Spaces:
Running
Running
File size: 7,325 Bytes
dd217c7 d37849f dd217c7 d37849f dd217c7 d37849f dd217c7 d37849f dd217c7 d37849f dd217c7 d37849f dd217c7 d37849f dd217c7 d37849f dd217c7 d37849f dd217c7 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 |
# Emilia Dataset: https://huggingface.co/datasets/amphion/Emilia-Dataset/tree/fc71e07
# if use updated new version, i.e. WebDataset, feel free to modify / draft your own script
# generate audio text map for Emilia ZH & EN
# evaluate for vocab size
import sys
import os
sys.path.append(os.getcwd())
from pathlib import Path
import json
from tqdm import tqdm
from concurrent.futures import ProcessPoolExecutor
from datasets.arrow_writer import ArrowWriter
from model.utils import (
repetition_found,
convert_char_to_pinyin,
)
out_zh = {
"ZH_B00041_S06226",
"ZH_B00042_S09204",
"ZH_B00065_S09430",
"ZH_B00065_S09431",
"ZH_B00066_S09327",
"ZH_B00066_S09328",
}
zh_filters = ["い", "て"]
# seems synthesized audios, or heavily code-switched
out_en = {
"EN_B00013_S00913",
"EN_B00042_S00120",
"EN_B00055_S04111",
"EN_B00061_S00693",
"EN_B00061_S01494",
"EN_B00061_S03375",
"EN_B00059_S00092",
"EN_B00111_S04300",
"EN_B00100_S03759",
"EN_B00087_S03811",
"EN_B00059_S00950",
"EN_B00089_S00946",
"EN_B00078_S05127",
"EN_B00070_S04089",
"EN_B00074_S09659",
"EN_B00061_S06983",
"EN_B00061_S07060",
"EN_B00059_S08397",
"EN_B00082_S06192",
"EN_B00091_S01238",
"EN_B00089_S07349",
"EN_B00070_S04343",
"EN_B00061_S02400",
"EN_B00076_S01262",
"EN_B00068_S06467",
"EN_B00076_S02943",
"EN_B00064_S05954",
"EN_B00061_S05386",
"EN_B00066_S06544",
"EN_B00076_S06944",
"EN_B00072_S08620",
"EN_B00076_S07135",
"EN_B00076_S09127",
"EN_B00065_S00497",
"EN_B00059_S06227",
"EN_B00063_S02859",
"EN_B00075_S01547",
"EN_B00061_S08286",
"EN_B00079_S02901",
"EN_B00092_S03643",
"EN_B00096_S08653",
"EN_B00063_S04297",
"EN_B00063_S04614",
"EN_B00079_S04698",
"EN_B00104_S01666",
"EN_B00061_S09504",
"EN_B00061_S09694",
"EN_B00065_S05444",
"EN_B00063_S06860",
"EN_B00065_S05725",
"EN_B00069_S07628",
"EN_B00083_S03875",
"EN_B00071_S07665",
"EN_B00071_S07665",
"EN_B00062_S04187",
"EN_B00065_S09873",
"EN_B00065_S09922",
"EN_B00084_S02463",
"EN_B00067_S05066",
"EN_B00106_S08060",
"EN_B00073_S06399",
"EN_B00073_S09236",
"EN_B00087_S00432",
"EN_B00085_S05618",
"EN_B00064_S01262",
"EN_B00072_S01739",
"EN_B00059_S03913",
"EN_B00069_S04036",
"EN_B00067_S05623",
"EN_B00060_S05389",
"EN_B00060_S07290",
"EN_B00062_S08995",
}
en_filters = ["ا", "い", "て"]
def deal_with_audio_dir(audio_dir):
audio_jsonl = audio_dir.with_suffix(".jsonl")
sub_result, durations = [], []
vocab_set = set()
bad_case_zh = 0
bad_case_en = 0
with open(audio_jsonl, "r") as f:
lines = f.readlines()
for line in tqdm(lines, desc=f"{audio_jsonl.stem}"):
obj = json.loads(line)
text = obj["text"]
if obj["language"] == "zh":
if obj["wav"].split("/")[1] in out_zh or any(f in text for f in zh_filters) or repetition_found(text):
bad_case_zh += 1
continue
else:
text = text.translate(
str.maketrans({",": "，", "!": "！", "?": "？"})
) # not "。" cuz much code-switched
if obj["language"] == "en":
if (
obj["wav"].split("/")[1] in out_en
or any(f in text for f in en_filters)
or repetition_found(text, length=4)
