U %d:@sddlZddlZddlZddlmZddlmZmZmZm Z ddl Z ddl Z ddl m Z ddl mZdddd d d d d ddd dd dddddddddddddddd d!d"gZde ee eeeeeeeeeeee d% d&dZde eeee eeeeeeee d' d(dZe jd)d*d+Ze e eeeeeeeeee d, d-dZde eeeeee d.d/d Ze eee d0d1d Zdeeed3d4d5Zde ee d6d7d8Ze e e d9d:d;Zdeeeeeeeee d<d=d Zeeeeee d>d?dZ eeeee d@dAdZ!e ee dBdCdZ"e ee dDdEdZ#e ee e dFdGdZ$eeeedHdIdJZ%de eeeee dLdMdZ&de eeeee dNdOdZ'de eee dRdSd Z(e eeee dTdUdVZ)dee e fee e feee e fdXdYdZZ*e eee d[d\d]Z+e ee d^d_d`Z,de eeeeee dedfdZ-de eeeee djdkdZ.e eee eeee dldmdZ/e0de eeeeeeeeee dndodZ1e2de j eeeeeeeeeeeeeeeeeee j d{d|d Z3d}dWd~de 4ddfeeeeeeeee j4ee jd ddZ5e eeee edddZ6de eeeeeeee dddZ7e j8j9eeedddZ:de eeeeeeeeee e d ddZ;de eeeeeeeee e dddZde ee eee dddZ?de j eee j dddZ@de j eee j dddZAe j e j ddddZBde e eee feeee dddZCde e eeee feeee dddZDe e ddd ZEde e eee feeee ddd!ZFe e e ddd"ZGdS)N)Sequence)OptionalTupleUnionList)Tensor) module_utils spectrograminverse_spectrogram griffinlimamplitude_to_DBDB_to_amplitudecompute_deltascompute_kaldi_pitchmelscale_fbanks linear_fbanks create_dctdetect_pitch_frequencymu_law_encodingmu_law_decoding phase_vocodermask_along_axismask_along_axis_iidsliding_window_cmnspectral_centroid apply_codecresample edit_distance pitch_shift rnnt_losspsdmvdr_weights_soudenmvdr_weights_rtfrtf_evd rtf_powerapply_beamformingTreflect) waveformpadwindown_fft hop_length win_lengthpower normalizedcenterpad_modeonesidedreturn_complexreturnc  Cs| dk rtd|dkr0tjj|||fd}|} |d| d}tj||||||| d| dd } | | dd| j d d} |r| | d  } |dk r|d kr| S| |S| S) a-Create a spectrogram or a batch of spectrograms from a raw audio signal. The spectrogram can be either magnitude-only or complex. .. devices:: CPU CUDA .. properties:: Autograd TorchScript Args: waveform (Tensor): Tensor of audio of dimension `(..., time)` pad (int): Two sided padding of signal window (Tensor): Window tensor that is applied/multiplied to each frame/window n_fft (int): Size of FFT hop_length (int): Length of hop between STFT windows win_length (int): Window size power (float or None): Exponent for the magnitude spectrogram, (must be > 0) e.g., 1 for energy, 2 for power, etc. If None, then the complex spectrum is returned instead. normalized (bool): Whether to normalize by magnitude after stft center (bool, optional): whether to pad :attr:`waveform` on both sides so that the :math:`t`-th frame is centered at time :math:`t \times \text{hop\_length}`. Default: ``True`` pad_mode (string, optional): controls the padding method used when :attr:`center` is ``True``. Default: ``"reflect"`` onesided (bool, optional): controls whether to return half of results to avoid redundancy. Default: ``True`` return_complex (bool, optional): Deprecated and not used. Returns: Tensor: Dimension `(..., freq, time)`, freq is ``n_fft // 2 + 1`` and ``n_fft`` is the number of Fourier bins, and time is the number of window hops (n_frame). Nz`return_complex` argument is now deprecated and is not effective.`torchaudio.functional.spectrogram(power=None)` always returns a tensor with complex dtype. Please remove the argument in the function call.rconstantFT inputr*r+r,r)r/r0r.r1r2@?)warningswarntorchnn functionalr(sizereshapestftshapepowsumsqrtabs)r'r(r)r*r+r,r-r.r/r0r1r2rCspec_frID/tmp/pip-unpacked-wheel-lbdmvq91/torchaudio/functional/functional.pyr 1s8/ ) r lengthr(r)r*r+r,r.r/r0r1r3c Cs|std|r*||d}|} |d| d| d}tj||||||d| |dk rp|d|nddd } |dk r|d kr| dd|| f} | | dd| j dd} | S) aCreate an inverse spectrogram or a batch of inverse spectrograms from the provided complex-valued spectrogram. .. devices:: CPU CUDA .. properties:: Autograd TorchScript Args: spectrogram (Tensor): Complex tensor of audio of dimension (..., freq, time). length (int or None): The output length of the waveform. pad (int): Two sided padding of signal. It is only effective when ``length`` is provided. window (Tensor): Window tensor that is applied/multiplied to each frame/window n_fft (int): Size of FFT hop_length (int): Length of hop between STFT windows win_length (int): Window size normalized (bool): Whether the stft output was normalized by magnitude center (bool, optional): whether the waveform was padded on both sides so that the :math:`t`-th frame is centered at time :math:`t \times \text{hop\_length}`. Default: ``True`` pad_mode (string, optional): controls the padding method used when :attr:`center` is ``True``. This parameter is provided for compatibility with the spectrogram function and is not used. Default: ``"reflect"`` onesided (bool, optional): controls whether spectrogram was done in onesided mode. Default: ``True`` Returns: Tensor: Dimension `(..., time)`. Least squares estimation of the original signal. z+Expected `spectrogram` to be complex dtype.r9r5r8FN) r7r*r+r,r)r/r.r1rKr2r) is_complex ValueErrorrDrErFr@rAr=istftrC) r rKr(r)r*r+r,r.r/r0r1rCr'rIrIrJr s,*   real_dtypecCsB|tjkrtjS|tjkr tjS|tjkr0tjStd|dS)NzUnexpected dtype )r=doubleZcdoublefloatZcfloathalfZ complex32rNrPrIrIrJ_get_complex_dtypes   rU) specgramr)r*r+r,r-n_itermomentumrK rand_initr3c  Csn|dkstd||dks,td||} |dgt| dd}|d|}| rtj|t|j |j d} ntj |dt|j |j d} tj d |j |j d} t |D]t} tj|| |||||d }tj|||||d d d d d d }|} |r| | |d|} | | d} |} qtj|| |||||d }|| dd|jdd}|S)aCompute waveform from a linear scale magnitude spectrogram using the Griffin-Lim transformation. .. devices:: CPU CUDA .. properties:: Autograd TorchScript Implementation ported from *librosa* [:footcite:`brian_mcfee-proc-scipy-2015`], *A fast Griffin-Lim