Spaces:
Sleeping
Sleeping
File size: 9,087 Bytes
957e2dc |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 |
import os
import math
import librosa as li
import numpy as np
import textgrid
import torch
from src.data import DataProperties, VoiceBoxDataset
from src.utils import ensure_dir
from src.constants import (
LIBRISPEECH_DATA_DIR,
LIBRISPEECH_CACHE_DIR,
SAMPLE_RATE,
LIBRISPEECH_EXT,
LIBRISPEECH_PHONEME_EXT,
LIBRISPEECH_PHONEME_DICT,
LIBRISPEECH_SIG_LEN,
HOP_LENGTH
)
from src.attacks.offline.perturbation.voicebox.voicebox import PitchEncoder
from os import path
from tqdm import tqdm
from pathlib import Path
from typing import Union, Iterable
################################################################################
# Cache and load LibriSpeech dataset
################################################################################
class LibriSpeechDataset(VoiceBoxDataset):
"""
A Dataset object for the LibriSpeech dataset subsets. The required data can
be downloaded by running the script `download_librispeech.sh`. This class
takes audio data from the specified directory and caches tensors to disk.
"""
def __init__(self,
split: str = 'test-clean',
data_dir: str = LIBRISPEECH_DATA_DIR,
cache_dir: str = LIBRISPEECH_CACHE_DIR,
sample_rate: int = SAMPLE_RATE,
audio_ext: str = LIBRISPEECH_EXT,
phoneme_ext: str = LIBRISPEECH_PHONEME_EXT,
signal_length: Union[float, int] = LIBRISPEECH_SIG_LEN,
scale: Union[float, int] = 1.0,
hop_length: int = HOP_LENGTH,
target: str = 'speaker',
features: Union[str, Iterable[str]] = None,
batch_format: str = 'dict',
*args,
**kwargs):
"""
Load, organize, and cache LibriSpeech dataset.
Parameters
----------
split (str):
data_dir (str): LibriSpeech root directory
cache_dir (str): root directory to which tensors will be saved
sample_rate (int): sample rate in Hz
audio_ext (str): extension for audio files within dataset
phoneme_ext (str): extension for phoneme alignment files within
dataset
signal_length (int): length of audio files in samples (if `int` given)
or seconds (if `float` given)
scale (float): range to which audio will be scaled
hop_length (int): hop size for computing frame-wise features (e.g.
pitch, loudness)
target (str): string specifying target type. Must be one of
`speaker` (speaker ID), `phoneme` (aligned phoneme
labels), or `transcript`
features (Iterable): strings specifying features to compute for each
audio file in the dataset. Must be subset of
`pitch`, `periodicity`, `loudness`
batch_format (str): format for returning batches. Must be either `dict`
or `tuple`
"""
self.phoneme_ext = phoneme_ext
self.phoneme_list = []
super().__init__(
split=split,
data_dir=data_dir,
cache_dir=cache_dir,
audio_ext=audio_ext,
signal_length=signal_length,
scale=scale,
target=target,
features=features,
sample_rate=sample_rate,
hop_length=hop_length,
batch_format=batch_format,
*args, **kwargs
)
def __str__(self):
"""Return string representation of dataset"""
return f'LibriSpeechDataset(split={self.split}, ' \
f'target={self.target}, features={self.features})'
@staticmethod
def _check_split(split: str):
"""Check for valid dataset split"""
if split not in [
'test-clean',
'test-other',
'dev-clean',
'dev-other',
'train-clean-100',
'train-clean-360',
'train-other-500'
]:
raise ValueError(f'Invalid split {split}')
return split
@staticmethod
def _check_target(target: str):
if target not in ['speaker', 'phoneme', 'transcript']:
raise ValueError(f'Invalid target type {target}')
return target
def _get_target_id(self):
"""Identifier for cached targets"""
if self.target in ['speaker', 'transcript']:
return f'{self.target}'
else:
return f'{self.sample_rate}-{self.hop_length}-{self.target}'
def _get_audio_list(self, *args, **kwargs):
"""
Scan for all audio files with given extension. Additionally, only select
audio files for which corresponding phoneme alignments exist.
"""
audio_files = [os.path.splitext(f)[0] for f in
(Path(self.data_dir) / self.split).rglob(
f'*.{self.audio_ext}')]
phoneme_files = [os.path.splitext(f)[0] for f in
(Path(self.data_dir) / self.split).rglob(
f'*.{self.phoneme_ext}')]
matching_files = list(set(audio_files) & set(phoneme_files))
return sorted(
[f + "." + self.audio_ext for f in matching_files]
)
def _build_target_cache(self):
"""Process and cache targets"""
target_id = self._get_target_id()
target_cache = list(
(Path(self.cache_dir) / self.split).rglob(
f'{target_id}.pt')
)
if len(target_cache) >= 1:
return
# speaker ID targets
if self.target == 'speaker':
targets = torch.zeros(
len(self.audio_list), dtype=torch.long
)
pbar = tqdm(self.audio_list, total=len(self.audio_list))
for i, audio_fn in enumerate(pbar):
pbar.set_description(
f'Loading Speaker IDs ({self.split}): '
f'{path.basename(audio_fn)}')
# extract speaker ID
targets[i] = int(Path(audio_fn).parts[-3])
# frame-aligned phoneme label targets
elif self.target == 'phoneme':
# retrieve phoneme alignment files
self.phoneme_list = [
os.path.splitext(f)[0] +
"." + self.phoneme_ext for f in self.audio_list]
targets = torch.zeros(len(self.phoneme_list),
self.num_frames,
dtype=torch.long)
pbar = tqdm(self.phoneme_list, total=len(self.phoneme_list))
for i, phoneme_fn in enumerate(pbar):
pbar.set_description(
f'Loading phoneme alignments ({self.split}): '
f'{path.basename(phoneme_fn)}')
# load interval labels from TextGrid format
tg = textgrid.TextGrid.fromFile(phoneme_fn)
if tg[0].name == 'phones':
phoneme_intervals = tg[0]
elif tg[1].name == 'phones':
phoneme_intervals = tg[1]
else:
raise ValueError("Could not find phonemes")
# compute number of frames in audio file given hop size,
# rounding up
num_frames = math.ceil(
tg.maxTime * self.sample_rate / self.hop_length)
ppg = torch.zeros(num_frames, dtype=torch.long)
# for each labeled interval, break up into frames with given hop
# size and assign phoneme labels
for interval in phoneme_intervals:
interval.minTime = math.ceil(
interval.minTime * self.sample_rate / self.hop_length)
interval.maxTime = math.ceil(
interval.maxTime * self.sample_rate / self.hop_length)
phoneme_idx = LIBRISPEECH_PHONEME_DICT[interval.mark]
ppg[interval.minTime:interval.maxTime+1] = phoneme_idx
targets[
i, :min(ppg.shape[-1], self.num_frames)
] = ppg[..., :self.num_frames]
# string transcript targets
elif self.target == 'transcript':
raise NotImplementedError()
else:
raise ValueError(f'Invalid target type {self.target}')
# cache targets to disk
torch.save(targets,
path.join(
self.cache_dir,
self.split,
f'{target_id}.pt'
))
|