espnet2.spk.encoder.xvector_encoder.XvectorEncoder

source

class espnet2.spk.encoder.xvector_encoder.XvectorEncoder(input_size: int, ndim: int = 512, output_size: int = 1500, kernel_sizes: List = [5, 3, 3, 1, 1], paddings: List = [2, 1, 1, 0, 0], dilations: List = [1, 2, 3, 1, 1], **kwargs)

Bases: AbsEncoder

X-vector encoder. Extracts frame-level x-vector embeddings from features.

This class implements the X-vector model for speaker recognition as described in the paper by D. Snyder et al., “X-vectors: Robust dnn embeddings for speaker recognition,” presented at IEEE ICASSP, 2018. The model takes input features and processes them through a series of convolutional layers to produce speaker embeddings.

layers

A list of convolutional layers, ReLU activations, and batch normalization layers.

• Type: nn.ModuleList

_output_size

The output embedding dimension.

• Type: int

• Parameters:

  • input_size (int) – Input feature dimension.
  • ndim (int , optional) – Dimensionality of the hidden representation. Defaults to 512.
  • output_size (int , optional) – Output embedding dimension. Defaults to 1500.
  • kernel_sizes (List , optional) – List of kernel sizes for each convolutional layer. Defaults to [5, 3, 3, 1, 1].
  • paddings (List , optional) – List of padding sizes for each convolutional layer. Defaults to [2, 1, 1, 0, 0].
  • dilations (List , optional) – List of dilation rates for each convolutional layer. Defaults to [1, 2, 3, 1, 1].
  • **kwargs – Additional keyword arguments.

######### Examples

>>> encoder = XvectorEncoder(input_size=40)
>>> input_tensor = torch.randn(10, 100, 40)  # (Batch, Sequence, Features)
>>> output = encoder(input_tensor)
>>> print(output.shape)  # Output shape will depend on the configuration

NOTE

This implementation is adapted for ESPnet-SPK by Jee-weon Jung, and cross-checked with the SpeechBrain implementation: https://github.com/speechbrain/speechbrain/blob/develop/speechbrain/lobes/models/Xvector.py

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

forward(x)

Forward pass of the X-vector encoder.

This method processes the input tensor through the layers of the encoder, transforming the input features into frame-level x-vector embeddings. The input tensor should have the shape (B, S, D), where B is the batch size, S is the sequence length, and D is the feature dimension. The output will have the shape (B, output_size, new_S), where new_S is determined by the convolutional layers.

• Parameters:x (torch.Tensor) – Input tensor of shape (B, S, D).
• Returns: Output tensor after passing through the encoder layers, with shape (B, output_size, new_S).
• Return type: torch.Tensor

######### Examples

>>> import torch
>>> encoder = XvectorEncoder(input_size=40)
>>> input_tensor = torch.randn(10, 100, 40)  # (B, S, D)
>>> output_tensor = encoder.forward(input_tensor)
>>> print(output_tensor.shape)
torch.Size([10, 1500, new_S])  # Output shape will vary based on new_S

output_size() → int

X-vector encoder. Extracts frame-level x-vector embeddings from features.

Paper: D. Snyder et al., “X-vectors: Robust dnn embeddings for speaker recognition,” in Proc. IEEE ICASSP, 2018.

• Parameters:
  • input_size – Input feature dimension.
  • ndim – Dimensionality of the hidden representation.
  • output_size – Output embedding dimension.

layers

A list of convolutional layers, activation functions, and batch normalization layers.

• Returns: The output embedding dimension.
• Return type: int

######### Examples

Creating an instance of the XvectorEncoder

encoder = XvectorEncoder(input_size=40, ndim=512, output_size=1500)

Accessing the output size

print(encoder.output_size()) # Outputs: 1500

NOTE

This class is adapted for ESPnet-SPK by Jee-weon Jung, and it is cross-checked with the SpeechBrain implementation.