speechbrain.lobes.downsampling module

Combinations of processing algorithms to implement downsampling methods.

Authors

• Salah Zaiem

Summary

Classes:

ConcatDownsampler	Concatenation downsampling with naive frame dropping.
Conv1DDownsampler	1D Convolutional downsampling with a learned convolution
Downsampler	Wrapper for downsampling techniques
PoolingDownsampler	1D Pooling downsampling (non-learned)
SignalDownsampler	Signal downsampling (Decimation)

Reference

class speechbrain.lobes.downsampling.Downsampler(*args, **kwargs)

Bases: Module

Wrapper for downsampling techniques

forward(x)

Downsampling function

Parameters:

x (tensor) – Speech samples of shape [B,n_samples] with B the batch size

Return type:

Downsampled outputs.

class speechbrain.lobes.downsampling.SignalDownsampler(downsampling_factor, initial_sampling_rate)

Bases: Downsampler

Signal downsampling (Decimation)

Parameters:

• downsampling_factor (int) – Factor of downsampling (i.e. ratio (length before ds / length after ds))

• initial_sampling_rate (int) – Sampling_rate of the input audios

Example

>>> sd = SignalDownsampler(2, 16000)
>>> a = torch.rand([8, 28000])
>>> a = sd(a)
>>> print(a.shape)
torch.Size([8, 14000])

class speechbrain.lobes.downsampling.Conv1DDownsampler(downsampling_factor, kernel_size)

Bases: Downsampler

1D Convolutional downsampling with a learned convolution

Parameters:

• downsampling_factor (int) – Factor of downsampling (i.e. ratio (length before ds / length after ds))

• kernel_size (int) – Kernel size of the 1D filter (must be an odd integer)

Example

>>> sd = Conv1DDownsampler(3, 161)
>>> a = torch.rand([8, 33000])
>>> a = sd(a)
>>> print(a.shape)
torch.Size([8, 10947])

class speechbrain.lobes.downsampling.PoolingDownsampler(downsampling_factor, kernel_size, padding=0, pool_type='avg')

Bases: Downsampler

1D Pooling downsampling (non-learned)

Parameters:

• downsampling_factor (int) – Factor of downsampling (i.e. ratio (length before ds / length after ds))

• kernel_size (int) – Kernel size of the 1D filter (must be an odd integer)

• padding (int) – The number of padding elements to apply.

• pool_type (string) – Pooling approach, must be within [“avg”,”max”]

Example

>>> sd = PoolingDownsampler(3, 41)
>>> a = torch.rand([8, 33000])
>>> a = sd(a)
>>> print(a.shape)
torch.Size([8, 10987])

class speechbrain.lobes.downsampling.ConcatDownsampler(downsampling_factor)

Bases: Downsampler

Concatenation downsampling with naive frame dropping. Frames are dropped to make the time dimension divisible by the downsampling_factor.

Parameters:

downsampling_factor (int) – Factor of downsampling (i.e. ratio (length before ds / length after ds))

Example

>>> down = ConcatDownsampler(2)
>>> a = torch.rand([8, 40, 40])
>>> a = down(a)
>>> print(a.shape)
torch.Size([8, 20, 80])

forward(x)

Downsamples x given the resampling factor.

Parameters:

x (torch.Tensor) – Factor of downsampling (i.e. ratio (length before ds / length after ds)).

Returns:

x – The downsampled tensor.

Return type:

torch.Tensor