what this dataset is. -->
 
-- **Language(s) (NLP):** English, Simplified Chinese, Mixed
+- **Language(s):** English, Simplified Chinese, Mixed
 - **License:** Creative Common Attribution Share-Alike 4.0 International (CC-BY-SA 4.0)
 
 ## Dataset Creation
@@ -77,15 +77,20 @@ https://huggingface.co/datasets/CAiRE/ASCEND
 <!-- 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
+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
 ```python
 import re
 
 def clean_transcripts(x):
-    x = re.sub(r"[嗯呃]|\s+(um|uh)", " ", " " + 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()
 
@@ -110,88 +115,102 @@ model = load_faster_whisper(
 ```
 5. Resolve UNKs with whisper-medium
 ```python
-from stable_whisper.audio.utils import SAMPLE_RATE, resample
 from sacrebleu.tokenizers.tokenizer_zh import TokenizerZh
+from whisper_normalizer.basic import BasicTextNormalizer
+import cn2an
 import json
+import jiwer
 from tqdm.auto import tqdm
-from dataclasses import dataclass, asdict
 
 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
 
-@dataclass
-class Entry:
-    id: str
-    transcription: str
-    adjustment: str
+adjusted = {split:dict() for split in data}
+double_check = {split:dict() for split in data}
 
-adjusted = {'validation':[], 'test':[], 'train':[]}
+UNK = "[UNK]"
 
-for split in unk:
+for split in data:
     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
+        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,
-            prefix=texts[0],                        # forced decoding the same prefix
+            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
+            word_timestamps=True  # though unused, timestamps reduce hallucination
         ).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,
+        adjustment = clean_transcripts(
+            result.text
+            .replace(keyword, " ")
+            .replace(header, " ")
         )
-        adjusted[split].append(entry)
-        if i % 5 == 0:
+        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(list(map(asdict, adjusted[split])), f)
+                json.dump(adjusted[split], f)
 ```
 6. Replace UNK utterances with resolved ones
 ```python
-adjusted = {}
+import json
+
+adjusted_transcripts = {}
 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
+        adjusted_transcripts[split] = json.load(f)
+
+UNK = "[UNK]"
+
+def fix_unk(ds):
+    def bad(orig, new):
+        return sacretok(new) in sacretok(orig)
+
+    for split in ds:
+        adjusted = 0
+        copied = 0
+        for sid, adjustment in adjusted_transcripts[split].items():
+            index = ds[split]['id'].index(sid)
+            transcription = ds[split][index]['transcription'].replace(UNK, "")
+            if bad(transcription, adjustment):
+                # adjustment worse than just removing UNK
+                # print("skipped:", transcription, "||", adjustment)
+                adjustment = transcription
+                copied += 1
+            ds[split][index]['transcription'] = adjustment
+            adjusted += 1
+        print(f"{split} adjusted {adjusted} samples, {copied} of which just removes UNKs.")
+
+    return ds
 
 data = fix_unk(data)
 data = data.filter(lambda x: x["transcription"] != "")
-data = data.sort("id")
-data.save_to_disk('./ASCEND_CLEAN')
+data = data.sort(["session_id","id"])
 ```
-> train adjusted 401 samples.
-
-> test adjusted 36 samples.
+> train adjusted 402 samples, 75 of which just removes UNKs.
+ 
+> test adjusted 36 samples, 9 of which just removes UNKs.
 
-> validation adjusted 61 samples.
+> validation adjusted 63 samples, 7 of which just removes UNKs.