algorithm* [:footcite:`6701851`] and *Signal estimation from modified short-time Fourier transform* [:footcite:`1172092`]. Args: specgram (Tensor): A magnitude-only STFT spectrogram of dimension `(..., freq, frames)` where freq is ``n_fft // 2 + 1``. window (Tensor): Window tensor that is applied/multiplied to each frame/window n_fft (int): Size of FFT, creates ``n_fft // 2 + 1`` bins hop_length (int): Length of hop between STFT windows. ( Default: ``win_length // 2``) win_length (int): Window size. (Default: ``n_fft``) power (float): Exponent for the magnitude spectrogram, (must be > 0) e.g., 1 for energy, 2 for power, etc. n_iter (int): Number of iteration for phase recovery process. momentum (float): The momentum parameter for fast Griffin-Lim. Setting this to 0 recovers the original Griffin-Lim method. Values near 1 can lead to faster convergence, but above 1 may not converge. length (int or None): Array length of the expected output. rand_init (bool): Initializes phase randomly if True, to zero otherwise. Returns: Tensor: waveform of `(..., time)`, where time equals the ``length`` parameter if given. zmomentum={} > 1 can be unstablerzmomentum={} < 0r5r8Ndtypedevicer*r+r,r)rKTr&Fr6gؗҜ<)AssertionErrorformatr@rAlistrDr=randrUr\r]fulltensorrangerOrBmul_divrGaddrC)rVr)r*r+r,r-rWrXrKrYrCZanglesZtprev_ZinverseZrebuiltr'rIrIrJr sZ*  )x multiplieramin db_multipliertop_dbr3c Cs|ttj||d}|||8}|dk r|}|dkrH|dnd}|d||d|d}t||jdd |dddd}||}|S) aTurn a spectrogram from the power/amplitude scale to the decibel scale. .. devices:: CPU CUDA .. properties:: Autograd TorchScript The output of each tensor in a batch depends on the maximum value of that tensor, and so may return different values for an audio clip split into snippets vs. a full clip. Args: x (Tensor): Input spectrogram(s) before being converted to decibel scale. Input should take the form `(..., freq, time)`. Batched inputs should include a channel dimension and have the form `(batch, channel, freq, time)`. multiplier (float): Use 10. for power and 20. for amplitude amin (float): Number to clamp ``x`` db_multiplier (float): Log10(max(reference value and amin)) top_db (float or None, optional): Minimum negative cut-off in decibels. A reasonable number is 80. (Default: ``None``) Returns: Tensor: Output tensor in decibel scale )minNrLrZr5r8)rqr8r5dim) r=log10clampr@rsrAmaxZamaxview)rkrlrmrnroZx_dbrCZpacked_channelsrIrIrJr Bs $ )rkrefr-r3cCs|ttdd||S)aTurn a tensor from the decibel scale to the power/amplitude scale. .. devices:: CPU CUDA .. properties:: TorchScript Args: x (Tensor): Input tensor before being converted to power/amplitude scale. ref (float): Reference which the output will be scaled by. power (float): If power equals 1, will compute DB to power. If 0.5, will compute DB to amplitude. Returns: Tensor: Output tensor in power/amplitude scale. $@皙?)r=rD)rkrxr-rIrIrJr mshtk)freq mel_scaler3cCs|dkrtd|dkr.dtd|dSd}d}|||}d }|||}td d }||kr~|t|||}|S) zConvert Hz to Mels. Args: freqs (float): Frequencies in Hz mel_scale (str, optional): Scale to use: ``htk`` or ``slaney``. (Default: ``htk``) Returns: mels (float): Frequency in Mels slaneyr{-mel_scale should be one of "htk" or "slaney".r{F@r:@r^竪P@@@皙@;@)rNmathrtlog)r|r}f_minf_spmels min_log_hz min_log_mellogsteprIrIrJ _hz_to_mels   r)rr}r3c Cs|dkrtd|dkr,dd|ddSd}d }|||}d }|||}td d }||k}|t||||||<|S) zConvert mel bin numbers to frequencies. Args: mels (Tensor): Mel frequencies mel_scale (str, optional): Scale to use: ``htk`` or ``slaney``. (Default: ``htk``) Returns: freqs (Tensor): Mels converted in Hz r~rr{rryrr:r^rrrr)rNrrr=exp) rr}rrfreqsrrrZlog_trIrIrJ _mel_to_hzs   r) all_freqsf_ptsr3cCs|dd|dd}|d|d}td}d|ddddf|dd}|ddddf|dd}t|t||}|S)aCreate a triangular filter bank. Args: all_freqs (Tensor): STFT freq points of size (`n_freqs`). f_pts (Tensor): Filter mid points of size (`n_filter`). Returns: fb (Tensor): The filter bank of size (`n_freqs`, `n_filter`). rZNr5rgr8rL) unsqueezer=zerosrvrp)rrZf_diffZslopesZzeroZ down_slopesZ up_slopesfbrIrIrJ_create_triangular_filterbanks $ r)n_freqsrf_maxn_mels sample_ratenormr}r3cCs|dk r|dkrtdtd|d|}t||d}t||d} t|| |d} t| |d} t|| } |dk r|dkrd| d|d| d|} | | d9} | jddjd k rt d |d |d | S) a"Create a frequency bin conversion matrix. .. devices:: CPU .. properties:: TorchScript Note: For the sake of the numerical compatibility with librosa, not all the coefficients in the resulting filter bank has magnitude of 1. .. image:: https://download.pytorch.org/torchaudio/doc-assets/mel_fbanks.png :alt: Visualization of generated filter bank Args: n_freqs (int): Number of frequencies to highlight/apply f_min (float): Minimum frequency (Hz) f_max (float): Maximum frequency (Hz) n_mels (int): Number of mel filterbanks sample_rate (int): Sample rate of the audio waveform norm (str or None, optional): If 'slaney', divide the triangular mel weights by the width of the mel band (area normalization). (Default: ``None``) mel_scale (str, optional): Scale to use: ``htk`` or ``slaney``. (Default: ``htk``) Returns: Tensor: Triangular filter banks (fb matrix) of size (``n_freqs``, ``n_mels``) meaning number of frequencies to highlight/apply to x the number of filterbanks. Each column is a filterbank so that assuming there is a matrix A of size (..., ``n_freqs``), the applied result would be ``A * melscale_fbanks(A.size(-1), ...)