def __init__(self, feature_extract_by: str = 'librosa', sample_rate: int = 16000,
n_mels: int = 80, frame_length: int = 20, frame_shift: int = 10,
del_silence: bool = False, input_reverse: bool = True,
normalize: bool = False, transform_method: str = 'mel',
time_mask_para: int = 70, freq_mask_para: int = 12, time_mask_num: int = 2, freq_mask_num: int = 2,
sos_id: int = 1, eos_id: int = 2, target_dict: dict = None, noise_augment: bool = False,
dataset_path: str = None, noiseset_size: int = 0, noise_level: float = 0.7) -> None:
super(SpectrogramParser, self).__init__(dataset_path, noiseset_size, sample_rate, noise_level, noise_augment)
self.del_silence = del_silence
self.input_reverse = input_reverse
self.normalize = normalize
self.sos_id = sos_id
self.eos_id = eos_id
self.target_dict = target_dict
self.spec_augment = SpecAugment(time_mask_para, freq_mask_para, time_mask_num, freq_mask_num)
if transform_method.lower() == 'mel':
self.transforms = MelSpectrogram(sample_rate, n_mels, frame_length, frame_shift, feature_extract_by)
elif transform_method.lower() == 'mfcc':
self.transforms = MFCC(sample_rate, n_mels, frame_length, frame_shift, feature_extract_by)
elif transform_method.lower() == 'spect':
self.transforms = Spectrogram(sample_rate, frame_length, frame_shift, feature_extract_by)
elif transform_method.lower() == 'fbank':
self.transforms = FilterBank(sample_rate, n_mels, frame_length, frame_shift)
else:
raise ValueError("Unsupported feature : {0}".format(transform_method))
def __init__(
self,
feature_extract_by:
str = 'librosa', # which library to use for feature extraction
sample_rate: int = 16000, # sample rate of audio signal.
n_mels: int = 80, # Number of mfc coefficients to retain.
frame_length: int = 20, # frame length for spectrogram
frame_shift: int = 10, # Length of hop between STFT windows.
del_silence:
bool = False, # flag indication whether to delete silence or not
input_reverse:
bool = True, # flag indication whether to reverse input or not
normalize:
bool = False, # flag indication whether to normalize spectrum or not
transform_method:
str = 'mel', # which feature to use [mel, fbank, spect, mfcc]
freq_mask_para:
int = 12, # hyper Parameter for Freq Masking to limit freq masking length
time_mask_num: int = 2, # how many time-masked area to make
freq_mask_num: int = 2, # how many freq-masked area to make
sos_id: int = 1, # start of sentence token`s identification
eos_id: int = 2, # end of sentence token`s identification
dataset_path: str = None, # noise dataset path
audio_extension: str = 'pcm' # audio extension
) -> None:
super(SpectrogramParser, self).__init__(dataset_path)
self.del_silence = del_silence
self.input_reverse = input_reverse
self.normalize = normalize
self.sos_id = sos_id
self.eos_id = eos_id
self.spec_augment = SpecAugment(freq_mask_para, time_mask_num,
freq_mask_num)
self.audio_extension = audio_extension
if transform_method.lower() == 'mel':
self.transforms = MelSpectrogram(sample_rate, n_mels, frame_length,
frame_shift, feature_extract_by)
elif transform_method.lower() == 'mfcc':
self.transforms = MFCC(sample_rate, n_mels, frame_length,
frame_shift, feature_extract_by)
elif transform_method.lower() == 'spect':
self.transforms = Spectrogram(sample_rate, frame_length,
frame_shift, feature_extract_by)
elif transform_method.lower() == 'fbank':
self.transforms = FilterBank(sample_rate, n_mels, frame_length,
frame_shift)
else:
raise ValueError(
"Unsupported feature : {0}".format(transform_method))