iZ7>ddlmZddlmZmZddlmZddlZddlmZddl m Z ddl m Z ddl mZmZeresd gZed d ed ededej&def dZed d edededej&def dZed dededed ededej&deeeeeffdZd,dedeedefdZdededefdZd-ded edefd!Zd"ed#ed$edefd%Z d.d&ed ed'ed ededeed(ed)edefd*Z d/d ed'ed ededeed(ed)eddfd+Zy)0) lru_cache)ceilpi)OptionalN)Tensor)pad)rank_zero_warn)_GAMMATONE_AVAILABLE_TORCHAUDIO_AVAILABLE,speech_reverberation_modulation_energy_ratiod)maxsizelow_freqfs n_filtersdevicereturncddlm}d}d}d}|||||z |z||zzd|z z}tj||S)Nr) centre_freqsg<;k"@g333338@r)gammatone.filtersrtorchtensor) rrrrrear_qmin_bwordererbss w/Volumes/fast/ai/experiments/voice-extract-mac/.venv/lib/python3.12/site-packages/torchmetrics/functional/audio/srmr.py _calc_erbsr $sR. E F E "i 2U :u Dvu} TZ[^cZc dD <<V ,, num_freqscutoffcfddlm}m}||||}|||}tj||S)Nr)rmake_erb_filtersr)rrr%rr)rr"r#rrr%cfsfcoefss r_make_erb_filtersr(/s0@ r9f -C b# &F <<v ..r!min_cfmax_cfnqc ||z d|dz z z}tj|tj}||d<td|D]}||dz |z||<dtdt dtfd} tj d tz|z|z D cgc] } | | | c} d } d td tdt dtttffd } |j|}| j|} | |||\} }|| | |fScc} w)N?rdtyperw0r,rcFtj|dz }||z }tj|d| gtj}tjd|z|dzzd|dzzdz d|z |dzzgtj}tj||gdS)Nrr/rdim)rtanrfloat64stack)r1r,b0bas r_make_modulation_filterzK_compute_modulation_filterbank_and_cutoffs.._make_modulation_filterCs YYrAv  !V LL"a"U]] ; LL1r6BE>QQY]a"fr1unNV[VcVc d{{Aq6q))r!r3r4r&rcdtz|z|z }tj|dz |z }|||zdtzz z }|||zdtzz z}||fS)Nr3)rrr6)r&rr,r1r9llrrs r _calc_cutoffszA_compute_modulation_filterbank_and_cutoffs.._calc_cutoffsLsb Vc\B  YYrAv  " BGq2v& ' BGq2v& '2v r!r) rzerosr7rangerintr8rfloattupleto)r)r*r+rr,rspacing_factorr&kr<r1mfbr@r>r?s r*_compute_modulation_filterbank_and_cutoffsrJ8s# vo3!a%=9N ++au}} -C CF 1a[QUn,A*F*s*v* ++B r@QR@Q".r15@QRXY ZC6uvv~9N &&& C &&& C 3A &FB R SsDxc|jr td|%|jd}|dzrt|dz dz}|dkr tdtj j ||d}t j ||j|jd}|d zdk(rd x|d<||d z<d |d |d znd |d<d |d |d zd ztj j||zd }|d d|jdfS) Nzx must be real.rzN must be positive.)r+r5F)r0r requires_gradr3rr4.) is_complex ValueErrorshaperrfftrAr0rifft)rKr+x_ffthys r_hilbertrXZs ||~*++y GGBK r6QV r!AAv.// IIMM!qbM )E AQWWQXXUKA1uz!qay!a1f !!q1ul uqyb)A S-AGGBK-   r!wavecoefscddlm}|j\}}|j|jj |d|}|j d|jdd}|dddf}|dddf}|ddd f}|ddd f}|ddd f} |ddd df} ||| |d } || | |d } || | |d } || | | d }||j dddz S)zTranslated from gammatone package. Args: wave: shape [B, time] coefs: shape [N, 10] Returns: Tensor: shape [B, N, time] rlfilterr/rrMN )rr)rr3r_)rr_)rr_T)batching)torchaudio.functional.filteringr]rRrFr0reshapeexpand)rYrZr] num_batchtimegainas1as2as3as4bsy1y2y3y4s r_erb_filterbankrsss8jjOIt 777 % - -iD AD ;;r5;;q>2 .D A;D 9 C 9 C 9 C 9 C q!A#vB r3 .B Rt ,B Rt ,B Rt ,B  QA& &&r!energydrangec"tj|ddjddj}|jddj}|d| dz zz}tj||k||}tj||kD||S)zNormalize energy to a dynamic range of 30 dB. Args: energy: shape [B, N_filters, 8, n_frames] drange: dynamic range in dB rTr5keepdimr3r`g$@)rmeanmaxvalueswhere)rtru peak_energy min_energys r_normalize_energyrs**VD9==!T=RYYK//a/6==Kt$77J [[*,j& AF ;;v +[& AAr!bw avg_energycutoffsc|d|kr |d|kDrd}n<|d|kr |d|kDrd}n)|d|kr |d|kDrd}n|d|krd}n tdtj|ddddftj|ddd|fz S)zCalculate srmr score.rar_rbz7Something wrong with the cutoffs compared to bw values.N)rQrsum)rrrkstars r_cal_srmr_scorers bwqzB !* b !* b r RSS 99Z2A2& '%))Jq!E'z4J*K KKr!predsn_cochlear_filtersnormfastc  trts tdddlm}ddlm} t||||||||j} t| dk(r|jddn|jd| d}|j\} } tj|sI|jtjtj|j j"z }|j%j#dd j&} tj(| dkD| tj*d | j | j, }||z }d }d }|rt/dd}g}|j1j3j5}t7| D]6}||||dd||}|j9tj*|8tj:|dj|j,}nCt=||||j,}tj$t?tA||}|}tC||z}tC||z}||rdnd}tE||d|d|j,\}}}}tGd| |z |zz}tjH|dztj|j, dd}| |jKdjMdd|jdd|dddddf|dddddfdd}dt#tC| |z |z| z || z f} tO|| dd}!|!jQd||}"|"dd|ddf|zdzjSd}#|r tU|#}#tjVtY||||j,}$tjZ|#d}%tjR|%j| dd}&tjR|%d}'|'d z|&jddz }(|(j]dj_d})tj`|)d!kDj_ddk(dddf}*|$|*}+g}t7| D]'}tc|+||%||"},|j9|,)tj:|},t| dkDr|,j| ddS|,S)#aCalculate `Speech-to-Reverberation Modulation Energy Ratio`_ (SRMR). SRMR is a non-intrusive metric for speech quality and intelligibility based on a modulation spectral representation of the speech signal. This code is translated from SRMRToolbox and `SRMRpy`_. Args: preds: shape ``(..., time)`` fs: the sampling rate n_cochlear_filters: Number of filters in the acoustic filterbank low_freq: determines the frequency cutoff for the corresponding gammatone filterbank. min_cf: Center frequency in Hz of the first modulation filter. max_cf: Center frequency in Hz of the last modulation filter. If None is given, then 30 Hz will be used for `norm==False`, otherwise 128 Hz will be used. norm: Use modulation spectrum energy normalization fast: Use the faster version based on the gammatonegram. Note: this argument is inherited from `SRMRpy`_. As the translated code is based to pytorch, setting `fast=True` may slow down the speed for calculating this metric on GPU. .. hint:: Usingsing this metrics requires you to have ``gammatone`` and ``torchaudio`` installed. Either install as ``pip install torchmetrics[audio]`` or ``pip install torchaudio`` and ``pip install git+https://github.com/detly/gammatone``. .. attention:: This implementation is experimental, and might not be consistent with the matlab implementation SRMRToolbox, especially the fast implementation. The slow versions, a) ``fast=False, norm=False, max_cf=128``, b) ``fast=False, norm=True, max_cf=30``, have a relatively small inconsistency. Returns: Scalar tensor with srmr value with shape ``(...)`` Raises: ModuleNotFoundError: If ``gammatone`` or ``torchaudio`` package is not installed Example: >>> from torch import randn >>> from torchmetrics.functional.audio import speech_reverberation_modulation_energy_ratio >>> preds = randn(8000) >>> speech_reverberation_modulation_energy_ratio(preds, 8000) tensor([0.3191], dtype=torch.float64) aspeech_reverberation_modulation_energy_ratio requires you to have `gammatone` and `torchaudio>=0.10` installed. Either install as ``pip install torchmetrics[audio]`` or ``pip install torchaudio>=0.10`` and ``pip install git+https://github.com/detly/gammatone``r) fft_gtgramr\rrrr)r*rrrrMTrwr.)r0rgMb?gMb?z:`fast=True` may slow down the speed of SRMR metric on GPU.gy@g{Gz?g{Gzd?r4rNrr3)r+rr,rF)clamprcconstant)rmodevalue.r Z)r)2r r ModuleNotFoundErrorgammatone.fftweightrrdr]_srmr_arg_validaterRlenreris_floating_pointrFr7finfor0rzabsr{r|rrr detachcpunumpyrBappendr8r(rXrsrrJrChamming_window unsqueezerfrunfoldrrflipudr ryflipcumsumnonzeror)-rrrrr)r*rrrr]rRrgrhmax_valsval_norm w_length_sw_inc_smfstemppreds_npr:gt_env_bgt_envr'w_lengthw_inc_mfr num_frameswmod_outpadding mod_out_pad mod_out_framertrr total_energy ac_energyac_percac_perc_cumsum k90perc_idxrscores- rr r sn !(<! j  /7 -   KKE$'J!OEMM!R r5QS99UEkkOIt  " "5 )'%++ekk*B*F*FFyy{2t4;;H{{1  SxGH H EJG ST<<>%%'--/y!A!(1+r5&BTV^_H KK X. /"Tq),,ELL,A"2'98ELLY8OE6$BCDJ$%H 3 E~B!qAr7AQ$/e334J X\u||TUXVXYA##BBHHQKBGL *!,I#o 4 4R ;;G\\"%,,R0N--"!4 < @)}raNFF)rrrarFF) functoolsrmathrrtypingrrrtorch.nn.functionalrtorchmetrics.utilitiesr torchmetrics.utilities.importsr r __doctest_skip__rDrCrr r(rErJrXrsrrrr rr!rrs$  #1 $8FG 3--C-C--RX-- 3/#/#/u/ell/W]// 3  %(.38;EJ\\ 6666 )*B!!8C=!F!2'&''F'> Bf Be Bv B L LF LV L L$!"RC RC RCRC RC  RC UO RC RC RC RCn!!$H $H$H$H  $H UO $H  $H $H $Hr!