o }oi}M@sddlZddlmZddlZddlZddlmZddlZddl Z ddl m Z m Z ddl mZmZdZdedefdd Zd d efd ejd ededededejf ddZdSdejdedejfddZd d defd ejdejd ededededefddZ  dTdejdejded ed!edejf d"d#Zdejdefd$d%ZdSd&edeedefd'd(Zd)edefd*d+ZdSd,edeedefd-d.Zdejd/ejdefd0d1Z 2 2  3  4dUd5ejd6ejd7e d8e d9eed:e d;eededefdd?Z#dVdejd@edAedBeedejf dCdDZ$dVdejd@edAedBeedejf dEdFZ%dWd5ejd6ejdedejfdGdHZ& 4dXd5ejd6ejdJededejf dKdLZ'de j"de j"fdMdNZ( 4dYde j"dOee j"dPe dede j"f dQdRZ)dS)ZN)Optional) rearrangereduce)pdist squareformgpu@xreturncCst|tjS)zcUnnormalized sinc. Args: x: input value Returns: Calculates sin(x)/x )npsincpirr \/home/ubuntu/.local/lib/python3.10/site-packages/nemo/collections/audio/parts/utils/audio.pysinc_unnormalizeds r sphericali mic_positions sample_ratefield fft_lengthsound_velocityc Cs|jddks Jd|jd}|dkrtd||dd}dtj|td||}t|||f}tt|} t|D]G} d|dd| | f<t| d|D]4} | | | f} |d krpt || ||dd| | f<ntd |d |dd| | f|dd| | f<qTqB|S) aCalculate a theoretical coherence matrix for given mic positions and field type. Args: mic_positions: 3D Cartesian coordinates of microphone positions, shape (num_mics, 3) field: string denoting the type of the soundfield sample_rate: sampling rate of the input signal in Hz fft_length: length of the fft in samples sound_velocity: speed of sound in m/s Returns: Calculated coherence with shape (num_subbands, num_mics, num_mics) z!Expecting 3D microphone positionsrz+Expecting at least 2 microphones, received ?NrzUnknown noise field .) shape ValueErrorr r arangezerosrrranger) rrrrrnum_mics num_subbands angular_freqdesired_coherence mic_distancepqdist_pqr r rtheoretical_coherence)s$      " r(缉ؗҜ<Sepsc Cs0|jdkr td|j\}}}|dkrtdtjt|ddd}tj|||ftd}t |D]d}d|d d ||f<t |d|D]Q}tj|d d d d |ft |d d d d |fdd} | t |d d |f|d d |f||d d ||f<t |d d ||f|d d ||f<qCq1|S) a0Estimate complex-valued coherence for the input STFT-domain signal. Args: S: STFT of the signal with shape (num_subbands, num_frames, num_channels) eps: small regularization constant Returns: Estimated coherence with shape (num_subbands, num_channels, num_channels) rz)Expecting the input STFT to be a 3D arrayr Expecting at least 2 microphonesraxis)dtyperN) ndim RuntimeErrorrrr meanabsrcomplexr conjugatesqrt) r*r+r! num_frames num_channelspsdestimated_coherencer%r& cross_psdr r rr:Vs   <<(r:cholesky noise_signalmethodc CsJ|ddksJ|jd}|dkrtdt|||||d}t|||dS)ay Args: mic_positions: 3D microphone positions, shape (num_mics, 3) noise_signal: signal used to generate the approximate noise field, shape (num_samples, num_mics). Different channels need to be independent. sample_rate: sampling rate of the input signal field: string denoting the type of the soundfield fft_length: length of the fft in samples method: coherence decomposition method sound_velocity: speed of sound in m/s Returns: Signal with coherence approximately matching the desired coherence, shape (num_samples, num_channels) References: E.A.P. Habets, I. Cohen and S. Gannot, 'Generating nonstationary multisensor signals under a spatial coherence constraint', Journal of the Acoustical Society of America, Vol. 124, Issue 5, pp. 2911-2917, Nov. 2008. rrr,)rrrrr)signalr#r>)rrr(transform_to_match_coherence) rr=rrrr>rr r#r r r generate_approximate_noise_fieldvs rA皙?r?r# ref_channelcorrcoef_thresholdcCs|jd}|jd}|jd|ksJ|jd|ksJd|d}|tj|dd}tjt|ddd}|t||t|}t|} t| dtt| |krlt d|dt|  dt j ||d } | ddd} t | } |d krtj|dd } | dd} n.