o .wiZ7@sddlmZddlmZmZddlmZddlZddlmZddl m Z ddl m Z ddl mZmZer6es9d gZed d d ed ededejdef ddZed d d edededejdef ddZed d dededed ededejdeeeeeffddZd:dedeedefddZd ed!edefd"d#Zd;d%ed&edefd'd(Zd)ed*ed+edefd,d-Z . / 0  1 1d) lru_cache)ceilpi)OptionalN)Tensor)pad)rank_zero_warn)_GAMMATONE_AVAILABLE_TORCHAUDIO_AVAILABLE,speech_reverberation_modulation_energy_ratiod)maxsizelow_freqfs n_filtersdevicereturnc CsJddlm}d}d}d}||||||||d|}tj||dS)Nr) centre_freqsg<;k"@g333338@r)gammatone.filtersrtorchtensor) rrrrrear_qmin_bwordererbsr_/home/ubuntu/sommelier/.venv/lib/python3.10/site-packages/torchmetrics/functional/audio/srmr.py _calc_erbs$s $r num_freqscutoffcCs4ddlm}m}||||}|||}tj||dS)Nr)rmake_erb_filtersr)rrr"rr)rr r!rrr"cfsfcoefsrrr_make_erb_filters/s  r%min_cfmax_cfnqc s||d|d}tj|tjd}||d<td|D] }||d|||<qdtdtdtfdd tjfd d d t||Ddd } dtdtdtdt ttffdd} |j |d}| j |d} | ||\} } || | | fS)N?rdtyperw0r)rcSszt|d}||}tj|d| gtjd}tjd||dd|ddd||dgtjd}tj||gddS)Nrr+rdim)rtanrfloat64stack)r-r)b0barrr_make_modulation_filterCs :zK_compute_modulation_filterbank_and_cutoffs.._make_modulation_filtercsg|]}|qSrr).0r-r7r)rr Jsz>_compute_modulation_filterbank_and_cutoffs..r.r/r#rcSsRdt||}t|d|}|||dt}|||dt}||fS)Nr.)rrr1)r#rr)r-r4llrrrrr _calc_cutoffsLs zA_compute_modulation_filterbank_and_cutoffs.._calc_cutoffsr) rzerosr2rangerintr3rfloattupleto) r&r'r(rr)rspacing_factorr#kmfbr=r;r<rr9r*_compute_modulation_filterbank_and_cutoffs8s*"   rGxcCs|rtd|dur|jd}|drt|dd}|dkr%tdtjj||dd}tj||j|jdd}|d dkrRd |d<||d <d |d |d <nd |d<d |d |d d <tjj ||dd }|d d|jdfS) Nzx must be real.rzN must be positive.)r(r0F)r,r requires_gradr.rr/.) is_complex ValueErrorshaperrfftr>r,rifft)rHr(x_ffthyrrr_hilbertZs"  rTwavecoefscCsddlm}|j\}}|j|jd|d|}|d|jdd}|dddf}|dddf}|ddd f}|ddd f}|ddd f} |ddd df} ||| |d d} || | |d d} || | |d d} || | | d d}||dddS)zTranslated from gammatone package. Args: wave: shape [B, time] coefs: shape [N, 10] Returns: Tensor: shape [B, N, time] rlfilterr+rrIN )rr)rr.rZ)rrZ)rrZT)batching)torchaudio.functional.filteringrXrNrCr,reshapeexpand)rUrVrX num_batchtimegainas1as2as3as4bsy1y2y3y4rrr_erb_filterbankss rn>@energydrangecCsbtj|dddjdddj}|jdddj}|d| d}t||k||}t||k||S)zNormalize energy to a dynamic range of 30 dB. Args: energy: shape [B, N_filters, 8, n_frames] drange: dynamic range in dB rTr0keepdimr.r[g$@)rmeanmaxvalueswhere)rprq peak_energy min_energyrrr_normalize_energys rzbw avg_energycutoffscCs|d|kr|d|krd}n+|d|kr|d|krd}n|d|kr-|d|kr-d}n |d|kr6d}ntdt|ddddft|ddd|fS)zCalculate srmr score.r\rZr]z7Something wrong with the cutoffs compared to bw values.N)rMrsum)r{r|r}kstarrrr_cal_srmr_scores 4r}r\Fpredsn_cochlear_filtersnormfastc- Cstrtstdddlm}ddlm} t|||||||d|j} t | dkr.| ddn| d| d}|j\} } t |sM| t jt |jj}|jddd j} t | dk| t jd | j| jd }||}d }d }|rtdd}g}|}t| D]}||||dd||}|t |qt j|ddj |jd}nt||||jd}t tt ||}|}t!||}t!||}|dur|rdnd}t"||d|d|jd\}}}}t#d| ||}t j$|dt j|jd dd}| |%d&dd|jdd|dddddf|dddddfddd}dtt!| ||| || f} t'|| ddd}!|!(d||}"|"dd|ddf|dj)dd}#|r^t*|#}#t +t,||||jd}$t j-|#dd}%t j)|% | ddd}&t j)|%dd}'|'d |& dd}(|(.d/d})t 0|)d!k/ddkdddf}*|$|*}+g}t| D]}t1|+||%||d"},||,qt |},t | dkr|,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_gtgramrWrrrr&r'rrrrITrrr*)r,rgMb?gMb?z:`fast=True` may slow down the speed of SRMR metric on GPU.gy@g{Gz?g{Gzd?r/rNrr.)r(rr)rF)clampr^constant)rmodevalue.r Z)r})2r r ModuleNotFoundErrorgammatone.fftweightrr_rX_srmr_arg_validaterNlenr`ris_floating_pointrCr2finfor,ruabsrvrwrrrdetachcpunumpyr?appendr3r%rTrnrrGr@hamming_window unsqueezerarunfoldrrzflipudrrtflipcumsumnonzeror)-rrrrr&r'rrrrXrNrbrcmax_valsval_norm w_length_sw_inc_smfstemppreds_npr5gt_env_bgt_envr$w_lengthw_inc_mfr} num_frameswmod_outpadding mod_out_pad mod_out_framerprr| total_energy ac_energyac_percac_perc_cumsum k90perc_idxr{scorerrrr s7   (       F"&$  $rcCst|tr |dkstd|t|tr|dks td|t|ttfr+|dks2td|t|ttfr=|dksDtd||durZt|ttfrS|dksZtd|t|tsctdt|tsltd dS) a9Validate the arguments for speech_reverberation_modulation_energy_ratio. Args: 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, norm: Use modulation spectrum energy normalization fast: Use the faster version based on the gammatonegram. rz;Expected argument `fs` to be an int larger than 0, but got zKExpected argument `n_cochlear_filters` to be an int larger than 0, but got zBExpected argument `low_freq` to be a float larger than 0, but got z@Expected argument `min_cf` to be a float larger than 0, but got Nz@Expected argument `max_cf` to be a float larger than 0, but got z+Expected argument `norm` to be a bool valuez+Expected argument `fast` to be a bool value) isinstancer@rMrAboolrrrrrEs"  r)N)ro)rrr\NFF)rrr\rFF) functoolsrmathrrtypingrrrtorch.nn.functionalrtorchmetrics.utilitiesrtorchmetrics.utilities.importsr r __doctest_skip__rAr@rrr%rBrGrTrnrzrrr rrrrrs     " " !