import os
from typing import Callable, Tuple, Dict, Optional
from pathlib import Path
import torch
import torchaudio
from torch.utils.data import Dataset
from torch import Tensor
from torchaudio.datasets.utils import (
download_url,
extract_archive
)
from torchvision.datasets.utils import verify_str_arg
import numpy as np
from random import choice
FOLDER_IN_ARCHIVE = "SpeechCommands"
URL = "speech_commands_v0.02"
HASH_DIVIDER = "_nohash_"
EXCEPT_FOLDER = "_background_noise_"
_CHECKSUMS = {
"https://storage.googleapis.com/download.tensorflow.org/data/speech_commands_v0.01.tar.gz":
"3cd23799cb2bbdec517f1cc028f8d43c",
"https://storage.googleapis.com/download.tensorflow.org/data/speech_commands_v0.02.tar.gz":
"6b74f3901214cb2c2934e98196829835",
}
VAL_RECORD = "validation_list.txt"
TEST_RECORD = "testing_list.txt"
TRAIN_RECORD = "training_list.txt"
[文档]def load_speechcommands_item(relpath: str, path: str) -> Tuple[Tensor, int, str, str, int]:
filepath = os.path.join(path, relpath)
label, filename = os.path.split(relpath)
speaker, _ = os.path.splitext(filename)
speaker_id, utterance_number = speaker.split(HASH_DIVIDER)
utterance_number = int(utterance_number)
# Load audio
waveform, sample_rate = torchaudio.load(filepath)
return waveform, sample_rate, label, speaker_id, utterance_number
[文档]class SPEECHCOMMANDS(Dataset):
def __init__(self,
label_dict: Dict,
root: str,
silence_cnt: Optional[int] = 0,
silence_size: Optional[int] = 16000,
transform: Optional[Callable] = None,
url: Optional[str] = URL,
split: Optional[str] = "train",
folder_in_archive: Optional[str] = FOLDER_IN_ARCHIVE,
download: Optional[bool] = False) -> None:
'''
:param label_dict: 标签与类别的对应字典
:type label_dict: Dict
:param root: 数据集的根目录
:type root: str
:param silence_cnt: Silence数据的数量
:type silence_cnt: int, optional
:param silence_size: Silence数据的尺寸
:type silence_size: int, optional
:param transform: A function/transform that takes in a raw audio
:type transform: Callable, optional
:param url: 数据集版本，默认为v0.02
:type url: str, optional
:param split: 数据集划分，可以是 ``"train", "test", "val"``，默认为 ``"train"``
:type split: str, optional
:param folder_in_archive: 解压后的目录名称，默认为 ``"SpeechCommands"``
:type folder_in_archive: str, optional
:param download: 是否下载数据，默认为False
:type download: bool, optional
SpeechCommands语音数据集，出自 `Speech Commands: A Dataset for Limited-Vocabulary Speech Recognition <https://arxiv.org/abs/1804.03209>`_，根据给出的测试集与验证集列表进行了划分，包含v0.01与v0.02两个版本。
数据集包含三大类单词的音频：
#. 指令单词，共10个，"Yes", "No", "Up", "Down", "Left", "Right", "On", "Off", "Stop", "Go". 对于v0.02，还额外增加了5个："Forward", "Backward", "Follow", "Learn", "Visual".
#. 0~9的数字，共10个："One", "Two", "Three", "Four", "Five", "Six", "Seven", "Eight", "Nine".
#. 辅助词，可以视为干扰词，共10个："Bed", "Bird", "Cat", "Dog", "Happy", "House", "Marvin", "Sheila", "Tree", "Wow".
