espnet2.diar.decoder.linear_decoder.LinearDecoder

source

class espnet2.diar.decoder.linear_decoder.LinearDecoder(encoder_output_size: int, num_spk: int = 2)

Bases: AbsDecoder

Linear decoder for speaker diarization.

This class implements a linear decoder that processes the output of an encoder for speaker diarization tasks. It inherits from the AbsDecoder class and uses a linear layer to map encoder outputs to the desired number of speakers.

num_spk

The number of speakers that the decoder can output.

• Type: int

• Parameters:

  • encoder_output_size (int) – The size of the encoder’s output feature vector.
  • num_spk (int , optional) – The number of speakers to decode. Defaults to 2.

• Returns: The decoded output tensor with shape [Batch, T, num_spk].

• Return type: torch.Tensor

######### Examples

>>> decoder = LinearDecoder(encoder_output_size=128, num_spk=3)
>>> input_tensor = torch.randn(10, 50, 128)  # Batch size 10, T=50, F=128
>>> ilens = torch.tensor([50] * 10)  # All sequences have length 50
>>> output = decoder(input_tensor, ilens)
>>> print(output.shape)
torch.Size([10, 50, 3])  # Output shape corresponds to num_spk

• Raises:ValueError – If input does not have the correct shape or dimensions.

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

#

forward(input

: Tensor, ilens: Tensor)

Forward pass of the LinearDecoder.

This method takes the input tensor representing the hidden space and the input lengths, and applies a linear transformation to decode the speaker representations.

• Parameters:
  • input (torch.Tensor) – A tensor of shape [Batch, T, F] representing the hidden space, where ‘Batch’ is the number of samples, ‘T’ is the time dimension, and ‘F’ is the feature dimension.
  • ilens (torch.Tensor) – A tensor of shape [Batch] representing the lengths of the input sequences.
• Returns: A tensor of shape [Batch, T, num_spk] representing the decoded speaker outputs, where ‘num_spk’ is the number of speakers.
• Return type: torch.Tensor

######### Examples

>>> decoder = LinearDecoder(encoder_output_size=256, num_spk=3)
>>> input_tensor = torch.randn(10, 20, 256)  # Batch of 10, 20 time steps, 256 features
>>> input_lengths = torch.tensor([20] * 10)  # All sequences are of length 20
>>> output = decoder.forward(input_tensor, input_lengths)
>>> output.shape
torch.Size([10, 20, 3])  # Decoded output for 3 speakers

NOTE

The input tensor should be properly normalized and prepared before passing to the forward method.

property num_spk

Linear decoder for speaker diarization.

This class implements a linear decoder that maps encoder outputs to a specified number of speakers. It is designed to work with the output of an encoder in speaker diarization tasks.

num_spk

The number of speakers the decoder can handle.

• Type: int

• Parameters:

  • encoder_output_size (int) – The size of the output from the encoder.
  • num_spk (int , optional) – The number of speakers to decode. Defaults to 2.

forward(input

torch.Tensor, ilens: torch.Tensor) -> torch.Tensor: Computes the forward pass of the linear decoder.

######### Examples

Initialize the LinearDecoder with encoder output size of 128 and 3 speakers

decoder = LinearDecoder(encoder_output_size=128, num_spk=3)

Create a dummy input tensor and input lengths

input_tensor = torch.randn(10, 20, 128) # [Batch, T, F] input_lengths = torch.randint(1, 21, (10,)) # Random lengths for each batch

Perform a forward pass

output = decoder.forward(input_tensor, input_lengths)

Access the number of speakers

number_of_speakers = decoder.num_spk # Should return 3

NOTE

The forward method expects the input tensor to have a shape of [Batch, T, F] where T is the time dimension and F is the feature dimension.

• Raises:ValueError – If input dimensions do not match the expected shape.