espnet2.enh.abs_enh.AbsEnhancement

source

class espnet2.enh.abs_enh.AbsEnhancement(*args, **kwargs)

Bases: Module, ABC

Abstract base class for audio enhancement models.

This class serves as a blueprint for implementing audio enhancement algorithms. It inherits from torch.nn.Module and requires the implementation of the forward and forward_rawwav methods, which process input tensors representing audio signals and their lengths.

None

• Parameters:None
• Returns: None
• Yields: None
• Raises:
  • NotImplementedError – If the derived class does not implement the
  • required methods. –

######### Examples

To create a custom enhancement model, subclass AbsEnhancement and implement the required methods:

``

`

python class MyEnhancementModel(AbsEnhancement):

def forward(self, input: torch.Tensor, ilens: torch.Tensor) -> Tuple: : # Implement the forward pass pass

def forward_rawwav(self, input: torch.Tensor, ilens: torch.Tensor) -> Tuple: : # Implement the raw waveform processing pass

``

`

NOTE

The output_size method is commented out but can be implemented in derived classes if needed to provide the size of the output tensor.

Initialize internal Module state, shared by both nn.Module and ScriptModule.

abstract forward(input: Tensor, ilens: Tensor) → Tuple[Tensor, Tensor, OrderedDict]

Computes the forward pass of the enhancement model.

This method takes an input tensor and its corresponding lengths and produces an output tensor along with updated lengths and additional information encapsulated in an OrderedDict. The specific implementation of this method should be provided in a derived class that inherits from AbsEnhancement.

• Parameters:
  • input (torch.Tensor) – The input tensor representing the audio signal to be enhanced. It is expected to have a shape of (batch_size, num_channels, signal_length).
  • ilens (torch.Tensor) – A tensor containing the lengths of the input sequences. It should have a shape of (batch_size,).
• Returns: A tuple containing: : - output (torch.Tensor): The enhanced audio signal tensor.
  • olens (torch.Tensor): A tensor with the lengths of the <br/> output sequences.
  • info (OrderedDict): An OrderedDict containing any additional information generated during the forward pass.
• Return type: Tuple[torch.Tensor, torch.Tensor, OrderedDict]
• Raises:NotImplementedError – If this method is called directly from the abstract class without an overriding implementation in a subclass.

######### Examples

>>> model = MyEnhancementModel()  # Assume MyEnhancementModel
>>> input_tensor = torch.randn(10, 1, 16000)  # Example input
>>> ilens = torch.tensor([16000] * 10)  # All sequences have
>>> output, olens, info = model.forward(input_tensor, ilens)

abstract forward_rawwav(input: Tensor, ilens: Tensor) → Tuple[Tensor, Tensor, OrderedDict]

Computes the forward pass for raw waveform input in the enhancement model.

This method processes the input raw waveform tensor and its corresponding lengths to produce the enhanced output, which can be further utilized for tasks such as speech enhancement or signal processing.

• Parameters:
  • input (torch.Tensor) – A tensor containing the raw waveform input data. The shape should be (batch_size, sequence_length).
  • ilens (torch.Tensor) – A tensor containing the lengths of the input sequences. The shape should be (batch_size,).
• Returns: A tuple containing: : - output (torch.Tensor): The enhanced output waveform tensor.
  • output_lengths (torch.Tensor): A tensor containing the lengths <br/> of the output sequences.
  • extras (OrderedDict): A dictionary containing any additional information or metrics computed during the forward pass.
• Return type: Tuple[torch.Tensor, torch.Tensor, OrderedDict]
• Raises:NotImplementedError – If the method is not implemented in a subclass.

######### Examples

>>> model = MyEnhancementModel()
>>> raw_waveform = torch.randn(2, 16000)  # Example raw waveform for 2 samples
>>> lengths = torch.tensor([16000, 12000])  # Lengths of each sample
>>> output, output_lengths, extras = model.forward_rawwav(raw_waveform, lengths)

NOTE

This method must be implemented in any subclass of AbsEnhancement.