griffin#

class diffsptk.GriffinLim(frame_length: int, frame_period: int, fft_length: int, *, center: bool = True, mode: str = 'constant', window: str = 'blackman', norm: str = 'power', symmetric: bool = True, n_iter: int = 100, alpha: float = 0.99, beta: float = 0.99, gamma: float = 1.1, init_phase: str = 'random', verbose: bool = False)[source]#

Griffin-Lim phase reconstruction module.

Parameters:
frame_lengthint >= 1

The frame length in samples, \(L\).

frame_periodint >= 1

The frame period in samples, \(P\).

fft_lengthint >= L

The number of FFT bins, \(N\).

centerbool

If True, pad the input on both sides so that the frame is centered.

window[‘blackman’, ‘hamming’, ‘hanning’, ‘bartlett’, ‘trapezoidal’, ‘rectangular’, ‘nuttall’]

The window type.

norm[‘none’, ‘power’, ‘magnitude’]

The normalization type of the window.

symmetricbool

If True, the window is symmetric, otherwise periodic.

n_iterint >= 1

The number of iterations for phase reconstruction.

alphafloat >= 0

The momentum factor, \(\alpha\).

betafloat >= 0

The momentum factor, \(\beta\).

gammafloat >= 0

The smoothing factor, \(\gamma\).

init_phase[‘zeros’, ‘random’]

The initial phase for the reconstruction.

verbosebool

If True, print the SNR at each iteration.

References

[1]

R. Nenov et al., “Faster than fast: Accelerating the Griffin-Lim algorithm,” Proceedings of ICASSP, 2023.

forward(y: Tensor, out_length: int | None = None) Tensor[source]#

Reconstruct a waveform from the spectrum using the Griffin-Lim algorithm.

Parameters:
yTensor [shape=(…, T/P, N/2+1)]

The power spectrum.

out_lengthint > 0 or None

The length of the output waveform.

Returns:
outTensor [shape=(…, T)]

The reconstructed waveform.

Examples

>>> x = diffsptk.ramp(1, 3)
>>> x
tensor([1., 2., 3.])
>>> stft_params = {"frame_length": 3, "frame_period": 1, "fft_length": 8}
>>> stft = diffsptk.STFT(**stft_params, out_format="power")
>>> griffin = diffsptk.GriffinLim(**stft_params, n_iter=10, init_phase="zeros")
>>> y = griffin(stft(x), out_length=3)
>>> y
tensor([ 1.0000,  2.0000, -3.0000])
diffsptk.functional.griffin(y: Tensor, *, out_length: int | None = None, frame_length: int = 400, frame_period: int = 80, fft_length: int = 512, center: bool = True, mode: str = 'constant', window: str = 'blackman', norm: str = 'power', symmetric: bool = True, n_iter: int = 100, alpha: float = 0.99, beta: float = 0.99, gamma: float = 1.1, init_phase: str = 'random', verbose: bool = False) Tensor[source]#

Reconstruct a waveform from the spectrum using the Griffin-Lim algorithm.

Parameters:
yTensor [shape=(…, T/P, N/2+1)]

The power spectrum.

out_lengthint > 0 or None

The length of the output waveform.

frame_lengthint >= 1

The frame length in samples, \(L\).

frame_periodint >= 1

The frame period in samples, \(P\).

fft_lengthint >= L

The number of FFT bins, \(N\).

centerbool

If True, pad the input on both sides so that the frame is centered.

window[‘blackman’, ‘hamming’, ‘hanning’, ‘bartlett’, ‘trapezoidal’, ‘rectangular’, ‘nuttall’]

The window type.

norm[‘none’, ‘power’, ‘magnitude’]

The normalization type of the window.

symmetricbool

If True, the window is symmetric, otherwise periodic.

n_iterint >= 1

The number of iterations for phase reconstruction.

alphafloat >= 0

The momentum factor, \(\alpha\).

betafloat >= 0

The momentum factor, \(\beta\).

gammafloat >= 0

The smoothing factor, \(\gamma\).

init_phase[‘zeros’, ‘random’]

The initial phase for the reconstruction.

verbosebool

If True, print the SNR at each iteration.

Returns:
outTensor [shape=(…, T)]

The reconstructed waveform.

See also

stft istft