``. Nrz$norm must be one of None or 'slaney'rrL)r}r9rrr^zIAt least one mel filterbank has all zero values. The value for `n_mels` (z4) may be set too high. Or, the value for `n_freqs` (z) may be set too low.) rNr=linspacerrrrrvvaluesanyr;r<)rrrrrrr}rZm_minZm_maxZm_ptsrrZenormrIrIrJrs )     )rrrn_filterrr3cCs2td|d|}t|||d}t||}|S)a)Creates a linear triangular filterbank. .. devices:: CPU .. properties:: TorchScript Note: For the sake of the numerical compatibility with librosa, not all the coefficients in the resulting filter bank has magnitude of 1. .. image:: https://download.pytorch.org/torchaudio/doc-assets/lin_fbanks.png :alt: Visualization of generated filter bank Args: n_freqs (int): Number of frequencies to highlight/apply f_min (float): Minimum frequency (Hz) f_max (float): Maximum frequency (Hz) n_filter (int): Number of (linear) triangular filter sample_rate (int): Sample rate of the audio waveform Returns: Tensor: Triangular filter banks (fb matrix) of size (``n_freqs``, ``n_filter``) meaning number of frequencies to highlight/apply to x the number of filterbanks. Each column is a filterbank so that assuming there is a matrix A of size (..., ``n_freqs``), the applied result would be ``A * linear_fbanks(A.size(-1), ...)``. rrL)r=rr)rrrrrrrrrIrIrJr$s# )n_mfccrrr3cCstt|}tt|d}ttjt||d|}|dkrT|d9}n<|dks`t|ddtd9<|tdt|9}| S)aCreate a DCT transformation matrix with shape (``n_mels``, ``n_mfcc``), normalized depending on norm. .. devices:: CPU .. properties:: TorchScript Args: n_mfcc (int): Number of mfc coefficients to retain n_mels (int): Number of mel filterbanks norm (str or None): Norm to use (either 'ortho' or None) Returns: Tensor: The transformation matrix, to be right-multiplied to row-wise data of size (``n_mels``, ``n_mfcc``). rZ?Nr9Zorthorr:) r=arangerSrcosrpir`rFt)rrrnkdctrIrIrJrRs   )rkquantization_channelsr3cCs~|d}|s&td|tj}tj||jd}t|t |t |t |}|dd|dtj }|S)a Encode signal based on mu-law companding. .. devices:: CPU CUDA .. properties:: TorchScript For more info see the `Wikipedia Entry `_ This algorithm expects the signal has been scaled to between -1 and 1 and returns a signal encoded with values from 0 to quantization_channels - 1. Args: x (Tensor): Input tensor quantization_channels (int): Number of channels Returns: Tensor: Input after mu-law encoding r:zbThe input Tensor must be of floating type. This will be an error in the v0.12 release.r\rZrLr) is_floating_pointr;r<tor=rSrer\signlog1prGZint64)rkrmux_murIrIrJrps ()rrr3cCsl|d}|s|tj}tj||jd}||dd}t|tt|t |d|}|S)aDecode mu-law encoded signal. .. devices:: CPU CUDA .. properties:: TorchScript For more info see the `Wikipedia Entry `_ This expects an input with values between 0 and quantization_channels - 1 and returns a signal scaled between -1 and 1. Args: x_mu (Tensor): Input tensor quantization_channels (int): Number of channels Returns: Tensor: Input after mu-law decoding r:rrL) rrr=rSrer\rrrGr)rrrrkrIrIrJrs ,)complex_specgramsrate phase_advancer3cCsp|dkr |S|}|dgt|dd}t|j}tjd|d||j|d}|d}|dddf}tj j |dd g}| d| }| d|d } |} | } |} | } | | |}|d tjt|d tj}||}tj||dddfgdd }t|d}|| d|| }t||}||dd|jdd}|S) aFGiven a STFT tensor, speed up in time without modifying pitch by a factor of ``rate``. .. devices:: CPU CUDA .. properties:: Autograd TorchScript Args: complex_specgrams (Tensor): A tensor of dimension `(..., freq, num_frame)` with complex dtype. rate (float): Speed-up factor phase_advance (Tensor): Expected phase advance in each bin. Dimension of `(freq, 1)` Returns: Tensor: Stretched spectrogram. The resulting tensor is of the same dtype as the input spectrogram, but the number of frames is changed to ``ceil(num_frame / rate)``. Example >>> freq, hop_length = 1025, 512 >>> # (channel, freq, time) >>> complex_specgrams = torch.randn(2, freq, 300, dtype=torch.cfloat) >>> rate = 1.3 # Speed up by 30% >>> phase_advance = torch.linspace( >>> 0, math.pi * hop_length, freq)[..., None] >>> x = phase_vocoder(complex_specgrams, rate, phase_advance) >>> x.shape # with 231 == ceil(300 / 1.3) torch.Size([2, 1025, 231]) r:r5r8Nrr]r\.rZrLrr)r@rArbr=realr\rr]Zangler>r?r(Z index_selectlongrGrrroundcatcumsumZpolarrC)rrrrCrQZ time_stepsalphasZphase_0Zcomplex_specgrams_0Zcomplex_specgrams_1Zangle_0Zangle_1Znorm_0Znorm_1phaseZ phase_accZmagZcomplex_specgrams_stretchrIrIrJrs0  "   ) mask_paramp axis_lengthr3cCs"|dkr |St|t||SdS)Nr:)rpint)rrrrIrIrJ_get_mask_paramsrr:) specgramsr mask_valueaxisrr3c Cs|dkrtdd|kr$dks6ntd|dt|||j|}|dkrT|S|j}|j}tj|jdd ||d |}tj|jdd ||d |||}|d } ||d } tj d ||||d } | |d }| | | k| | k@|}| |d }|S)aApply a mask along ``axis``. .. devices:: CPU CUDA .. properties:: Autograd TorchScript Mask will be applied from indices ``[v_0, v_0 + v)``, where ``v`` is sampled from ``uniform(0, max_v)`` and ``v_0`` from ``uniform(0, specgrams.size(axis) - v)``, with ``max_v = mask_param`` when ``p = 1.0`` and ``max_v = min(mask_param, floor(specgrams.size(axis) * p))`` otherwise. Args: specgrams (Tensor): Real spectrograms `(batch, channel, freq, time)` mask_param (int): Number of columns to be masked will be uniformly sampled from [0, mask_param] mask_value (float): Value to assign to the masked columns axis (int): Axis to apply masking on (2 -> frequency, 3 -> time) p (float, optional): maximum proportion of columns that can be masked. (Default: 1.0) Returns: Tensor: Masked spectrograms of dimensions `(batch, channel, freq, time)` )rL-Only Frequency and Time masking are supportedr^r:,The value of p must be between 0.0 and 1.0 ( given).rZNrLr.NNrr5) rNrrCr]r\r=rcr@rr transpose masked_fill) rrrrrr]r\value min_value