espnet.nets.beam_search_timesync_streaming.BeamSearchTimeSyncStreaming

source

class espnet.nets.beam_search_timesync_streaming.BeamSearchTimeSyncStreaming(sos: int, beam_size: int, scorers: ~typing.Dict[str, ~espnet.nets.scorer_interface.ScorerInterface], weights: ~typing.Dict[str, float], token_list=<class 'dict'>, pre_beam_ratio: float = 1.5, blank: int = 0, hold_n: int = 0)

Bases: Module

Time synchronous beam search algorithm.

Initialize beam search.

• Parameters:
  • beam_size – num hyps
  • sos – sos index
  • ctc – CTC module
  • pre_beam_ratio – pre_beam_ratio * beam_size = pre_beam pre_beam is used to select candidates from vocab to extend hypotheses
  • decoder – decoder ScorerInterface
  • ctc_weight – ctc_weight
  • blank – blank index

cached_score(h: Tuple[int], cache: dict, scorer: ScorerInterface, recompute_cache: bool = False) → Any

Retrieve decoder/LM scores which may be cached.

forward(x: Tensor, maxlenratio: float = 0.0, minlenratio: float = 0.0, start_idx: int = 0, is_final: bool = False, incremental_decode: bool = False) → List[Hypothesis]

Perform beam search.

• Parameters:enc_output (torch.Tensor)
• Returns: list[Hypothesis]

joint_score(hyps: Any, ctc_score_dp: Any, recompute_cache: bool = False) → Any

Calculate joint score for hyps.

reset(enc_output: Tensor)

Reset object for a new utterance.

time_step(p_ctc: Any, ctc_score_dp: Any, hyps: Any, recompute_cache: bool = False) → Any

Execute a single time step.