|d krtj|dd \}}t|d d d d d f|} | dd} ntd |t| dd df| | dd df<t j| dddt|d}|}|S)aTransform the input multichannel signal to match the desired coherence. Note: It's assumed that channels are independent. Args: signal: independent noise signals with shape (num_samples, num_channels) desired_coherence: desired coherence with shape (num_subbands, num_channels, num_channels) method: decomposition method used to construct the transformation matrix ref_channel: reference channel for power normalization of the input signal corrcoef_threshold: used to detect input signals with high correlation between channels Returns: Signal with coherence approximately matching the desired coherence, shape (num_samples, num_channels) References: E.A.P. Habets, I. Cohen and S. Gannot, 'Generating nonstationary multisensor signals under a spatial coherence constraint', Journal of the Acoustical Society of America, Vol. 124, Issue 5, pp. 2911-2917, Nov. 2008. rrrr-gz2Input channels are correlated above the threshold z:. Max abs off-diagonal element of the coefficient matrix: r)n_fftr<NevdzUnknown method .)length)rr r2r3r6corrcoef transpose fill_diagonalanyr1maxlibrosastft zeros_likelinalgr<swapaxeseigrmatmulistftlen)r?r#r>rCrDr8r!r signal_powercorrcoef_matrixr*XLAwVrr r rr@s:      $r@cCsttt|dS)zCalculate RMS value for the input signal. Args: x: input signal Returns: RMS of the input signal. r)r r6r2r3r r r rrmss r]magcCdt||S)zConvert magnitude ratio from linear scale to dB. Args: mag: linear magnitude value eps: small regularization constant Returns: Value in dB. r log10)r^r+r r rmag2db rcdbcCs d|dS)zConvert value in dB to linear magnitude ratio. Args: db: magnitude ratio in dB Returns: Magnitude ratio in linear scale. r`r )rer r rdb2mags rgpowercCr_)zConvert power ratio from linear scale to dB. Args: power: power ratio in linear scale eps: small regularization constant Returns: Power in dB. rfra)rhr+r r rpow2dbrdrisegmentcCsHt|t|krtdt|dt|tjj||dd}t|S)aGet starting point of `segment` in `signal`. We assume that `segment` is a sub-segment of `signal`. For example, `signal` may be a 10 second audio signal, and `segment` could be the signal between 2 seconds and 5 seconds. This function will then return the index of the sample where `segment` starts (at 2 seconds). Args: signal: numpy array with shape (num_samples,) segment: numpy array with shape (num_samples,) Returns: Index of the start of `segment` in `signal`. z4segment must be shorter than signal: len(segment) = z, len(signal) = valid)mode)rUrscipyr? correlater argmax)r?rjccr r rget_segment_start*s  rqFT:0yE>estimatetargetscale_invariantconvolution_invariantconvolution_filter_length remove_meansdr_maxc Cs|r|rtd|r|t|}|t|}|s |r(|dkr(t|||d}n |r2t||||d}tt|d}tt||d} |durW| d| d|} dt|| ||} | S)aKCalculate signal-to-distortion ratio. SDR = 10 * log10( ||t||_2^2 / (||e-t||_2^2 + alpha * ||t||^2) where alpha = 10^(-sdr_max/10) Optionally, apply scale-invariant scaling to target signal. Args: estimate: estimated signal target: target signal Returns: SDR in dB. zRArguments scale_invariant and convolution_invariant cannot be used simultaneously.r)rsrtr+)rsrt filter_lengthr+rNrf)rr r2scale_invariant_target_numpy"convolution_invariant_target_numpyr3rb) rsrtrurvrwrxryr+ target_powdistortion_powsdrr r rcalculate_sdr_numpyAs"rcCs*tjtj|jd}t||d||S)zWrap angle in radians to [-pi, pi] Args: x: angle in radians Returns: Angle in radians wrapped to [-pi, pi] )devicer)torchtensormathr r remainder)rr r r r wrap_to_piss rrzdelayn_stepsc Cs|jdkr td|j|durt|}tt||g}|t|}dkr4t|t|g}n|d|}tj |t|dt|dgS)aConstruct a causal convolutional matrix from x delayed by `delay` samples. Args: x: input signal, shape (N,) filter_length: length of the filter in samples delay: delay the signal by a number of samples n_steps: total number of time steps (rows) for the output matrix Returns: Convolutional matrix, shape (n_steps, filter_length) rz=Expecting one-dimensional signal. Received signal with shape Nr) r0rrrUr hstackrrmrPtoeplitz)rrzrrx_padpad_lenr r r convmtx_numpys  &rc Cs^|jdkr td|jg}t|jdD]}|t|dd|f|||dqt|S)aConstruct a causal multi-channel convolutional matrix from `x` delayed by `delay` samples. Args: x: input signal, shape (N, M) filter_length: length of the filter in samples delay: delay the signal by a number of samples n_steps: total number of time steps (rows) for the output matrix Returns: Multi-channel convolutional matrix, shape (n_steps, M * filter_length) rz=Expecting two-dimensional signal. Received signal with shape rN)rzrr)r0rrrappendrr r)rrzrrmc_mtxmr r rconvmtx_mc_numpys $ rcCs^|j|jkrdksJdJdt||}tt|d}|||}||S)aCalculate convolution-invariant target for a given estimated signal. Calculate scaled target obtained by solving min_scale || scale * target - estimate ||^2 Args: estimate: one-dimensional estimated signal, shape (T,) target: one-dimensional target signal, shape (T,) eps: regularization constans Returns: Scaled target signal, shape (T,) r%Only one-dimensional inputs supportedr)r0r r2r3)rsrtr+estimate_dot_targetr}scaler r rr{s ( r{ư>diag_regcCs(|j|jkrdksJdJddttt|t|d}tjj||d}tjj||d}tjjt |d|d}tjj| ||d} |d|}| d|} |durm|d||d|7<t j |} tj | | } |tjj| |d} tjj| |d} | dt|S)aCalculate convolution-invariant target for a given estimated signal. Calculate target filtered with a linear f obtained by solving min_filter || conv(filter, target) - estimate ||^2 Args: estimate: one-dimensional estimated signal target: one-dimensional target signal filter_length: length of the (convolutive) filter diag_reg: multiplicative factor for relative diagonal loading eps: absolute diagonal loading rrr)nNr)r0rceillog2rUr fftrfftirfftr3conjrmrPrsolve)rsrtrzrr+rETEtt_corrte_corrTTfiltT_filt target_filtr r rr|s($   r|cCsF|d}tj|dddfjdd|gdd}|d|djddS)zCreate Toeplitz matrix for one-dimensional signals along the last dimension. Args: x: tensor with shape (..., T) Returns: Tensor with shape (..., T, T) .rN)r)dimsdim)sizercatflipunfold)rrGr r rrs &rmasknormalize_maskcCs|jdkr td|jt|d}td||}|dur'|jdd}|S|j|jdkr;td |jd |j|j|jdd krQtd |jd |j|r^||jd d d|}|d|}t |dd}|S)aQCalculate covariance matrix of the input signal. If a mask is provided, the covariance matrix is calculated by weighting by the provided time-frequency mask and summing over the time dimension. The mask is normalized by default. If a mask is not provided, the covariance matrix is calculated by averaging over the time dimension. The provided mask can be real-valued or complex-valued, or a binary or boolean mask. Args: x: input signal with shape `(..., channel, freq, time)` mask: Time-frequency mask with shape `(..., freq, time)`. Default is `None`. normalize_mask: if `True`, normalize the mask by dividing by the sum of the mask across time. Default is `True`. eps: regularization constant. Default is `1e-10`. Returns: Covariance matrix with shape (..., freq, channel, channel) rzBInput signal must have at least 3 dimensions. Input signal shape: z... c f t -> ... f t cz...tm,...tn->...tmnNrrzwMask must have the same number of dimensions as the input signal, excluding the channel dimension. Input signal shape: z, mask shape: rzhMask must have the same shape as the input signal, excluding the channel dimension. Input signal shape: T)rkeepdim).NNz... f t m n -> ... f m nsum) r0rrrreinsumrr2rr)rrrr+p_xxr r rcovariance_matrix s(     r)r))r<rrB)FFNTNrr)rN)rr)rrr)NTrr)*rtypingrrMnumpyr numpy.typingnptrmreinopsrrscipy.spatial.distancerrSOUND_VELOCITYfloatrNDArraystrintr(r:rAr@ndarrayr]rcrgrirqboolrTensorrrrr{r|rrr r r rs   -$ 0 U     2( (" 0