espnet2.tok.output.SpeechTokenizerOutput

source

class espnet2.tok.output.SpeechTokenizerOutput(continuous: Tensor, assignment: Tensor, token_ids: Tensor, lengths: Tensor, soft_assignment: Tensor | None = None, quantized_features: Tensor | None = None)

Bases: object

Outputs of a trainable speech tokenizer.

• Parameters:
  • continuous – Continuous frontend features of shape (B, T, D).
  • assignment – Hard straight-through cluster assignments of shape (B, T, K). The forward values are one-hot during training, while gradients propagate through their soft counterparts.
  • token_ids – Integer cluster indices of shape (B, T).
  • lengths – Valid token lengths of shape (B,).
  • soft_assignment – Optional soft cluster assignments of shape (B, T, K) for analysis and auxiliary regularization.
  • quantized_features – Optional differentiable quantized features of shape (B, T, D). A future VQ implementation can expose its straight-through vectors here so downstream losses can reach the frontend.

assignment : Tensor

continuous : Tensor

lengths : Tensor

quantized_features : Tensor | None = None

soft_assignment : Tensor | None = None

token_ids : Tensor