Compute and Display Audio Mel-spectrogram in Python – Python Tutorial

Audio mel-spectrogram is a classic feature for deep learning. In this tutorial, we will introduce how to get and display it using python.

librosa.feature.melspectrogram()

This function can compute a mel-scaled spectrogram.

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)

Here are some important parameters:

y: the audio data, it may (,n) shape.

sr: the audio sample rate.

hop_length: number of samples between successive frames. It will affect the result.

win_length: Each frame of audio is windowed by window()

From the source code, we can find the relation between hop_length and win_length is:

    # By default, use the entire frame
    if win_length is None:
        win_length = n_fft

    # Set the default hop, if it's not already specified
    if hop_length is None:
        hop_length = int(win_length // 4)

    fft_window = get_window(window, win_length, fftbins=True)

We will use an example to explain this function.

Read a wav file

import librosa
import numpy as np

audio_file =r'D:\1481637021654134785_sep.wav'
audio_data, sr = librosa.load(audio_file, sr= 8000, mono=True)
print(audio_data.shape)

In this example code, we use librosa.load() to read audio data. Here is the detail.

Understand librosa.load() is Between -1.0 and 1.0 – Librosa Tutorial

Run this code, we will get:

(182015,)

It means the sample poit is 182015 in this file.

Compute Mel-spectrogram

We will use librosa.feature.melspectrogram() to compute mel-spectrogram. Here is an example:

melspectrum = librosa.feature.melspectrogram(y=audio_data, sr=sr, hop_length= 512, window='hann', n_mels=256)
print(melspectrum.shape)

Run this code, we will get:

(256, 356)

If we change parameters hop_length and n_mels, how about the result?

melspectrum = librosa.feature.melspectrogram(y=audio_data, sr=sr, hop_length= 200, window='hann', n_mels=128)
print(melspectrum.shape)  #(128, 911)

The result will be 128*911.

From above we can find: the mel-spectrogram is a matrix. It is:

[n_mels, len(audio_data)//hop_length +1]

For example, if n_mels = 128, hop_length = 200,

len(audio_data)//hop_length +1 = 182015//200 + 1 = 911.

Display Mel-spectrogram

When we have computed Mel-spectrogram, we can display it. Here is an example:

import matplotlib.pyplot as plt
import librosa.display

fig, ax = plt.subplots()

S_dB = librosa.power_to_db(melspectrum, ref=np.max)

img = librosa.display.specshow(S_dB, x_axis='time',
                         y_axis='mel', sr=sr,
                         ax=ax)

fig.colorbar(img, ax=ax, format='%+2.0f dB')

ax.set(title='Mel-frequency spectrogram')
plt.show()

As to function: librosa.display.specshow() shoud be same to librosa.feature.melspectrogram().

So we should set hop_length = 512, then run this code, we will get an image as follows: