espnet.nets.pytorch_backend.transducer.error_calculator.ErrorCalculator

source

class espnet.nets.pytorch_backend.transducer.error_calculator.ErrorCalculator(decoder: RNNDecoder | CustomDecoder, joint_network: JointNetwork, token_list: List[int], sym_space: str, sym_blank: str, report_cer: bool = False, report_wer: bool = False)

Bases: object

CER and WER computation for Transducer model.

• Parameters:
  • decoder – Decoder module.
  • joint_network – Joint network module.
  • token_list – Set of unique labels.
  • sym_space – Space symbol.
  • sym_blank – Blank symbol.
  • report_cer – Whether to compute CER.
  • report_wer – Whether to compute WER.

Construct an ErrorCalculator object for Transducer model.

calculate_cer(hyps: Tensor, refs: Tensor) → float

Calculate sentence-level CER score.

• Parameters:
  • hyps – Hypotheses sequences. (B, L)
  • refs – References sequences. (B, L)
• Returns: Average sentence-level CER score.

calculate_wer(hyps: Tensor, refs: Tensor) → float

Calculate sentence-level WER score.

• Parameters:
  • hyps – Hypotheses sequences. (B, L)
  • refs – References sequences. (B, L)
• Returns: Average sentence-level WER score.

convert_to_char(hyps: Tensor, refs: Tensor) → Tuple[List, List]

Convert label ID sequences to character.

• Parameters:
  • hyps – Hypotheses sequences. (B, L)
  • refs – References sequences. (B, L)
• Returns: Character list of hypotheses. char_hyps: Character list of references.
• Return type: char_hyps