mindspore.dataset.audio.TimeMasking

class mindspore.dataset.audio.TimeMasking(iid_masks=False, time_mask_param=0, mask_start=0, mask_value=0.0)

Apply masking to a spectrogram in the time domain.

Note

The shape of the audio waveform to be processed needs to be <…, freq, time>.

Parameters

• iid_masks (bool, optional) – Whether to apply different masks to each example/channel. Default: False.

• time_mask_param (int, optional) – When iid_masks is True, length of the mask will be uniformly sampled from [0, time_mask_param]; When iid_masks is False, directly use it as length of the mask. The value should be in range of [0, time_length], where time_length is the length of audio waveform in time domain. Default: 0.

• mask_start (int, optional) – Starting point to apply mask, only works when iid_masks is True. The value should be in range of [0, time_length - time_mask_param], where time_length is the length of audio waveform in time domain. Default: 0.

• mask_value (float, optional) – Value to assign to the masked columns. Default: 0.0.

Raises

• TypeError – If iid_masks is not of type bool.

• TypeError – If time_mask_param is not of type int.

• ValueError – If time_mask_param is greater than the length of audio waveform in time domain.

• TypeError – If mask_start is not of type int.

• ValueError – If mask_start a negative number.

• TypeError – If mask_value is not of type float.

• ValueError – If mask_value is a negative number.

• RuntimeError – If input tensor is not in shape of <…, freq, time>.

Supported Platforms:

CPU

Examples

>>> import numpy as np
>>> import mindspore.dataset as ds
>>> import mindspore.dataset.audio as audio
>>>
>>> # Use the transform in dataset pipeline mode
>>> waveform = np.random.random([5, 16, 2])  # 5 samples
>>> numpy_slices_dataset = ds.NumpySlicesDataset(data=waveform, column_names=["audio"])
>>> transforms = [audio.TimeMasking(time_mask_param=1)]
>>> numpy_slices_dataset = numpy_slices_dataset.map(operations=transforms, input_columns=["audio"])
>>> for item in numpy_slices_dataset.create_dict_iterator(num_epochs=1, output_numpy=True):
...     print(item["audio"].shape, item["audio"].dtype)
...     break
(16, 2) float64
>>>
>>> # Use the transform in eager mode
>>> waveform = np.random.random([16, 2])  # 1 sample
>>> output = audio.TimeMasking(time_mask_param=1)(waveform)
>>> print(output.shape, output.dtype)
(16, 2) float64

Tutorial Examples:

• Illustration of audio transforms