espnet2.utils.griffin_lim.griffin_lim

source

espnet2.utils.griffin_lim.griffin_lim(spc: ndarray, n_fft: int, n_shift: int, win_length: int | None = None, window: str | None = 'hann', n_iter: int | None = 32) → ndarray

Convert linear spectrogram into waveform using Griffin-Lim.

This function reconstructs a waveform from a linear spectrogram using the Griffin-Lim algorithm, which iteratively refines the phase information to recover the audio signal.

• Parameters:
  • spc – Linear spectrogram (T, n_fft // 2 + 1).
  • n_fft – The number of FFT points.
  • n_shift – Shift size in points.
  • win_length – Window length in points (optional).
  • window – Window function type (optional, default is “hann”).
  • n_iter – The number of iterations (optional, default is 32).
• Returns: Reconstructed waveform (N,).

Examples

Example of using griffin_lim to convert a linear spectrogram back to audio

import numpy as np spc = np.random.rand(100, 1025) # Example linear spectrogram waveform = griffin_lim(spc, n_fft=2048, n_shift=512)

NOTE

This function requires librosa library version 0.7.0 or higher for the optimized Griffin-Lim algorithm. If an older version is used, a warning will be logged, and a slower implementation will be used.

• Raises:
  • ValueError – If the input spectrogram shape does not match the expected
  • dimensions. –