espnet2.utils.griffin_lim.logmel2linear

source

espnet2.utils.griffin_lim.logmel2linear(lmspc: ndarray, fs: int, n_fft: int, n_mels: int, fmin: int | None = None, fmax: int | None = None) → ndarray

Convert log Mel filterbank to linear spectrogram.

This function transforms a log Mel spectrogram into a linear spectrogram using the inverse of the Mel filterbank. It computes the inverse Mel basis and applies it to the log Mel spectrogram.

• Parameters:
  • lmspc (np.ndarray) – Log Mel filterbank with shape (T, n_mels).
  • fs (int) – Sampling frequency.
  • n_fft (int) – The number of FFT points.
  • n_mels (int) – The number of Mel basis.
  • fmin (Optional *[*int ]) – Minimum frequency to analyze. If None, defaults to 0.
  • fmax (Optional *[*int ]) – Maximum frequency to analyze. If None, defaults to fs / 2.
• Returns: Linear spectrogram with shape (T, n_fft // 2 + 1).
• Return type: np.ndarray

Examples

>>> import numpy as np
>>> lmspc = np.random.rand(100, 40)  # Example log Mel spectrogram
>>> fs = 16000
>>> n_fft = 2048
>>> n_mels = 40
>>> linear_spc = logmel2linear(lmspc, fs, n_fft, n_mels)
>>> print(linear_spc.shape)  # Should be (100, 1025) for 2048 FFT points

NOTE

The function uses the librosa library to create the Mel filterbank and compute its pseudoinverse.