mask_startmask_endmaskrIrIrJrs$(   )rVrrrrr3c Cs>|dkrtdd|kr$dks6ntd|dt|||j|}|dkrT|S|}|dgt|d d }td|}td|||}| }|| } tj d |j||j |j d } | |k| | k@} |dkr| d} | ||kst|| |}||d d |jd d }|S) aApply a mask along ``axis``. .. devices:: CPU CUDA .. properties:: Autograd TorchScript Mask will be applied from indices ``[v_0, v_0 + v)``, where ``v`` is sampled from ``uniform(0, max_v)`` and ``v_0`` from ``uniform(0, specgrams.size(axis) - v)``, with ``max_v = mask_param`` when ``p = 1.0`` and ``max_v = min(mask_param, floor(specgrams.size(axis) * p))`` otherwise. All examples will have the same mask interval. Args: specgram (Tensor): Real spectrogram `(channel, freq, time)` mask_param (int): Number of columns to be masked will be uniformly sampled from [0, mask_param] mask_value (float): Value to assign to the masked columns axis (int): Axis to apply masking on (1 -> frequency, 2 -> time) p (float, optional): maximum proportion of columns that can be masked. (Default: 1.0) Returns: Tensor: Masked spectrogram of dimensions `(channel, freq, time)` )rZrLrr^r:rrrZr5r8Nrr)rNrrCr@rArbr=rcrsqueezerr]r\rr`r) rVrrrrrCrrrrrrIrIrJr:s*     replicate)rVr,moder3c Cs|j}|j}|}|dd|d}|dks2t|dd}||dd|dd}tjjj|||f|d}tj | |dd||d |j ddd}tjjj |||j dd|} | |} | S)aCompute delta coefficients of a tensor, usually a spectrogram: .. devices:: CPU CUDA .. properties:: TorchScript .. math:: d_t = \frac{\sum_{n=1}^{\text{N}} n (c_{t+n} - c_{t-n})}{2 \sum_{n=1}^{\text{N}} n^2} where :math:`d_t` is the deltas at time :math:`t`, :math:`c_t` is the spectrogram coeffcients at time :math:`t`, :math:`N` is ``(win_length-1)//2``. Args: specgram (Tensor): Tensor of audio of dimension `(..., freq, time)` win_length (int, optional): The window length used for computing delta (Default: ``5``) mode (str, optional): Mode parameter passed to padding (Default: ``"replicate"``) Returns: Tensor: Tensor of deltas of dimension `(..., freq, time)` Example >>> specgram = torch.randn(1, 40, 1000) >>> delta = compute_deltas(specgram) >>> delta2 = compute_deltas(delta) rZr5rrL)rr)groups) r]r\r@rAr`r=r>r?r(rrepeatrCconv1d) rVr,rr]r\rCrZdenomkerneloutputrIrIrJrzs  * )r'r frame_timefreq_lowr3cCs>d}tt||}tt||}|d}tt||}||||} tjj|d| f}g} td|dD]} |dd| f d||dd|ddf} |d| df d||dd|ddf} | |  d|tj | ddd d|tj | ddd d}| |dq|t| d}|S) a Compute Normalized Cross-Correlation Function (NCCF). .. math:: \phi_i(m) = \frac{\sum_{n=b_i}^{b_i + N-1} w(n) w(m+n)}{\sqrt{E(b_i) E(m+b_i)}}, where :math:`\phi_i(m)` is the NCCF at frame :math:`i` with lag :math:`m`, :math:`w` is the waveform, :math:`N` is the length of a frame, :math:`b_i` is the beginning of frame :math:`i`, :math:`E(j)` is the energy :math:`\sum_{n=j}^{j+N-1} w^2(n)`. & .>r5rrZ.NrL)rrs)rrceilr@r=r>r?r(rfunfoldrErrDappendrr)r'rrrEPSILONZlags frame_sizeZwaveform_lengthZ num_of_framesrZ output_lagZlags1s2Z output_framesnccfrIrIrJ _compute_nccfs( .,  rGz?)abthreshr3cCsP|d||dk}||d||d}||d||d}||fS)zb Take value from first if bigger than a multiplicative factor of the second, elementwise. rrZrI)rrrrrindicesrIrIrJ _combine_maxsr)rr freq_highr3cCsvtt||}t|d|dfd}|jdd}t|d||fd}t||}|d}||7}|d7}|S)z For each frame, take the highest value of NCCF, apply centered median smoothing, and convert to frequency. Note: If the max among all the lags is very close to the first half of lags, then the latter is taken. .Nr5rLrZ)rrrr=rvrCr)rrrZlag_minbestZ half_sizerTrrIrIrJ_find_max_per_frames  r)rr,r3cCsv|dd}tjjj||dfddd}tj||d|fdgdd |dd |f<|d|d}t|d\}}|S) zH Apply median smoothing to the 1D tensor over the given window. rZrLrr4r^)rr.r5rrN)r=r>r?r(rrrZmedian)rr,Z pad_lengthZrollrrjrIrIrJ_median_smoothings  .r{Gz?UH )r'rrr,rrr3c Cst|}|dg|dd}t||||}t|||}t||}d} || |tj} | |ddt| j dd} | S)aDetect pitch frequency. .. devices:: CPU CUDA .. properties:: TorchScript It is implemented using normalized cross-correlation function and median smoothing. Args: waveform (Tensor): Tensor of audio of dimension `(..., freq, time)` sample_rate (int): The sample rate of the waveform (Hz) frame_time (float, optional): Duration of a frame (Default: ``10 ** (-2)``). win_length (int, optional): The window length for median smoothing (in number of frames) (Default: ``30``). freq_low (int, optional): Lowest frequency that can be detected (Hz) (Default: ``85``). freq_high (int, optional): Highest frequency that can be detected (Hz) (Default: ``3400``). Returns: Tensor: Tensor of freq of dimension `(..., frame)` r5Nr) rbr@rArrrrr=rSrC) r'rrr,rrrCrrrr|rIrIrJrs   $XdF)rV cmn_windowmin_cmn_windowr/ norm_varsr3cCs|j}|dd\}}|d||}|jd}|j} |j} d} } tj||| | d} tj||| | d}tj|||| | d}t|D]$}d}d}|r||d}||}n||}|d}|dkr||8}d}|s||krt|d|}||kr|||8}|}|dkrd}| dkrv|dd|||ddf}| t|d7} |r|t |dddddddf7}nt|| kr|dd| ddf}| |8} |r||d8}|| kr|dd| ddf}| |7} |r||d7}||}|} |} |dd|ddf| ||dd|ddf<|r|dkr^tj||| | d|dd|ddf<q|}||}|| d|d8}t |d}|dd|ddf|9<q||dd||f}t |dkr| d}|S) a Apply sliding-window cepstral mean (and optionally variance) normalization per utterance. .. devices:: CPU CUDA .. properties:: TorchScript Args: specgram (Tensor): Tensor of spectrogram of dimension `(..., time, freq)` cmn_window (int, optional): Window in frames for running average CMN computation (int, default = 600) min_cmn_window (int, optional): Minimum CMN window used at start of decoding (adds latency only at start). Only applicable if center == false, ignored if center==true (int, default = 100) center (bool, optional): If true, use a window centered on the current frame (to the extent possible, modulo