Spaces:
Running
on
Zero
Running
on
Zero
File size: 7,677 Bytes
82bc972 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 |
# construct manifest file for training, note that we only have one train split
# also create neighbors folder for each sample, which is simply done through speaker label in the original manifest where each file has rows
# path\tdistance\tduration
# where distance is always 0 because we don't know the distance between the samples
# waiting on Yushen Chen to provide data filtering approach
import sys, copy
import os, random, numpy as np, socket
import json
import tqdm
from multiprocessing import Pool
import glob, os
from collections import defaultdict
def write_jsonl(data, fn):
with open(fn, "w") as file:
for entry in data:
file.write(json.dumps(entry, ensure_ascii=False) + "\n")
def read_jsonl(file_path):
cur_data = []
with open(file_path, 'r', encoding='utf-8-sig') as file:
for line in file:
cur_data.append(json.loads(line.strip()))
return cur_data
def repetition_found(text, length=2, tolerance=10):
pattern_count = defaultdict(int)
for i in range(len(text) - length + 1):
pattern = text[i : i + length]
pattern_count[pattern] += 1
for pattern, count in pattern_count.items():
if count > tolerance:
return True
return False
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"]
from multiprocessing import Pool
def process_meta_item(item, root, sub_root, audio_folder, audio_ext, text_ext):
global filtered_duration, filtered_count, total_duration, total_count
# Data filtering following Yushen's approach
if (
item["wav"].split("/")[-1] in out_en
or any(t in item["text"] for t in en_filters)
or repetition_found(item["text"], length=4)
):
return None, item["duration"], 1, 0, 0, (None, None) # Return filtered results
# Trim leading space from text if exists
if item["text"].startswith(" "):
item["text"] = item["text"][1:]
# write text to text file
text_fn = os.path.join(root, sub_root, audio_folder, item["wav"].replace(audio_ext, text_ext))
os.makedirs(os.path.dirname(text_fn), exist_ok=True)
with open(text_fn, "w") as f:
f.write(item["text"])
# spk2info[item["speaker"]].append(item)
return (
f"{item['wav']}\t{item['duration']}\n",
0,
0,
item["duration"],
1,
(item['speaker'], item)
) # Return processed results
def parallel_process_meta(meta, root, sub_root, audio_folder, num_workers, audio_ext, text_ext):
with Pool(num_workers) as pool:
results = pool.starmap(
process_meta_item,
[(item, root, sub_root, audio_folder, audio_ext, text_ext) for item in meta],
)
processed_items = []
spkitem = []
filtered_duration = 0
filtered_count = 0
total_duration = 0
total_count = 0
for result in results:
if result[0]: # If the item was processed
processed_items.append(result[0])
filtered_duration += result[1]
filtered_count += result[2]
total_duration += result[3]
total_count += result[4]
spkitem.append(result[5])
return processed_items, filtered_duration, filtered_count, total_duration, total_count, spkitem
def main(
root: str = "/data/scratch/pyp/datasets/emilia",
sub_root: str = "preprocessed",
audio_folder: str = "audio",
manifest_folder: str = "manifest_for_codec",
neighbors_folder: str = "neighbors",
audio_ext: str = ".mp3",
text_ext: str = ".txt",
num_workers: int = 8, # Specify the number of workers
):
# Find the segments that are untarred
all_fns = [
item
for item in glob.glob(f"{root}/{sub_root}/{audio_folder}/*")
if os.path.basename(item).startswith("EN_") and os.path.isdir(item)
]
print(f"found {len(all_fns)} untarred segments")
print(f"{all_fns[:3]}")
res = []
total_duration = 0
total_count = 0
filtered_duration = 0
filtered_count = 0
for fn in tqdm.tqdm(all_fns, desc="overall progress"):
spk2info = defaultdict(list)
metafn = os.path.join(root, "EN", os.path.basename(fn) + ".jsonl")
meta = read_jsonl(metafn)
# Parallel process metadata
processed_items, fd, fc, td, tc, spkitem = parallel_process_meta(
meta, root, sub_root, audio_folder, num_workers, audio_ext, text_ext
)
# Aggregate results
res.extend(processed_items)
filtered_duration += fd
filtered_count += fc
total_duration += td
total_count += tc
for spk, item in spkitem:
if spk:
spk2info[spk].append(item)
# Save neighbor files
for spk in spk2info:
for item in spk2info[spk]:
neighbor_fn = os.path.join(
root,
sub_root,
neighbors_folder,
item["wav"].replace(audio_ext, text_ext),
)
os.makedirs(os.path.dirname(neighbor_fn), exist_ok=True)
tobe_write = [f"{neighbor_item['wav'].replace(audio_ext, text_ext)}\t0\t{neighbor_item['duration']}\n" for neighbor_item in spk2info[spk] if neighbor_item["wav"] != item["wav"]]
if tobe_write:
with open(neighbor_fn, "w") as f:
f.writelines(tobe_write)
print(
f"total duration: {total_duration / 3600:.2f} hours, total count: {total_count}"
)
print(
f"filtered duration: {filtered_duration / 3600:.2f} hours, filtered count: {filtered_count}"
)
save_fn = os.path.join(root, sub_root, manifest_folder, "train.txt")
os.makedirs(os.path.dirname(save_fn), exist_ok=True)
with open(save_fn, "w") as f:
for item in res:
f.write(item)
if __name__ == "__main__":
import fire
fire.Fire(main) |