The Difference librosa.filters.mel() and librosa.feature.melspectrogram() – Librosa Tutorial

To compute mel in librosa, we can use librosa.filters.mel() and librosa.feature.melspectrogram(). In this tutorial, we will introduce the difference between them.

librosa.filters.mel()

It is defined as:

librosa.filters.mel(*, sr, n_fft, n_mels=128, fmin=0.0, fmax=None, htk=False, norm='slaney', dtype=<class 'numpy.float32'>)

It will create a Mel filter-bank and produce a linear transformation matrix to project FFT bins onto Mel-frequency bins.

Notice: It creates a Mel filter-bank does not FBank, you can not use it as audio feature.

For example:

import librosa
import numpy as np
import matplotlib.pyplot as plt

def plot_mel_fbank(fbank, title=None):
    fig, axs = plt.subplots(1, 1)
    axs.set_title(title or "Filter bank")
    axs.imshow(fbank, aspect="auto")
    axs.set_ylabel("frequency bin")
    axs.set_xlabel("mel bin")
    plt.show(block=False)
sr = 8000

mels = librosa.filters.mel(sr=sr, n_fft = 512, fmin=0.0, fmax=sr / 2.0,n_mels=80)
print(mels.shape)
plot_mel_fbank(mels)

Run this code, you will see:

(80, 257)

We can find this function only returns a weight, it can not process any audio data.

librosa.feature.melspectrogram()

It is defined as:

librosa.feature.melspectrogram(*, y=None, sr=22050, S=None, n_fft=2048, hop_length=512, win_length=None, window='hann', center=True, pad_mode='constant', power=2.0, **kwargs)

It can compute a mel-scaled spectrogram.

Notice: The result of this function can be used as the feature of an audio.

In order to understand how to use this function, you can read:

Compute Audio Log Mel Spectrogram Feature: A Step Guide – Python Audio Processing

We will use an example to show the effect of mel spectrogram.

import librosa
import numpy as np
import matplotlib.pyplot as plt

def plot_spectrogram(spec, title=None, ylabel="freq_bin", aspect="auto", xmax=None):
    fig, axs = plt.subplots(1, 1)
    axs.set_title(title or "Spectrogram (db)")
    axs.set_ylabel(ylabel)
    axs.set_xlabel("frame")
    im = axs.imshow(librosa.power_to_db(spec), origin="lower", aspect=aspect)
    if xmax:
        axs.set_xlim((0, xmax))
    fig.colorbar(im, ax=axs)
    plt.show(block=False)

audio_file = 'speech-01-002.flac'
sr = 8000
audio_data, sr = librosa.load(audio_file, sr= sr, mono=True)
print(audio_data.shape)
win_length = int(0.025 * sr)
hop_length = int(0.01 * sr)
melspectrum = librosa.feature.melspectrogram(y=audio_data, sr=sr, hop_length= hop_length, win_length = win_length, window='hann', n_fft = 512, n_mels=80)
print(melspectrum.shape)
plot_spectrogram(melspectrum)

Run this code, you will see:

(389724,)
(80, 4872)

For the source code of librosa.feature.melspectrogram(), we can find:

    # Build a Mel filter
    mel_basis = filters.mel(sr=sr, n_fft=n_fft, **kwargs)

    return np.einsum("...ft,mf->...mt", S, mel_basis, optimize=True)

librosa.filters.mel() is used in librosa.feature.melspectrogram().