end effects). If false, window is to the left. (bool, default = false) norm_vars (bool, optional): If true, normalize variance to one. (bool, default = false) Returns: Tensor: Tensor matching input shape `(..., freq, time)` r8Nr5rr[rLrZg) rCrwr\r]r=rrfrvrErrDlenr)rVrrr/rZ input_shapeZ num_framesZ num_featsZ num_channelsr\r]Zlast_window_startZlast_window_endZcur_sumZ cur_sumsqZ cmn_specgramrZ window_startZ window_endZ input_partZframe_to_removeZ frame_to_addZ window_framesZvariancerIrIrJrDsx       (    0 &   )r'rr(r)r*r+r,r3c Cs^t||||||ddd}tjd|dd|d|jdd}d } ||j| d |j| d S) aCompute the spectral centroid for each channel along the time axis. .. devices:: CPU CUDA .. properties:: Autograd TorchScript The spectral centroid is defined as the weighted average of the frequency values, weighted by their magnitude. Args: waveform (Tensor): Tensor of audio of dimension `(..., time)` sample_rate (int): Sample rate of the audio waveform pad (int): Two sided padding of signal window (Tensor): Window tensor that is applied/multiplied to each frame/window n_fft (int): Size of FFT hop_length (int): Length of hop between STFT windows win_length (int): Window size Returns: Tensor: Dimension `(..., time)` r:F)r(r)r*r+r,r-r.rrLrZ)Zstepsr])r5rZr8rr)r r=rr]rArE) r'rr(r)r*r+r,rVrZfreq_dimrIrIrJrs &)r'rrachannels_first compressionencodingbits_per_sampler3c Cs^t}tjj|||||||||dtjjj|||d\}} | |krZt|| |}|S)a^ Apply codecs as a form of augmentation. .. devices:: CPU Args: waveform (Tensor): Audio data. Must be 2 dimensional. See also ```channels_first```. sample_rate (int): Sample rate of the audio waveform. format (str): File format. channels_first (bool, optional): When True, both the input and output Tensor have dimension `(channel, time)`. Otherwise, they have dimension `(time, channel)`. compression (float or None, optional): Used for formats other than WAV. For more details see :py:func:`torchaudio.backend.sox_io_backend.save`. encoding (str or None, optional): Changes the encoding for the supported formats. For more details see :py:func:`torchaudio.backend.sox_io_backend.save`. bits_per_sample (int or None, optional): Changes the bit depth for the supported formats. For more details see :py:func:`torchaudio.backend.sox_io_backend.save`. Returns: Tensor: Resulting Tensor. If ``channels_first=True``, it has `(channel, time)` else `(time, channel)`. r)rra) ioBytesIO torchaudiobackendZsox_io_backendsaveseekloadr) r'rrarrrrbytesZ augmentedsrrIrIrJrs !  9@ry2rz{Gzt?XrZ)r'r frame_length frame_shiftmin_f0max_f0 soft_min_f0penalty_factorlowpass_cutoffresample_frequency delta_pitch nccf_ballastlowpass_filter_widthupsample_filter_widthmax_frames_latencyframes_per_chunksimulate_first_pass_onlinerecompute_frame snip_edgesr3cCsl|j}|d|d}tjj|||||||||| | | | | |||||}||dd|jdd}|S)ad Extract pitch based on method described in *A pitch extraction algorithm tuned for automatic speech recognition* [:footcite:`6854049`]. .. devices:: CPU .. properties:: TorchScript This function computes the equivalent of `compute-kaldi-pitch-feats` from Kaldi. Args: waveform (Tensor): The input waveform of shape `(..., time)`. sample_rate (float): Sample rate of `waveform`. frame_length (float, optional): Frame length in milliseconds. (default: 25.0) frame_shift (float, optional): Frame shift in milliseconds. (default: 10.0) min_f0 (float, optional): Minimum F0 to search for (Hz) (default: 50.0) max_f0 (float, optional): Maximum F0 to search for (Hz) (default: 400.0) soft_min_f0 (float, optional): Minimum f0, applied in soft way, must not exceed min-f0 (default: 10.0) penalty_factor (float, optional): Cost factor for FO change. (default: 0.1) lowpass_cutoff (float, optional): Cutoff frequency for LowPass filter (Hz) (default: 1000) resample_frequency (float, optional): Frequency that we down-sample the signal to. Must be more than twice lowpass-cutoff. (default: 4000) delta_pitch( float, optional): Smallest relative change in pitch that our algorithm measures. (default: 0.005) nccf_ballast (float, optional): Increasing this factor reduces NCCF for quiet frames (default: 7000) lowpass_filter_width (int, optional): Integer that determines filter width of lowpass filter, more gives sharper filter. (default: 1) upsample_filter_width (int, optional): Integer that determines filter width when upsampling NCCF. (default: 5) max_frames_latency (int, optional): Maximum number of frames of latency that we allow pitch tracking to introduce into the feature processing (affects output only if ``frames_per_chunk > 0`` and ``simulate_first_pass_online=True``) (default: 0) frames_per_chunk (int, optional): The number of frames used for energy normalization. (default: 0) simulate_first_pass_online (bool, optional): If true, the function will output features that correspond to what an online decoder would see in the first pass of decoding -- not the final version of the features, which is the default. (default: False) Relevant if ``frames_per_chunk > 0``. recompute_frame (int, optional): Only relevant for compatibility with online pitch extraction. A non-critical parameter; the frame at which we recompute some of the forward pointers, after revising our estimate of the signal energy. Relevant if ``frames_per_chunk > 0``. (default: 500) snip_edges (bool, optional): If this is set to false, the incomplete frames near the ending edge won't be snipped, so that the number of frames is the file size divided by the