o oi%@sddlmZmZddlZddlZddlmZddlmZddl m Z ddl m Z edjZdd ed ed ed efd dZefd edededededed eeeeeffddZddZdddZddZdS))ListTupleN)logger)Tensor) functional)resamplefloatx window_lengthstepreturncCsz|jd||||f|jdd}g}t|D] }|||q|d|d|d||d<|||S)aReturns a tensor with chunks of overlapping windows of the first dim of x. Args: x (Tensor): Input of shape [N, B, H, W] window_length (int): Length of each window step (int): Step/hop of each window w.r.t. the original signal x Returns: windowed tensor (Tensor): Output tensor with shape [(N - window_length + step) // step, window_length, B, H, W] rr N)shaperangedimappendstrideinsert as_strided)r r r rrirE/home/ubuntu/SoloSpeech/.venv/lib/python3.10/site-packages/df/stoi.py as_windoweds(  ry dyn_rangeframelenhopepscsjd}t|d\}}||} t|| ft||| f}j\} } |}tjddjjddd ddt t | dt j ddt |} tj| ddd|| ddkfd d t| D fd d t| D d d Dfd d Dfdd t| D} fdd t| D}dd Ddd D tjjjddfdd D} fdd  D}fdd t| Dfdd Ddd t| D} dd t|D}t| D]A}ddd|fr=| ||d| |<|||d||<ddd|fr]| |d| | |<||d| ||<q| |fS)aRemove the silent frames from each signal in the batch. Note: This implementation is based on https://github.com/mpariente/pystoi The overlap add code is based on https://github.com/pytorch/pytorchaudio Args: x (Tensor): Reference signal of shape [batch, samples]. y (Tensor): Second signal of the same shape where the same frames are removed. dyn_range (int): Dynamic range / energy in [dB] to determin which frames to remove. framelen (int): Length for a frame that might be removed. hop (int): Hop between to subsequent frames. Returns: x (List[Tensor]): x without silent frames. Since each signal might have a different number of removed frames, the resulting signals will have different lengths. Thus a batch is return as list of tensors. y (List[Tensor]): y without silent frames. r Fperiodicdevicedtype)rrc.g|]}d|fd|fdqS.r# masked_selectview.0r)rmaskx_wrr Q.z(remove_silent_frames..cr%r&r'r*)rr,y_wrrr.Rr/cSsg|]}|jdqSrrr+r rrrr.Sscs"g|] }|jddqS)rr r2r3)rrrrr.Us"c g|] }tj|jdqS)r!torchzerosr!r*n_no_silr rrr.V cr4r5r6r*r9rrr.Wr;cSg|] }|dqSr1t unsqueezer3rrrr.YcSr<r1r=r+rrrrr.Zr@r!r"c g|] }tj|dqS)rFconv_transpose1dsqueezer3eyerrrr.]r;crCrDrErArIrrr.^r;cs$g|]}d|fdqS)r r)repeatr?r*) n_no_sil_wwrrr.`s$crCrDrE)r+w_rIrrr.br;cSg|]\}}||qSrr)r+r rMrrrr.ccSrOrr)r+rrMrrrr.drPN)rdivmodrFpadr>r7 hann_windowr!r"r)rlog10normnpsqrtmaxr?rrJzip)r rrrrrrR pad_frontmodpad_endB_ x_energiesx_no_sily_no_silrr) rJrrr,r:rLrMr r-r0rremove_silent_frames#sJ $$ rbcCstd||d}|dt|dd}tt|t}td||}|tdd|dd}|tdd|dd}t|t |f} tt |D]2} t t ||| } || || <| } t t ||| } || || <| } d| | | | f<qR| |fS)a{Returns the 1/3 octave band matrix and its center frequencies # Arguments : fs : sampling rate nfft : FFT size num_bands : number of 1/3 octave bands min_freq : center frequency of the lowest 1/3 octave band # Returns : obm : Octave Band Matrix cf : center frequencies # Credit: https://github.com/mpariente/pystoi rr Nrgr(?g@) rVlinspaceintarrayrastyperpowerr8lenargminsquare)fsnfft num_bandsmin_freqfkcffreq_low freq_highobmrf_binfl_iifh_iirrrthirdoctqs    ryTc Cs|dur|d}tj|d|j|jddd}||}t||d|df}tj|||||ddd}|r@||d }|S)NrFrr r#)centerreturn_complex) r7rSr!r"rFrRstftsumpowrW) r win_sizefft_sizehop_size normalizedwindowws missing_lenspecrrr_stftsrcCs|j|jks Jd|dksJd|jd}d}d}d}d}d }d } d } |jd } tj| |j|jd } t||||\} }t| |} t |||}t |||}t |||||d\}}t | D]}|| |krtt dqdt|||||dd}t|||||dd}|dd}|dd}t| |}t| |}|jd| kr|d| dddd }|d| dddd }n|d dd }|d dd }tj|dddtj|dddt}||}d| d}t||d|}||jddd}||jddd}|tj|dddt}|tj|dddt}||}|jd }|}t|||| |<qd| S)aPytorch STOI implementation. Should only used for validation/developement, use pystoi for reporting test results. Arguments: x (Tensor): Target signal y (Tensor): Degraded signal fs_source (int): Sampling rate of input signals zInputs must have the same shaperz7Expected input shape of [batch_size, samples], but got i'(ig.rrBzs8      N !