espnet2.spk.pooling.stat_pooling.StatsPooling

source

class espnet2.spk.pooling.stat_pooling.StatsPooling(input_size: int = 1536)

Bases: AbsPooling

Aggregates frame-level features to single utterance-level feature.

This class implements the statistical pooling method as proposed in D. Snyder et al., “X-vectors: Robust dnn embeddings for speaker recognition”. It takes frame-level embeddings and computes the mean and standard deviation across the time dimension, concatenating them to form a single utterance-level feature.

_output_size

The dimensionality of the output feature, which is double the input size.

• Type: int

• Parameters:input_size (int) – Dimensionality of the input frame-level embeddings. This is determined by the encoder hyperparameter. The output dimensionality will be double the input_size.

• Raises:ValueError – If task_tokens is provided during the forward pass, as this pooling method is not suitable for task-specific tokens.

######### Examples

>>> pooling_layer = StatsPooling(input_size=512)
>>> frame_level_embeddings = torch.randn(10, 512)  # (time, features)
>>> output = pooling_layer.forward(frame_level_embeddings)
>>> print(output.shape)  # Output will be (10, 1024), where 1024 = 2 * 512

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

forward(x, task_tokens: Tensor | None = None)

Args: x (torch.Tensor): Input tensor of shape (batch_size, num_frames,

input_size), where num_frames is the number of frames and input_size is the dimensionality of the input embeddings.

task_tokens (torch.Tensor, optional): A tensor of task tokens. If : provided, a ValueError will be raised since this pooling method does not support task tokens.

• Returns: Output tensor of shape (batch_size, output_size), : where output_size is double the input_size. The output contains the mean and standard deviation of the input embeddings across the frame dimension.
• Return type: torch.Tensor
• Raises:ValueError – If task_tokens is not None, as this pooling method does not support task-specific tokens.

######### Examples

>>> pooling = StatsPooling(input_size=1536)
>>> input_tensor = torch.rand(10, 5, 1536)  # Batch of 10, 5 frames
>>> output = pooling.forward(input_tensor)
>>> output.shape
torch.Size([10, 3072])  # Output size is 2 * input_size

NOTE

This method aggregates frame-level features into a single utterance-level feature by calculating the mean and standard deviation across the frame dimension.

output_size()

Aggregates frame-level features to single utterance-level feature.

Proposed in D. Snyder et al., “X-vectors: Robust dnn embeddings for speaker recognition”.

input_size

Dimensionality of the input frame-level embeddings. Determined by encoder hyperparameter. For this pooling layer, the output dimensionality will be double of the input_size.

• Type: int

• Parameters:input_size – Dimensionality of the input frame-level embeddings.

• Returns: The output size, which is double the input_size.

• Return type: int

######### Examples

>>> pooling = StatsPooling(input_size=128)
>>> pooling.output_size()
256

NOTE

The output size is fixed at double the input size for this pooling layer.