espnet2.legacy.nets.ctc_prefix_score.CTCPrefixScoreTH

source

class espnet2.legacy.nets.ctc_prefix_score.CTCPrefixScoreTH(x, xlens, blank, eos, margin=0)

Bases: object

Batch processing of CTCPrefixScore.

which is based on Algorithm 2 in WATANABE et al. “HYBRID CTC/ATTENTION ARCHITECTURE FOR END-TO-END SPEECH RECOGNITION,” but extended to efficiently compute the label probablities for multiple hypotheses simultaneously See also Seki et al. “Vectorized Beam Search for CTC-Attention-Based Speech Recognition,” In INTERSPEECH (pp. 3825-3829), 2019.

Construct CTC prefix scorer.

• Parameters:
  • x (torch.Tensor) – input label posterior sequences (B, T, O)
  • xlens (torch.Tensor) – input lengths (B,)
  • blank (int) – blank label id
  • eos (int) – end-of-sequence id
  • margin (int) – margin parameter for windowing (0 means no windowing). NOTE: a positive margin now also windows the recursion when no attention weights are given, centred on the frame where the prefix’s own forward probability peaks. Previously a positive margin without att_w fell through to the exact full-utterance recursion, so a caller that already sets one – an espnet1-style config with a decoder that supplies no attention weights, for instance – gets approximate scores where it used to get exact ones. Set margin=0 to keep the exact behaviour.

extend_prob(x)

Extend CTC prob.

• Parameters:x (torch.Tensor) – input label posterior sequences (B, T, O)

extend_state(state)

Compute CTC prefix state.

:param state : CTC state :return ctc_state

index_select_state(state, best_ids)

Select CTC states according to best ids.

:param state : CTC state :param best_ids : index numbers selected by beam pruning (B, W) :return selected_state