icqt#

diffsptk.ICQT#

alias of InverseConstantQTransform

class diffsptk.InverseConstantQTransform(frame_period, sample_rate, *, f_min=32.7, n_bin=84, n_bin_per_octave=12, tuning=0, filter_scale=1, norm=1, sparsity=0.01, window='hann', scale=True, res_type='kaiser_best', **kwargs)[source]#

Perform inverse constant-Q transform based on the librosa implementation.

Parameters:

frame_periodint >= 1

Frame period in samples, $P$.

sample_rateint >= 1

Sample rate in Hz.

f_minfloat > 0

Minimum center frequency in Hz.

n_binint >= 1

Number of CQ-bins, $K$.

n_bin_per_octaveint >= 1

number of bins per octave, $B$.

tuningfloat

Tuning offset in fractions of a bin.

filter_scalefloat > 0

Filter scale factor.

normfloat

Type of norm used in basis function normalization.

sparsityfloat in [0, 1)

Sparsification factor.

windowstr

Window function for the basis.

scalebool

If True, scale the CQT response by the length of filter.

res_type[‘kaiser_best’, ‘kaiser_fast’] or None

Resampling type.

**kwargsadditional keyword arguments

See torchaudio.transforms.Resample.

forward(c, out_length=None)[source]#

Compute inverse constant-Q transform.

Parameters:

cTensor [shape=(…, T/P, K)]

CQT complex output.

out_lengthint or None

Length of output waveform.

Returns:

outTensor [shape=(…, T)]

Reconstructed waveform.

Examples

>>> x = diffsptk.sin(99)
>>> cqt = diffsptk.CQT(100, 8000, n_bin=4)
>>> icqt = diffsptk.ICQT(100, 8000, n_bin=4)
>>> y = icqt(cqt(x), out_length=x.size(0))
>>> y.shape
torch.Size([100])