[AI from Scratch] Episode 303: Introduction to LibROSA — Basics of the Audio Processing Library

Recap and Today’s Theme

Hello! In the previous episode, we covered the basics of audio data, discussing key concepts such as sampling rate and bit depth. Understanding these fundamentals allows you to handle audio data properly and manage its quality.

Today, we’ll introduce LibROSA, a Python library for audio processing. LibROSA is a powerful library that offers various functions for loading, analyzing, transforming, and extracting features from audio data. In this episode, we will walk through how to use LibROSA to manipulate audio data.

What is LibROSA?

LibROSA is a Python library designed for audio and music analysis. It simplifies tasks such as loading and analyzing audio, extracting features like MFCC, visualizing waveforms, and performing time stretching or pitch shifting. It is widely used in fields like music information retrieval, speech recognition, and audio signal processing.

Main Features of LibROSA

• Loading and playing audio files
• Extracting basic audio features (e.g., MFCC, spectrograms)
• Analyzing pitch and tempo
• Time stretching and pitch shifting audio data
• Visualizing waveforms

LibROSA is used by both beginners and experts for various applications, ranging from research to product development in audio processing.

Installing LibROSA

First, let’s install LibROSA. You can use the following command to install LibROSA and its dependencies:

pip install librosa

Once installed, you are ready to load and process audio files with LibROSA.

Loading Audio Files

To load an audio file using LibROSA, the librosa.load() function is used. This function reads the audio data and returns both the waveform and the sampling rate. Here is a basic example of loading an audio file:

import librosa

# Specify the path to the audio file
audio_path = 'example.wav'

# Load the audio file
y, sr = librosa.load(audio_path, sr=None)

# Display the results
print(f'Sampling Rate: {sr}')
print(f'Waveform Length: {len(y)}')

• librosa.load(audio_path, sr=None): Loads the specified audio file and returns the waveform (y) and sampling rate (sr). Setting sr=None loads the file at its original sampling rate.

The waveform is returned as a one-dimensional NumPy array containing the amplitude information of the audio signal.

Playing Audio Files

To play the loaded audio data, you can use IPython.display.Audio, which is convenient when working in a Jupyter Notebook environment.

from IPython.display import Audio

# Play the audio
Audio(data=y, rate=sr)

This code allows you to play the audio directly within your notebook, making it easy to verify that the audio file was loaded correctly.

Visualizing Waveforms

LibROSA provides simple methods for visualizing audio waveforms. Using the librosa.display module, you can easily create graphs to show how the audio signal changes over time.

import librosa.display
import matplotlib.pyplot as plt

# Visualize the waveform
plt.figure(figsize=(10, 4))
librosa.display.waveshow(y, sr=sr)
plt.title('Waveform')
plt.xlabel('Time (seconds)')
plt.ylabel('Amplitude')
plt.show()

• librosa.display.waveshow(): Displays the waveform, showing how the amplitude of the audio signal changes over time.

This visualization helps you better understand the time-domain characteristics of the audio signal, such as loudness variations.

Extracting Audio Features

Extracting features from audio data is important for tasks like speech recognition and music analysis. LibROSA allows you to extract features such as spectrograms and MFCC (Mel-Frequency Cepstral Coefficients).

1. Displaying Spectrograms

A spectrogram shows the frequency components of an audio signal over time, which is useful for analyzing the sound’s characteristics.

# Generate a spectrogram using Short-Time Fourier Transform (STFT)
D = librosa.stft(y)
S_db = librosa.amplitude_to_db(abs(D))

# Display the spectrogram
plt.figure(figsize=(10, 4))
librosa.display.specshow(S_db, sr=sr, x_axis='time', y_axis='log')
plt.colorbar(format='%+2.0f dB')
plt.title('Spectrogram')
plt.show()

• librosa.stft(): Performs a Short-Time Fourier Transform (STFT) to generate the spectral data.
• librosa.amplitude_to_db(): Converts the amplitude spectrum to decibel (dB) scale for easier visualization.

2. Extracting MFCC (Mel-Frequency Cepstral Coefficients)

MFCC is a widely used feature in speech recognition systems, and LibROSA makes it easy to extract MFCCs from audio data.

# Extract MFCC features
mfccs = librosa.feature.mfcc(y=y, sr=sr, n_mfcc=13)

# Display the MFCCs
plt.figure(figsize=(10, 4))
librosa.display.specshow(mfccs, sr=sr, x_axis='time')
plt.colorbar()
plt.title('MFCC')
plt.xlabel('Time (seconds)')
plt.ylabel('MFCC coefficients')
plt.show()

• librosa.feature.mfcc(): Extracts MFCC from the audio data. The parameter n_mfcc defines how many MFCC coefficients to extract.
• librosa.display.specshow(): Displays the MFCCs over time, allowing you to visually analyze the audio’s features.

Manipulating Audio Data

LibROSA also provides functions for modifying audio data, such as time stretching and pitch shifting.

1. Time Stretching

To change the speed of the audio, you can use librosa.effects.time_stretch().

# Speed up the audio by 1.5 times
y_fast = librosa.effects.time_stretch(y, 1.5)

# Play the modified audio
Audio(data=y_fast, rate=sr)

2. Pitch Shifting

To change the pitch of the audio, use librosa.effects.pitch_shift().

# Shift the pitch up by two semitones
y_shifted = librosa.effects.pitch_shift(y, sr, n_steps=2)

# Play the modified audio
Audio(data=y_shifted, rate=sr)

These operations allow you to manipulate audio data for tasks like sound design or testing recognition systems with modified audio inputs.

Summary

In this episode, we introduced LibROSA and covered the basics of using this powerful library to process audio data. With LibROSA, you can easily load, play, visualize, extract features, and manipulate audio. In the next episode, we’ll explore more advanced topics, such as visualizing waveform data and analyzing the characteristics of audio signals.

Next Episode Preview

In the next episode, we’ll dive deeper into waveform visualization, explaining how to graph audio signals and analyze them visually. We’ll learn how to use spectrograms and other visualization techniques to better understand audio data.

---

Notes

• MFCC: A feature commonly used in speech recognition and music information processing, which extracts key characteristics of audio.
• Spectrogram: A graph that shows the frequency components of audio data over time, often used for analyzing audio signals.