Introduction
The main idea of this project is to analyze the features of a piece of music, and generate a synchronized visualization of the music. Audio analysis and visualization are all implemented in Matlab. DSP tools including filtering, DFT, auto-correlation are applied in this project.
Data
The only type of data we are using in this project is audio data from music files. The files are all downloaded legally from the internet. Figure 1 below shows the time-domain plot of a piece of music we are using.
Figure 1. Time-domain Plot of the Audio Signal
In order to analyze the frequency content of the audio signal, we take the length-N discrete Fourier transform of the signal. The plot of the DFT magnitude is shown in Figure 2 below:
Figure 2. Frequency Representation of the Audio Signal
In order to study the frequency content of the music at a particular time, we take the spectrogram (Hamming window is used for windowing) of the audio signal. The spectrogram is plotted as shown in Figure 3. The x, y axes represent time and frequency magnitude, respectively. And the color of the plot represents the intensity of the frequency.
Figure 3. Spectrogram of the Audio Signal
The spectrogram in Figure 3 gives us some basic intuition into the music signal. However, in order to develop a robust and accurate pitch-detection algorithm, we have to apply some more advanced techniques including auto-correlation, cepstrum, and wavelet transform. Up to now, the auto-correlation method is found to be most robust for single-pitch detection (Only one pitch presents at the same time). Figure 4 below shows the detection results for four consecutive piano notes from “Little Star”.
Figure 4. Detection Results of Auto-Correlation Method for Four Consecutive Notes
