o %i @sFdZddlmZddlZddlZeeZ ddd d ZdddZ dS)u CTC scoring: compute log-likelihood of reference text given audio log-probabilities. This is NOT ASR — we don't decode. We ask "how likely is this text given this audio?" ) annotationsN log_probs torch.Tensor target_tokens list[int]blank_idintreturntuple[float, float]c Cs|jd}t|}|dkrdS||krtdtdfS|}tjj|ddd}||dtj |gtj dtj |gtj dtj |gtj d}| }||} t |d t | d fS) a Compute CTC log-likelihood of target_tokens given log_probs. Args: log_probs: [T, V] log-softmax output from CTC model (on any device) target_tokens: list of token IDs for the reference text blank_id: blank token index (usually 0) Returns: (raw_score, normalized_score) where normalized = raw / len(tokens) Higher is better. Typical range: raw ~ -50 to -5, norm ~ -2.0 to -0.1 r)r z-infnoneT)blank reduction zero_infinity)dtype) shapelenfloatdetachcputorchnnCTCLoss unsqueezetensorlongitemround) rrrTS log_probs_cpu ctc_loss_fnloss raw_scorenormalized_scorer'1/home/ubuntu/transcripts/validations/ctc_score.pycompute_ctc_scores&  r) referencestr hypothesisrc Cs|s|sdSdS|s dSt|}t|}t|}t|}tt|d}td|dD];}|d}||d<td|dD])} || } ||d|| dkrW||| <ndt||| || d|| <| }q@q/|dkru|||SdS)z Compute Character Error Rate between reference and hypothesis. Returns 0.0 (perfect match) to 1.0+ (completely wrong). Used as CTC score fallback when log-probabilities aren't accessible. r g?rr)liststriprrangemin) r*r, ref_chars hyp_charsnmdpiprevjtempr'r'r(character_error_rate=s(     r:)r)rrrrrrr r )r*r+r,r+r r) __doc__ __future__rloggingr getLogger__name__loggerr)r:r'r'r'r(s   /