Audio

Correlation, Fourier, spectrograms

In this chapter, you will learn to

• Represent a sound as a signal
• Go from time to frequencies (Fourier, spectrograms)
• Make a machine recognise sounds

Tools: NumPy, SciPy (signal)

4 exercises

Each exercise is graded automatically, in your browser. The full statements are available once you have an account.

1. What this is for in real life

   A sound is just a curve that goes up and down hundreds of times a second, and yet we recognise a voice, a note or a song in it. - Shazam. A few seconds picked up in a noisy café are enough to recognise a track…

2. The ingredients of sound

   Your Shazam is making progress. Now you must analyze a mystery audio signal signal that is a mixture of several notes (pure frequencies). The signal was sampled at sr = 1000 Hz for 1 second (1000 samples). Goal…

3. The sound fingerprint

   Your Shazam now has to identify tracks in a database. The problem: a track lasts several seconds, and the frequencies change over time (the melody!). A single frequency analysis over the whole signal mixes…

4. The machine that listens

   You have dissected sound by hand (correlation, Fourier, spectrograms). Here is what that makes it possible to build: Whisper (OpenAI), a speech recognition model that transcribes speech into text. Open the Colab…