Lab 9: 1D Signals and Audio#

Objective#

Learn about:

  • Audio as 1D signals
  • Time and frequency representations
  • Characteristics of audio files
  • Basic audio processing

Setup#

The lab directory has some starter code and files in the lab09 subdirectory, so merge in the pull request and pull changes to your local computer.

I’ve downloaded two files from freesound.org. Do not listen to them - try to figure out which is which by looking at the signals! The files are:

As the attribution suggests, one is breaking glass and the other is a siren.

Task 1: Load and inspect the data#

As usual, it’s a good idea to do some EDA before diving in!

In the starter code, I’ve included two modules for loading audio:

  • wavfile.read() reads a wav file and interprets it as-is, returning a tuple of the sample rate (int) and the signal data (a numpy ndarray).
  • librosa.load() reads a wav file and by default does some basic preprocessing.

  1. Load both of the files using wavfile.read() (ignoring the warnings that pop up). Try to determine the the…

    • Sampling rates
    • Number of Channels
    • Data types
    • Bit depths If there are differences between the files, how does this present a challenge in processing?

  2. Next, load both using librosa.load(). How do the resulting signals differ?

  3. Working with the librosa data, plot the signals in the time domain on the same graph.

  4. Compute the frequency domain representation of both sounds and plot the magnitude spectrum.

    Matplotlib has a function plt.magnitude_spectrum which computes the FFT, scales it by the number of samples, and plots the magnitude in the positive half of the frequency spectrum. Neat. Just remember to pass the Fs parameter so that you get an accurate frequency axis.

  5. Based on the time and frequency plots, can you guess which signal is breaking glass and which is a siren? Remember, high in frequency = short in time.

Task 2: Basic audio processing#

Now that you’ve got the two clips loaded, you can start playing around with some processing. If you’re on your own computer or have a wired headset you can try listening the sounds as you change them with:

import IPython.display as ipd
ipd.Audio(sound, rate=sampling_rate)

which will display a little “play” button in the notebook.

Here’s some suggestions of things to try:

  1. Resample at various sampling rates, e.g. using librosa.resample. How does this affect both sounds? Pay particular attention to the frequency domain!

  2. One useful feature in audio signals is the onset, or when a sound starts. This is more useful for longer signals (like finding the beat in songs), but it could also be useful for aligning the start of a sound in a dataset. Try using librosa.onset.onset_detect to find the time or sample for a “new note” in each signal. Do the results make sense to you?

  3. Try adding “noise” to your sounds, such as random numbers sampled from a normal distribution.

Submit#

As usual, commit and push what you’ve done so I can see your work and give feedback (and points!).