):
bad_case_en += 1
continue
if tokenizer == "pinyin":
text = convert_char_to_pinyin([text], polyphone=polyphone)[0]
duration = obj["duration"]
sub_result.append({"audio_path": str(audio_dir.parent / obj["wav"]), "text": text, "duration": duration})
durations.append(duration)
vocab_set.update(list(text))
return sub_result, durations, vocab_set, bad_case_zh, bad_case_en
def main():
assert tokenizer in ["pinyin", "char"]
result = []
duration_list = []
text_vocab_set = set()
total_bad_case_zh = 0
total_bad_case_en = 0
# process raw data
executor = ProcessPoolExecutor(max_workers=max_workers)
futures = []
for lang in langs:
dataset_path = Path(os.path.join(dataset_dir, lang))
[
futures.append(executor.submit(deal_with_audio_dir, audio_dir))
for audio_dir in dataset_path.iterdir()
if audio_dir.is_dir()
]
for futures in tqdm(futures, total=len(futures)):
sub_result, durations, vocab_set, bad_case_zh, bad_case_en = futures.result()
result.extend(sub_result)
duration_list.extend(durations)
text_vocab_set.update(vocab_set)
total_bad_case_zh += bad_case_zh
total_bad_case_en += bad_case_en
executor.shutdown()
# save preprocessed dataset to disk
if not os.path.exists(f"data/{dataset_name}"):
os.makedirs(f"data/{dataset_name}")
print(f"\nSaving to data/{dataset_name} ...")
# dataset = Dataset.from_dict({"audio_path": audio_path_list, "text": text_list, "duration": duration_list}) # oom
# dataset.save_to_disk(f"data/{dataset_name}/raw", max_shard_size="2GB")
with ArrowWriter(path=f"data/{dataset_name}/raw.arrow") as writer:
for line in tqdm(result, desc="Writing to raw.arrow ..."):
writer.write(line)
# dup a json separately saving duration in case for DynamicBatchSampler ease
with open(f"data/{dataset_name}/duration.json", "w", encoding="utf-8") as f:
json.dump({"duration": duration_list}, f, ensure_ascii=False)
# vocab map, i.e. tokenizer
# add alphabets and symbols (optional, if plan to ft on de/fr etc.)
# if tokenizer == "pinyin":
# text_vocab_set.update([chr(i) for i in range(32, 127)] + [chr(i) for i in range(192, 256)])
with open(f"data/{dataset_name}/vocab.txt", "w") as f:
for vocab in sorted(text_vocab_set):
f.write(vocab + "\n")
print(f"\nFor {dataset_name}, sample count: {len(result)}")
print(f"For {dataset_name}, vocab size is: {len(text_vocab_set)}")
print(f"For {dataset_name}, total {sum(duration_list)/3600:.2f} hours")
if "ZH" in langs:
print(f"Bad zh transcription case: {total_bad_case_zh}")
if "EN" in langs:
print(f"Bad en transcription case: {total_bad_case_en}\n")
if __name__ == "__main__":
max_workers = 32
tokenizer = "pinyin" # "pinyin" | "char"
polyphone = True
langs = ["ZH", "EN"]
dataset_dir = "<SOME_PATH>/Emilia_Dataset/raw"
dataset_name = f"Emilia_{'_'.join(langs)}_{tokenizer}"
print(f"\nPrepare for {dataset_name}\n")
main()
# Emilia ZH & EN
# samples count 37837916 (after removal)
# pinyin vocab size 2543 (polyphone)
# total duration 95281.87 (hours)
# bad zh asr cnt 230435 (samples)
# bad eh asr cnt 37217 (samples)
# vocab size may be slightly different due to jieba tokenizer and pypinyin (e.g. way of polyphoneme)
# please be careful if using pretrained model, make sure the vocab.txt is same
|