def PrepareGoogleSpeechCmd(version=2, forceDownload=False, task='20cmd'):
"""
Prepares Google Speech commands dataset version 2 for use
tasks: 20cmd, 12cmd, leftright or 35word
Returns full path to training, validation and test file list and file categories
"""
allowedTasks = ['12cmd', 'leftright', '35word', '20cmd']
if task not in allowedTasks:
raise Exception('Task must be one of: {}'.format(allowedTasks))
basePath = None
if version == 2:
_DownloadGoogleSpeechCmdV2(forceDownload)
basePath = 'sd_GSCmdV2'
elif version == 1:
_DownloadGoogleSpeechCmdV1(forceDownload)
basePath = 'sd_GSCmdV1'
else:
raise Exception('Version must be 1 or 2')
if task == '12cmd':
GSCmdV2Categs = {
'unknown': 0,
'silence': 1,
'_unknown_': 0,
'_silence_': 1,
'_background_noise_': 1,
'yes': 2,
'no': 3,
'up': 4,
'down': 5,
'left': 6,
'right': 7,
'on': 8,
'off': 9,
'stop': 10,
'go': 11
}
numGSCmdV2Categs = 12
elif task == 'leftright':
GSCmdV2Categs = {
'unknown': 0,
'silence': 0,
'_unknown_': 0,
'_silence_': 0,
'_background_noise_': 0,
'left': 1,
'right': 2
}
numGSCmdV2Categs = 3
elif task == '35word':
GSCmdV2Categs = {
'unknown': 0,
'silence': 0,
'_unknown_': 0,
'_silence_': 0,
'_background_noise_': 0,
'yes': 2,
'no': 3,
'up': 4,
'down': 5,
'left': 6,
'right': 7,
'on': 8,
'off': 9,
'stop': 10,
'go': 11,
'zero': 12,
'one': 13,
'two': 14,
'three': 15,
'four': 16,
'five': 17,
'six': 18,
'seven': 19,
'eight': 20,
'nine': 1,
'backward': 21,
'bed': 22,
'bird': 23,
'cat': 24,
'dog': 25,
'follow': 26,
'forward': 27,
'happy': 28,
'house': 29,
'learn': 30,
'marvin': 31,
'sheila': 32,
'tree': 33,
'visual': 34,
'wow': 35
}
numGSCmdV2Categs = 36
elif task == '20cmd':
GSCmdV2Categs = {
'unknown': 0,
'silence': 0,
'_unknown_': 0,
'_silence_': 0,
'_background_noise_': 0,
'yes': 2,
'no': 3,
'up': 4,
'down': 5,
'left': 6,
'right': 7,
'on': 8,
'off': 9,
'stop': 10,
'go': 11,
'zero': 12,
'one': 13,
'two': 14,
'three': 15,
'four': 16,
'five': 17,
'six': 18,
'seven': 19,
'eight': 20,
'nine': 1
}
numGSCmdV2Categs = 21
print('Converting test set WAVs to numpy files')
audioUtils.WAV2Numpy(basePath + '/test/')
print('Converting training set WAVs to numpy files')
audioUtils.WAV2Numpy(basePath + '/train/')
# read split from files and all files in folders
testWAVs = pd.read_csv(basePath + '/train/testing_list.txt',
sep=" ",
header=None)[0].tolist()
valWAVs = pd.read_csv(basePath + '/train/validation_list.txt',
sep=" ",
header=None)[0].tolist()
testWAVs = [
os.path.join(basePath + '/train/', f + '.npy') for f in testWAVs
if f.endswith('.wav')
]
valWAVs = [
os.path.join(basePath + '/train/', f + '.npy') for f in valWAVs
if f.endswith('.wav')
]
allWAVs = []
for root, dirs, files in os.walk(basePath + '/train/'):
allWAVs += [root + '/' + f for f in files if f.endswith('.wav.npy')]
trainWAVs = list(set(allWAVs) - set(valWAVs) - set(testWAVs))
testWAVsREAL = []
for root, dirs, files in os.walk(basePath + '/test/'):
testWAVsREAL += [
root + '/' + f for f in files if f.endswith('.wav.npy')
]
# get categories
testWAVlabels = [_getFileCategory(f, GSCmdV2Categs) for f in testWAVs]
valWAVlabels = [_getFileCategory(f, GSCmdV2Categs) for f in valWAVs]
trainWAVlabels = [_getFileCategory(f, GSCmdV2Categs) for f in trainWAVs]
testWAVREALlabels = [
_getFileCategory(f, GSCmdV2Categs) for f in testWAVsREAL
]
# background noise should be used for validation as well
backNoiseFiles = [
trainWAVs[i] for i in range(len(trainWAVlabels))
if trainWAVlabels[i] == GSCmdV2Categs['silence']
]
backNoiseCats = [
GSCmdV2Categs['silence'] for i in range(len(backNoiseFiles))
]
if numGSCmdV2Categs == 12:
valWAVs += backNoiseFiles
valWAVlabels += backNoiseCats
# build dictionaries
testWAVlabelsDict = dict(zip(testWAVs, testWAVlabels))
valWAVlabelsDict = dict(zip(valWAVs, valWAVlabels))
trainWAVlabelsDict = dict(zip(trainWAVs, trainWAVlabels))
testWAVREALlabelsDict = dict(zip(testWAVsREAL, testWAVREALlabels))
