o ߥi @s8ddlZddlZddlZGdddejjZddZdS)NcsBeZdZ d dejdejdeffdd Zdejfdd ZZS) GlobalCMVNTmeanistdnorm_varcs<t|j|jks J||_|d||d|dS)z Args: mean (torch.Tensor): mean stats istd (torch.Tensor): inverse std, std which is 1.0 / std rrN)super__init__shaperregister_buffer)selfrrr __class__^/home/ubuntu/.local/lib/python3.10/site-packages/modelscope/models/audio/kws/nearfield/cmvn.pyrs  zGlobalCMVN.__init__xcCs||j}|jr ||j}|S)z Args: x (torch.Tensor): (batch, max_len, feat_dim) Returns: (torch.Tensor): normalized feature )rrr)r rr r rforward(s  zGlobalCMVN.forward)T) __name__ __module__ __qualname__torchTensorboolrr __classcell__r r r rrsrcCsd}d}t|}|}t|D]\}}|ddkrT|d}t|dks+J||d}td|d} | d} d d | D}t|t |dksSJq|d dkr|d}t|dksjJ||d}td|d} | d} d d | D}t|t |dksJq|d dkr|d}t|dksJ||d}td|d} | d}t|t |dt |dksJt|}qqWdn1swYt ||g}t |d|f}|S)z Load the kaldi format cmvn stats file and no need to calculate Args: cmvn_file: cmvn stats file in kaldi format Returns: a numpy array of [means, vars] NAddShift z [\[](.*?)[\]]rcSsg|]}dt|qS)rfloat.0sr r r Ksz#load_kaldi_cmvn..RescalecSsg|]}t|qSr rrr r rr"SsSplice) open readlines enumeratefindstripsplitlenrefindallintnparraytile) cmvn_filemeansvariancef all_linesidxlinesegs next_line means_str means_listvars_str vars_list splice_str splice_list copy_timescmvnr r rload_kaldi_cmvn6sF     $ rD)r-numpyr0rnnModulerrDr r r rs