o ei@sZdZddlZddlZddlZddlmZeedjZ ddZ dd d Z dd dZ dS)zLibrary for computing STOI computation. Reference: "End-to-End Waveform Utterance Enhancement for Direct Evaluation Metrics Optimization by Fully Convolutional Neural Networks", TASLP, 2018 Authors: Szu-Wei, Fu 2020 N)check_torchaudio_backendfloatcCstd||d}|dt|dd}ttt|t}t d||}|t dd|dd}|t dd|dd}t |t |} tt |D]2} t t ||| } || || <| } t t ||| } || || <| } d| | | | f<qT| S)aSReturns the 1/3 octave band matrix. Arguments --------- fs : int Sampling rate. nfft : int FFT size. num_bands : int Number of 1/3 octave bands. min_freq : int Center frequency of the lowest 1/3 octave band. Returns ------- obm : tensor Octave Band Matrix. rNgr(?g@)torchlinspaceint from_numpynparrayrangeastyperpowzeroslenargminsquare)fsnfft num_bandsmin_freqfkcffreq_low freq_highobmif_binfl_iifh_iir"]/home/ubuntu/transcripts/venv/lib/python3.10/site-packages/speechbrain/nnet/loss/stoi_loss.pythirdocts   r$(c Csttt|dtj}|dt|jd|| t|jd||j }||t|jd|||| t|jd|||j }t ||jd|jd}||dddddf<||dddddf<dt t t|d|ddt} t| } t| | |dk} |dt|jd|| t|jd||j } ||t|jd|||| t|jd|||j } t || jd| jd}| |dddddf<| |dddddf<|j d|dd| fjd|dd| f}|j d|dd| fjd|dd| f}tj|d|df|d|ddf||dddfj |||dffdd}tj|d|df|d|ddf||dddfj |||dffdd}||gS) aRemoves silent frames from the STOI computation. This function can be used as a loss function for training with SGD-based updates. Arguments --------- x: torch.Tensor The clean (reference) waveforms. y: torch.Tensor The degraded (enhanced) waveforms. dyn_range: int Dynamic range used for mask computation. N: int Window length. K: int Step size. Returns ------- list with 2 elements, x and y with silence removed. rrNrg0@)axis)r unsqueezer r hanningtorr shapereshapeTrlog10sqrtmatmulsmallValmaxsqueezerepeatcatflatten)xy dyn_rangeNKwX1X2Xenergy Max_energymskY1Y2Yx_sily_silr"r"r#removeSilentFrames<sN 6$  6$00,, rKmeancCstj|dd}tj|dd}|jd}d}d}d}tddd d d }d } t|} tj|d|j } t d|D]} || dt || |jd f} || dt || |jd f}| | | |} }t | |\}}tjj ddddd|}tjj ddddd|}tt||d}tt||d}t |jd|d }td |d}td |d}t d|D]6}|dd|||f||d |d d ddf<|dd|||f||d |d d ddf<qtj|dddtj|dddt}||}t|||| }|tj|ddd}|tj|dddt}|tj|ddd}|tj|dddt}t||}|||| | <q8|dkrX|  S| S)aCompute the STOI score and return -1 * that score. This function can be used as a loss function for training with SGD-based updates. Arguments --------- y_pred_batch : torch.Tensor The degraded (enhanced) waveforms. y_true_batch : torch.Tensor The clean (reference) waveforms. lens : torch.Tensor The relative lengths of the waveforms within the batch. reduction : str The type of reduction ("mean" or "batch") to use. Returns ------- The computed STOI loss. Example ------- >>> a = torch.sin(torch.arange(16000, dtype=torch.float32)).unsqueeze(0) >>> b = a + 0.001 >>> -stoi_loss(b, a, torch.ones(1)) tensor(0.7...) r))dimri>g.@i'i)rrrrg `~@rr&r'r)n_fft win_length hop_lengthpowerg+=NT)rMkeepdimrL)rr6r.r$r torchaudio transformsResampler-devicer r rK Spectrogramr2r3normr4minrLsum) y_pred_batch y_true_batchlens reduction batch_sizerr=J octave_bandcD resamplerry_truey_pred y_sil_true y_sil_pred stft_true stft_predOCT_trueOCT_predMrBrHmalphaayr;xnyndr"r"r# stoi_losssf  ""46  rw)r%r&r')rL) __doc__numpyr rrV%speechbrain.utils.torch_audio_backendrfinfoepsr4r$rKrwr"r"r"r#s   )K