# a tweak here: we will heavily underuse silence samples because there are few files.
# we can add them to the training list to reuse them multiple times
# note that since we already added the files to the label dicts we don't
# need to do it again
# for i in range(200):
# trainWAVs = trainWAVs + backNoiseFiles
# info dictionary
trainInfo = {'files': trainWAVs, 'labels': trainWAVlabelsDict}
valInfo = {'files': valWAVs, 'labels': valWAVlabelsDict}
testInfo = {'files': testWAVs, 'labels': testWAVlabelsDict}
testREALInfo = {'files': testWAVsREAL, 'labels': testWAVREALlabelsDict}
gscInfo = {
'train': trainInfo,
'test': testInfo,
'val': valInfo,
'testREAL': testREALInfo
}
print('Done preparing Google Speech commands dataset version {}'.format(
version))
return gscInfo, numGSCmdV2Categs
def prepareKeyword(basePath):
numKeyWordCategs, keywords79 = getCategs()
print('Converting train set WAVs to numpy files')
trainFolders = os.listdir(basePath + '/train')
trainWAVs = []
#traintxts = []
for trainFolder in trainFolders:
root = basePath + '/train/' + trainFolder
if not os.path.isdir(root):
continue
# 如果.wav.npy文件存在，表明.wav文件转换成了.npy无需再次转换
npyFile = [
os.path.join(root, f) for f in os.listdir(root)
if f.endswith('.wav.npy')
]
#traintxt = [os.path.join(root, f) for f in os.listdir(root) if "seg" in f] # 通过seg.txt去对应utterance.wav，删除没有.wav文件的冗余seg.txt
#traintxts.extend(traintxt)
if npyFile:
trainWAVs.extend(npyFile)
continue
# 如果.wav.npy文件不存在，则转换.wav文件
audioUtils.WAV2Numpy(root)
npyFile = [
os.path.join(root, f) for f in os.listdir(root)
if f.endswith('.wav.npy')
]
trainWAVs.extend(npyFile)
#trainWAVtxt = [os.path.join(os.path.dirname(f), os.path.basename(f).split("seg")[0] + "utterance.wav.npy") for f in traintxts]
#wav_txt(trainWAVtxt)
#trainNpy = [len(np.load(x)) for x in trainWAVs]
#print("最大长度：%s， 最小长度：%s，平均长度：%s" % (max(trainNpy),min(trainNpy), sum(trainNpy)/len(trainNpy)))
trainWAVs = bad_audio_fn(deepcopy(trainWAVs))
print('Converting test set WAVs to numpy files')
testFolders = os.listdir(basePath + '/test')
testWAVs = []
#testtxts = []
for testFolder in testFolders:
if testFolder != "0800": # 一个文件一个文件测试
root = basePath + '/test/' + testFolder
if not os.path.isdir(root):
continue
# 如果.wav.npy文件存在，表明.wav文件转换成了.npy无需再次转换
npyFile = [
os.path.join(root, f) for f in os.listdir(root)
if f.endswith('.wav.npy')
]
#testtxt = [os.path.join(root, f) for f in os.listdir(root) if "seg" in f] # 通过seg.txt去对应utterance.wav，删除没有.wav文件的冗余seg.txt
#testtxts.extend(testtxt)
if npyFile:
testWAVs.extend(npyFile)
continue
# 如果.wav.npy文件不存在，则转换.wav文件
audioUtils.WAV2Numpy(root)
npyFile = [
os.path.join(root, f) for f in os.listdir(root)