frame-shift. This makes different types of features give the same number of frames. (default: True) Returns: Tensor: Pitch feature. Shape: `(batch, frames 2)` where the last dimension corresponds to pitch and NCCF. r5Nr8)rCrAr=opsrZkaldi_ComputeKaldiPitch)r'rrrrrrrrrrrrrrrrrr rCresultrIrIrJrs2W sinc_interpolationcpu) orig_freqnew_freqgcdrrolloffresampling_methodbetar]r\c  Cst||krt||ks td|dkr6td|t||}t||}|dksZtg} t||} | |9} t||| } |dk r|ntj } tj | | ||| d} t |D]}| || || }| | |}|dkrt |tj|dd}nF|dkrd}tt|}t|td ||dt|}|tj9}t|dktd |t||}||| |q| |}t| |d d |} |dkr| jtjd } | | fS) NaFrequencies must be of integer type to ensure quality resampling computation. To work around this, manually convert both frequencies to integer values that maintain their resampling rate ratio before passing them into the function. Example: To downsample a 44100 hz waveform by a factor of 8, use `orig_freq=8` and `new_freq=1` instead of `orig_freq=44100` and `new_freq=5512.5`. For more information, please refer to https://github.com/pytorch/audio/issues/1487.)r$Z kaiser_windowzInvalid resampling method: {}rrr$rLgQaTi-@rZr:r5r)r ExceptionrNrar`rprrr=Zfloat64rrfZclamp_rrrerSZi0rFwherersinrgrstackrwZfloat32)r&r'r(rr)r*r+r]r\ZkernelsZ base_freqwidthZ idx_dtypeidxirr)Z beta_tensorrscalerIrIrJ_get_sinc_resample_kernellsB         * (   r4)r'r&r'r(rr0c Cs|std|jdt||}t||}|}|d|d}|j\}}tjj ||||f}tjj j |dddf||d} | dd |d} tt|||} | dd| f} | |dd| jdd} | S)Nz?Expected floating point type for waveform tensor, but received .r5)ZstriderZrL.)r TypeErrorr\rr@rwrCr=r>r?r(rrrArr) r'r&r'r(rr0rCZnum_wavsrK resampledZ target_lengthrIrIrJ_apply_sinc_resample_kernels     r8)r'r&r'rr)r*r+r3c Csj|dkr|dkst||kr |Stt|t|}t||||||||j|j \}} t|||||| } | S)aResamples the waveform at the new frequency using bandlimited interpolation. [:footcite:`RESAMPLE`]. .. devices:: CPU CUDA .. properties:: Autograd TorchScript Note: ``transforms.Resample`` precomputes and reuses the resampling kernel, so using it will result in more efficient computation if resampling multiple waveforms with the same resampling parameters. Args: waveform (Tensor): The input signal of dimension `(..., time)` orig_freq (int): The original frequency of the signal new_freq (int): The desired frequency lowpass_filter_width (int, optional): Controls the sharpness of the filter, more == sharper but less efficient. (Default: ``6``) rolloff (float, optional): The roll-off frequency of the filter, as a fraction of the Nyquist. Lower values reduce anti-aliasing, but also reduce some of the highest frequencies. (Default: ``0.99``) resampling_method (str, optional): The resampling method to use. Options: [``sinc_interpolation``, ``kaiser_window``] (Default: ``'sinc_interpolation'``) beta (float or None, optional): The shape parameter used for kaiser window. Returns: Tensor: The waveform at the new frequency of dimension `(..., time).` r^)r`rr(rr4r]r\r8) r'r&r'rr)r*r+r(rr0r7rIrIrJrs"# )seq1seq2r3c Cst|}tt|d}ddt|dD}tdt|dD]}||d<td|dD]j}||d||dkr||d||<qZ||dd}||dd}||d} t||| ||<qZ||}}q@t|dS)a4 Calculate the word level edit (Levenshtein) distance between two sequences. .. devices:: CPU The function computes an edit distance allowing deletion, insertion and substitution. The result is an integer. For most applications, the two input sequences should be the same type. If two strings are given, the output is the edit distance between the two strings (character edit distance). If two lists of strings are given, the output is the edit distance between sentences (word edit distance). Users may want to normalize the output by the length of the reference sequence. Args: seq1 (Sequence): the first sequence to compare. seq2 (Sequence): the second sequence to compare. Returns: int: The distance between the first and second sequences. rZcSsg|]}dqS)rrI).0rjrIrIrJ 3sz!edit_distance..rr5)rrbrfrpr) r9r:Z len_sent2ZdoldZdnewr2jZ substitutionZ insertionZdeletionrIrIrJrs   ) r'rn_stepsbins_per_octaver*r,r+r)r3c CsHt|||||||}dt| |} t|t|| |} t| |S)a# Shift the pitch of a waveform by ``n_steps`` steps. .. devices:: CPU CUDA .. properties:: TorchScript Args: waveform (Tensor): The input waveform of shape `(..., time)`. sample_rate (int): Sample rate of `waveform`. n_steps (int): The (fractional) steps to shift `waveform`. bins_per_octave (int, optional): The number of steps per octave (Default: ``12``). n_fft (int, optional): Size of FFT, creates ``n_fft // 2 + 1`` bins (Default: ``512``). win_length (int or None, optional): Window size. If None, then ``n_fft`` is used. (Default: ``None``). hop_length (int or None, optional): Length of hop between STFT windows. If None, then ``win_length // 4`` is used (Default: ``None``). window (Tensor or None, optional): Window tensor that is applied/multiplied to each frame/window. If None, then ``torch.hann_window(win_length)`` is used (Default: ``None``). Returns: Tensor: The pitch-shifted audio waveform of shape `(..., time)`. r9)_stretch_waveformrSrr_fix_waveform_shaper@) r'rr@rAr*r,r+r)waveform_stretchrwaveform_shiftrIrIrJrEs! )r'r@rAr*r,r+r)r3c Cs|dkr|d}|dkr|}|dkr4tj||jd}|}|d|d}|d}dt| |} tj|||||dddddd } tjd tj || j d | jd d } t | | | } t t || } tj| ||||| d}|S)z Pitch shift helper function to preprocess and stretch waveform before resampling step. Args: See pitch_shift arg descriptions. Returns: Tensor: The preprocessed waveform stretched prior to resampling. N)Z window_lengthr]r5r9Tr&Fr6rr8)r].Nr_)r=Z hann_windowr]r@rArSrBrrrrCrrrrO)r'r@rAr*r,r+r)rCori_lenrrHrZ spec_stretchZ len_stretchrDrIrIrJrBusD $ rB)rErCr3cCsj|d}|d}||kr.