v0.01版本包含共计30类，64,727个音频片段，v0.02版本包含共计35类，105,829个音频片段。更详细的介绍参见前述论文，以及数据集的README。
代码实现基于torchaudio并扩充了功能，同时也参考了 `原论文的实现 <https://github.com/romainzimmer/s2net/blob/b073f755e70966ef133bbcd4a8f0343354f5edcd/data.py>`_。
'''
self.split = verify_str_arg(split, "split", ("train", "val", "test"))
self.label_dict = label_dict
self.transform = transform
self.silence_cnt = silence_cnt
self.silence_size = silence_size
if silence_cnt < 0:
raise ValueError(f"Invalid silence_cnt parameter: {silence_cnt}")
if silence_size <= 0:
raise ValueError(f"Invalid silence_size parameter: {silence_size}")
if url in [
"speech_commands_v0.01",
"speech_commands_v0.02",
]:
base_url = "https://storage.googleapis.com/download.tensorflow.org/data/"
ext_archive = ".tar.gz"
url = os.path.join(base_url, url + ext_archive)
basename = os.path.basename(url)
archive = os.path.join(root, basename)
basename = basename.rsplit(".", 2)[0]
folder_in_archive = os.path.join(folder_in_archive, basename)
self._path = os.path.join(root, folder_in_archive)
self.noise_list = sorted(str(p) for p in Path(self._path).glob('_background_noise_/*.wav'))
if download:
if not os.path.isdir(self._path):
if not os.path.isfile(archive):
checksum = _CHECKSUMS.get(url, None)
download_url(url, root, hash_value=checksum, hash_type="md5")
extract_archive(archive, self._path)
elif not os.path.isdir(self._path):
raise FileNotFoundError("Audio data not found. Please specify \"download=True\" and try again.")
if self.split == "train":
record = os.path.join(self._path, TRAIN_RECORD)
if os.path.exists(record):
with open(record, 'r') as f:
self._walker = list([line.rstrip('\n') for line in f])
else:
print("No training list, generating...")
walker = sorted(str(p) for p in Path(self._path).glob('*/*.wav'))
walker = filter(lambda w: HASH_DIVIDER in w and EXCEPT_FOLDER not in w, walker)
walker = map(lambda w: os.path.relpath(w, self._path), walker)
walker = set(walker)
val_record = os.path.join(self._path, VAL_RECORD)
with open(val_record, 'r') as f:
val_walker = set([line.rstrip('\n') for line in f])
test_record = os.path.join(self._path, TEST_RECORD)
with open(test_record, 'r') as f:
test_walker = set([line.rstrip('\n') for line in f])
walker = walker - val_walker - test_walker
self._walker = list(walker)
with open(record, 'w') as f:
f.write('\n'.join(self._walker))
print("Training list generated!")
labels = [self.label_dict.get(os.path.split(relpath)[0]) for relpath in self._walker]
label_weights = 1. / np.unique(labels, return_counts=True)[1]
if self.silence_cnt == 0:
label_weights /= np.sum(label_weights)
self.weights = torch.DoubleTensor([label_weights[label] for label in labels])
else:
silence_weight = 1. / self.silence_cnt
total_weight = np.sum(label_weights) + silence_weight
label_weights /= total_weight
self.weights = torch.DoubleTensor([label_weights[label] for label in labels] + [silence_weight / total_weight] * self.silence_cnt)
else:
if self.split == "val":
record = os.path.join(self._path, VAL_RECORD)
else:
record = os.path.join(self._path, TEST_RECORD)
with open(record, 'r') as f:
self._walker = list([line.rstrip('\n') for line in f])
def __getitem__(self, n: int) -> Tuple[Tensor, int]:
if n < len(self._walker):
fileid = self._walker[n]
waveform, sample_rate, label, speaker_id, utterance_number = load_speechcommands_item(fileid, self._path)
else:
# Silence data are randomly and dynamically generated from noise data
# Load random noise
noisepath = os.path.join(self._path, choice(self.noise_list))
waveform, sample_rate = torchaudio.load(noisepath)
# Random crop
offset = np.random.randint(waveform.shape[1] - self.silence_size)
waveform = waveform[:, offset:offset + self.silence_size]
label = "_silence_"
m = waveform.abs().max()
if m > 0:
waveform /= m
if self.transform is not None:
waveform = self.transform(waveform)
label = self.label_dict.get(label)
return waveform, label
def __len__(self) -> int:
return len(self._walker) + self.silence_cnt