if f.endswith('.wav.npy')
]
testWAVs.extend(npyFile)
#testWAVtxt = [os.path.join(os.path.dirname(f), os.path.basename(f).split("seg")[0] + "utterance.wav.npy") for f in testtxts]
#wav_txt(testWAVtxt)
#testNpy = [len(np.load(x)) for x in testWAVs]
#print("最大长度：%s， 最小长度：%s，平均长度：%s" % (max(testNpy),min(testNpy), sum(testNpy)/len(testNpy)))
testWAVs = bad_audio_fn(deepcopy(testWAVs))
print('Converting dev set WAVs to numpy files')
devFolders = os.listdir(basePath + '/dev')
devWAVs = []
#devtxts = []
for devFolder in devFolders:
root = basePath + '/dev/' + devFolder
if not os.path.isdir(root):
continue
# 如果.wav.npy文件存在，表明.wav文件转换成了.npy无需再次转换
npyFile = [
os.path.join(root, f) for f in os.listdir(root)
if f.endswith('.wav.npy')
]
#devtxt = [os.path.join(root, f) for f in os.listdir(root) if "seg" in f] # 通过seg.txt去对应utterance.wav，删除没有.wav文件的冗余seg.txt
#devtxts.extend(devtxt)
if npyFile:
devWAVs.extend(npyFile)
continue
# 如果.wav.npy文件不存在，则转换.wav文件
audioUtils.WAV2Numpy(root)
npyFile = [
os.path.join(root, f) for f in os.listdir(root)
if f.endswith('.wav.npy')
]
devWAVs.extend(npyFile)
#devWAVtxt = [os.path.join(os.path.dirname(f), os.path.basename(f).split("seg")[0] + "utterance.wav.npy") for f in devtxts]
#wav_txt(devWAVtxt)
#devNpy = [len(np.load(x)) for x in devWAVs]
#print("最大长度：%s， 最小长度：%s，平均长度：%s" % (max(devNpy),min(devNpy), sum(devNpy)/len(devNpy)))
devWAVs = bad_audio_fn(deepcopy(devWAVs))
# 准备训练数据的Label
trainLabelFiles = [
os.path.join(os.path.dirname(f),
os.path.basename(f).split("utterance")[0] + "seg.txt")
for f in trainWAVs
]
trainLabels = encodeLabel(trainLabelFiles, keywords79)
testLabelFiles = [
os.path.join(os.path.dirname(f),
os.path.basename(f).split("utterance")[0] + "seg.txt")
for f in testWAVs
]
testLabels = encodeLabel(testLabelFiles, keywords79)
devLabelFiles = [
os.path.join(os.path.dirname(f),
os.path.basename(f).split("utterance")[0] + "seg.txt")
for f in devWAVs
]
devLabels = encodeLabel(devLabelFiles, keywords79)
testWAVlabelsDict = dict(zip(testWAVs, testLabels))
devWAVlabelsDict = dict(zip(devWAVs, devLabels))
trainWAVlabelsDict = dict(zip(trainWAVs, trainLabels))
#correspond(testWAVlabelsDict, keywords79, basePath + "/testList.txt")
#correspond(devWAVlabelsDict, keywords79, basePath + "/devList.txt")
#correspond(trainWAVlabelsDict, keywords79, basePath + "/trainList.txt") # 将train,dev,test数据集的文件路径及标签写到txt文件中
trainInfo = {'files': trainWAVs, 'labels': trainWAVlabelsDict}
valInfo = {'files': devWAVs, 'labels': devWAVlabelsDict}
testInfo = {'files': testWAVs, 'labels': testWAVlabelsDict}
gscInfo = {'train': trainInfo, 'test': testInfo, 'val': valInfo}
print("Done Keywords data set prepare.")