|dd|f}ntjj|d||g}||dd|jdd}|S)a? PitchShift helper function to process after resampling step to fix the shape back. Args: waveform_shift(Tensor): The waveform after stretch and resample shape (List[int]): The shape of initial waveform Returns: Tensor: The pitch-shifted audio waveform of shape `(..., time)`. r5.Nr)r@r=r>r?r(rwrC)rErCrHZ shift_lenrIrIrJrCs  rCr5mean)logitstargets logit_lengthstarget_lengthsblankru reductionc Csh|dkrtd|dkr&|jd|}tjjj||||||d\}}|dkrT|S|dkrd|S|S)a.Compute the RNN Transducer loss from *Sequence Transduction with Recurrent Neural Networks* [:footcite:`graves2012sequence`]. .. devices:: CPU CUDA .. properties:: Autograd TorchScript The RNN Transducer loss extends the CTC loss by defining a distribution over output sequences of all lengths, and by jointly modelling both input-output and output-output dependencies. Args: logits (Tensor): Tensor of dimension `(batch, max seq length, max target length + 1, class)` containing output from joiner targets (Tensor): Tensor of dimension `(batch, max target length)` containing targets with zero padded logit_lengths (Tensor): Tensor of dimension `(batch)` containing lengths of each sequence from encoder target_lengths (Tensor): Tensor of dimension `(batch)` containing lengths of targets for each sequence blank (int, optional): blank label (Default: ``-1``) clamp (float, optional): clamp for gradients (Default: ``-1``) reduction (string, optional): Specifies the reduction to apply to the output: ``'none'`` | ``'mean'`` | ``'sum'``. (Default: ``'mean'``) Returns: Tensor: Loss with the reduction option applied. If ``reduction`` is ``'none'``, then size `(batch)`, otherwise scalar. )nonerIrEz3reduction should be one of 'none', 'mean', or 'sum'rr5)rJrKrLrMrNrurIrE)rNrCr=r!rrrIrE) rJrKrLrMrNrurOZcostsrjrIrIrJrs"" 绽|=)rVr normalizeepsr3cCs|dd}td||g}|dk r|jdd|jddkrX|jd|jdks`tdd|jd|jd |r||jdd d |}||d }|jdd }|S)a/Compute cross-channel power spectral density (PSD) matrix. .. devices:: CPU CUDA .. properties:: Autograd TorchScript Args: specgram (torch.Tensor): Multi-channel complex-valued spectrum. Tensor with dimensions `(..., channel, freq, time)`. mask (torch.Tensor or None, optional): Time-Frequency mask for normalization. Tensor with dimensions `(..., freq, time)`. (Default: ``None``) normalize (bool, optional): If ``True``, normalize the mask along the time dimension. (Default: ``True``) eps (float, optional): Value to add to the denominator in mask normalization. (Default: ``1e-15``) Returns: torch.Tensor: The complex-valued PSD matrix of the input spectrum. Tensor with dimensions `(..., freq, channel, channel)` rqr8z...ct,...et->...tceNr5zSThe dimensions of mask except the channel dimension should be the same as specgram.Found z for mask and  for specgram.T)rsZkeepdimrrr)rr=einsumconjrCr`rE)rVrrRrSr rIrIrJr s   r8)r7dim1dim2r3cCsL|jdkstd|j||j|ks.tdtj|d||d}|jddS)aCompute the trace of a Tensor along ``dim1`` and ``dim2`` dimensions. Args: input (torch.Tensor): Tensor with dimensions `(..., channel, channel)`. dim1 (int, optional): The first dimension of the diagonal matrix. (Default: ``-1``) dim2 (int, optional): The second dimension of the diagonal matrix. (Default: ``-2``) Returns: Tensor: The trace of the input Tensor. rLz/The dimension of the tensor must be at least 2.z3The size of ``dim1`` and ``dim2`` must be the same.r)rXrYr5rr)ndimr`rCr=ZdiagonalrE)r7rXrYrIrIrJ_compute_mat_trace(s r[Hz>:0yE>)matregrSr3cCsZ|d}tj||j|jd}t|jd|}||}|||dddddf}|S)aPerform Tikhonov regularization (only modifying real part). Args: mat (torch.Tensor): Input matrix with dimensions `(..., channel, channel)`. reg (float, optional): Regularization factor. (Default: 1e-8) eps (float, optional): Value to avoid the correlation matrix is all-zero. (Default: ``1e-8``) Returns: Tensor: Regularized matrix with dimensions `(..., channel, channel)`. r5r[r.N)r@r=eyer\r]r[r)r^r_rSCr`epsilonrIrIrJ_tik_reg;s rc)psd_spsd_nr3cCs|jdkr|jdkstd|r,|s4tdd|jd|jd|j|jksptd|jd|jd |jd |jd kstd |jd d S)aAssertion checks of the PSD matrices of target speech and noise. Args: psd_s (torch.Tensor): The complex-valued power spectral density (PSD) matrix of target speech. Tensor with dimensions `(..., freq, channel, channel)`. psd_n (torch.Tensor): The complex-valued power spectral density (PSD) matrix of noise. Tensor with dimensions `(..., freq, channel, channel)`. rzNExpected at least 3D Tensor (..., freq, channel, channel) for psd_s and psd_n.zJThe type of psd_s and psd_n must be ``torch.cfloat`` or ``torch.cdouble``.rTz for psd_s and for psd_n.z...9Expected 'int' or 'Tensor' for reference_channel. Found: r5)rhrcr=linalgsolver[jit isinstancerrrr\rVr6type) rdrerirjrkrS numeratorwsbeamform_weightsrIrIrJr!is,   ")rtfrerirjrkrSr3c Cs|jdkstd|jd|jdks8td|jd|rH|sPtdd|jd|jd |j|jd d kstd d|jd|jd |jd |jd kstd|jd|rt||d}tj|| d  d }t d| |g}||j d |}|d k rtj|tr<|d|f } nTtj|tr|||j}t d| |dd d d fg} ntdt|d|| d}|S)aSCompute the Minimum Variance Distortionless Response (*MVDR* [:footcite:`capon1969high`]) beamforming weights based on the relative transfer function (RTF) and power spectral density (PSD) matrix of noise. .. devices:: CPU CUDA .. properties:: Autograd TorchScript Given the relative transfer function (RTF) matrix or the steering vector of target speech :math:`\bm{v}`, the PSD matrix of noise :math:`\bf{\Phi}_{\textbf{NN}}`, and a one-hot vector that represents the reference channel :math:`\bf{u}`, the method computes the MVDR beamforming weight martrix :math:`\textbf{w}_{\text{MVDR}}`. The formula is defined as: .. math:: \textbf{w}_{\text{MVDR}}(f) = \frac{{{\bf{\Phi}_{\textbf{NN}}^{-1}}(f){\bm{v}}(f)}} {{\bm{v}^{\mathsf{H}}}(f){\bf{\Phi}_{\textbf{NN}}^{-1}}(f){\bm{v}}(f)} where :math:`(.)