return gscInfo, numKeyWordCategs + 1, keywords79
def convertWAV2Numpy():
print('Converting test set WAVs to numpy files')
audioUtils.WAV2Numpy(basePath + '/test/')
print('Converting training set WAVs to numpy files')
audioUtils.WAV2Numpy(basePath + '/train/')
def PrepareGoogleSpeechCmd(version, forceDownload=False, task='35word'):
"""
Prepare google speech command data version2
tasks: Just '35word' for our project
Returns full path to training, validation and test file list and file categories
"""
basePath = None
_DownloadGoogleSpeechCmdV2(forceDownload)
basePath = 'sd_GSCmdV2'
# categories in task '35word'
if task == '35word':
GSCmdV2Categs = {
'unknown': 0,
'silence': 0,
'_unknown_': 0,
'_silence_': 0,
'_background_noise_': 0,
'yes': 2,
'no': 3,
'up': 4,
'down': 5,
'left': 6,
'right': 7,
'on': 8,
'off': 9,
'stop': 10,
'go': 11,
'zero': 12,
'one': 13,
'two': 14,
'three': 15,
'four': 16,
'five': 17,
'six': 18,
'seven': 19,
'eight': 20,
'nine': 1,
'backward': 21,
'bed': 22,
'bird': 23,
'cat': 24,
'dog': 25,
'follow': 26,
'forward': 27,
'happy': 28,
'house': 29,
'learn': 30,
'marvin': 31,
'sheila': 32,
'tree': 33,
'visual': 34,
'wow': 35
}
numGSCmdV2Categs = 36
print(
'Converting test WAVs to numpy files, Data augmentation for test set')
test_aug = audioUtils.WAV2Numpy(basePath + '/test/')
print(
'Converting training set WAVs to numpy files, Data augmentation for train set'
)
train_aug = audioUtils.WAV2Numpy(basePath + '/train/')
# read split from files and all files in folders
testing = pd.read_csv(basePath + '/train/testing_list.txt',
sep=" ",
header=None)[0].tolist()
validation = pd.read_csv(basePath + '/train/validation_list.txt',
sep=" ",
header=None)[0].tolist()
testing = [
os.path.join(basePath + '/train/', f + '.npy') for f in testing
if f.endswith('.wav')
]
validation = [
os.path.join(basePath + '/train/', f + '.npy') for f in validation
if f.endswith('.wav')
]
if test_aug is not None:
testing.extend(test_aug) # add augmentated test files name
allWAVs = []
for Path, dirs, files in os.walk(basePath + '/train/'):
allWAVs += [Path + '/' + f for f in files if f.endswith('.wav.npy')]
if train_aug is not None:
allWAVs.extend(train_aug) #add augmentated train files name
training = list(set(allWAVs) - set(validation) - set(testing))
testWAVsREAL = []
for Path, dirs, files in os.walk(basePath + '/test/'):
testWAVsREAL += [
Path + '/' + f for f in files if f.endswith('.wav.npy')
]
# get categories
testing_label = [_getFileCategory(f, GSCmdV2Categs) for f in testing]
validation_label = [_getFileCategory(f, GSCmdV2Categs) for f in validation]
training_label = [_getFileCategory(f, GSCmdV2Categs) for f in training]
testWAVREALlabels = [
_getFileCategory(f, GSCmdV2Categs) for f in testWAVsREAL
]
# use background noise as validation
BN_list = [
training[i] for i in range(len(training_label))
if training_label[i] == GSCmdV2Categs['silence']
]
BN_categ = [GSCmdV2Categs['silence'] for i in range(len(BN_list))]
if numGSCmdV2Categs == 12:
validation += BN_list
validation_label += BN_categ
# build dictionaries
testing_dict = dict(zip(testing, testing_label))
validation_dict = dict(zip(validation, validation_label))
training_dict = dict(zip(training, training_label))
testing_dict_real = dict(zip(testWAVsREAL, testWAVREALlabels))
trainInfo = {'files': training, 'labels': training_dict}
valInfo = {'files': validation, 'labels': validation_dict}
testInfo = {'files': testing, 'labels': testing_dict}
testREALInfo = {'files': testWAVsREAL, 'labels': testing_dict_real}
gscInfo = {
'train': trainInfo,
'test': testInfo,
'val': valInfo,
'testREAL': testREALInfo
}
print('Done preparing Google Speech commands dataset version {}'.format(
version))
return gscInfo, numGSCmdV2Categs