^{\mathsf{H}}` denotes the Hermitian Conjugate operation. Args: rtf (torch.Tensor): The complex-valued RTF vector of target speech. Tensor with dimensions `(..., freq, channel)`. psd_n (torch.Tensor): The complex-valued power spectral density (PSD) matrix of noise. Tensor with dimensions `(..., freq, channel, channel)`. reference_channel (int or torch.Tensor): Specifies the reference channel. If the dtype is ``int``, it represents the reference channel index. If the dtype is ``torch.Tensor``, its shape is `(..., channel)`, where the ``channel`` dimension is one-hot. diagonal_loading (bool, optional): If ``True``, enables applying diagonal loading to ``psd_n``. (Default: ``True``) diag_eps (float, optional): The coefficient multiplied to the identity matrix for diagonal loading. It is only effective when ``diagonal_loading`` is set to ``True``. (Default: ``1e-7``) eps (float, optional): Value to add to the denominator in the beamforming weight formula. (Default: ``1e-8``) Returns: torch.Tensor: The complex-valued MVDR beamforming weight matrix with dimensions `(..., freq, channel)`. rLz@Expected at least 2D Tensor (..., freq, channel) for rtf. Found r5rzKExpected at least 3D Tensor (..., freq, channel, channel) for psd_n. Found zHThe type of rtf and psd_n must be ``torch.cfloat`` or ``torch.cdouble``.rTz for rtf and rfNr5z_The dimensions of rtf and the dimensions withou the last dimension of psd_n should be the same.r8z;The last two dimensions of psd_n should be the same. Found rlz...d,...d->....rmrnrG)rZr`rCrMr\rcr=rorprrrVrWrrqrrrrrr6rs) rwrerirjrkrSrt denominatorrvr3rIrIrJr"s@.    $ )rdr3cCs\|std|jd|jd|jdks@td|jdtj|\}}|d}|S)aEstimate the relative transfer function (RTF) or the steering vector by eigenvalue decomposition. .. devices:: CPU CUDA .. properties:: TorchScript Args: psd_s (Tensor): The complex-valued power spectral density (PSD) matrix of target speech. Tensor of dimension `(..., freq, channel, channel)` Returns: Tensor: The estimated complex-valued RTF of target speech. Tensor of dimension `(..., freq, channel)` zGThe type of psd_s must be ``torch.cfloat`` or ``torch.cdouble``. Found r5r5r8rg).r5)rMr`r\rCr=roZeigh)rdrjvrwrIrIrJr#s r)rdrerirWrjrkr3c Cst|||dkstd|r*t||d}tj||}tj|trT|d|f}nPtj|t r| |j }t d||dddddfg}nt dt|d|d }|d krt|d D]}t||}qt||}n t||}|d S) aEstimate the relative transfer function (RTF) or the steering vector by the power method. .. devices:: CPU CUDA .. properties:: Autograd TorchScript Args: psd_s (torch.Tensor): The complex-valued power spectral density (PSD) matrix of target speech. Tensor with dimensions `(..., freq, channel, channel)`. psd_n (torch.Tensor): The complex-valued power spectral density (PSD) matrix of noise. Tensor with dimensions `(..., freq, channel, channel)`. reference_channel (int or torch.Tensor): Specifies the reference channel. If the dtype is ``int``, it represents the reference channel index. If the dtype is ``torch.Tensor``, its shape is `(..., channel)`, where the ``channel`` dimension is one-hot. diagonal_loading (bool, optional): If ``True``, enables applying diagonal loading to ``psd_n``. (Default: ``True``) diag_eps (float, optional): The coefficient multiplied to the identity matrix for diagonal loading. It is only effective when ``diagonal_loading`` is set to ``True``. (Default: ``1e-7``) Returns: torch.Tensor: The estimated complex-valued RTF of target speech. Tensor of dimension `(..., freq, channel)`. rz/The number of iteration must be greater than 0.rl.rmNrnr5r5rL)rhr`rcr=rorprqrrrrrr\rVr6rsrrfmatmulr) rdrerirWrjrkphirwrjrIrIrJr$s$   "  )rvrVr3cCs|jdd|jddks$tdd|jd|jd|rJ|sRtdd|jd|jdtd ||g}|S) aApply the beamforming weight to the multi-channel noisy spectrum to obtain the single-channel enhanced spectrum. .. devices:: CPU CUDA .. properties:: Autograd TorchScript .. math:: \hat{\textbf{S}}(f) = \textbf{w}_{\text{bf}}(f)^{\mathsf{H}} \textbf{Y}(f) where :math:`\textbf{w}_{\text{bf}}(f)` is the beamforming weight for the :math:`f`-th frequency bin, :math:`\textbf{Y}` is the multi-channel spectrum for the :math:`f`-th frequency bin. Args: beamform_weights (Tensor): The complex-valued beamforming weight matrix. Tensor of dimension `(..., freq, channel)` specgram (Tensor): The multi-channel complex-valued noisy spectrum. Tensor of dimension `(..., channel, freq, time)` Returns: Tensor: The single-channel complex-valued enhanced spectrum. Tensor of dimension `(..., freq, time)` Nr8rqzUThe dimensions except the last two dimensions of beamform_weights should be the same rTz for beamform_weights and rUzXThe type of beamform_weights and specgram must be ``torch.cfloat`` or ``torch.cdouble``.z...fc,...cft->...ft)rCr`rMr\r=rVrW)rvrVZspecgram_enhancedrIrIrJr%Ps)Tr&TN)Tr&T)N)r{)r{)Nr{)r:)r:)rr)r)rrrr)rrFF)TNNN)rryr r ryrzr r r rrZrrrFrT)r#rr$N)r>r?NNN)r>r?NNN)r5r5rI)NTrQ)r5r8)r\r])Tr\r])NTr\r])rTr\)Hrrr;collections.abcrtypingrrrrr=rrZtorchaudio._internalrZ _mod_utils__all__rrSboolstrr r r\rUr r r rrrrrrrrrrrrrrrrrrrrZ requires_soxrZrequires_kaldirr]r4r8rrqZunusedrrrBrCrr r[rcrhr!r"r#r$r%rIrIrIrJs   , a L  d +""  I .!J  @ @5/.  / ^ -+u Y " 9- 3 5  ; ,  B R  >