Repository for my MUMT 605 final project
Full source code available!
The algorithm for extracting HPCP (Harmonic Pitch Class Profile) from audio signals is implemented in show_hpcp.m, which reads a single audio file, computes the instantaneous HPCP for every frame and then generates a plot of the global HPCP vector. This part of the algorithm is functionized in get_hpcp.m.
The key profiles used in key estimation can be plotted with show_profile.m, where the 12 discrete values of both major and minor key profiles are linearly interpolated to get all 36 key profile values. This part is functionized in get_profile.m. These two scripts both call interp_profile.m, which is a function for linear interpolation.
To generate a key estimation for one single audio excerpt, run show_single.m. This script is dependent on the functions get_profile.m and get_hpcp.m. It generates a plot of the correlation values with all 24 key profiles. The highest correlation value is circled in red. The label corresponding to the highest correlation value gives the estimated key.
The key estimation process is functionized in estm_key.m. It makes use of get_hpcp.m and get_profile.m, and calculates a correlation value for each of the 24 key profile vectors. The highest correlation value (tonalness) and its corresponding label are returned.
Putting it all together, the experiment on the test dataset is performed by running main.m. It loops through all audio files in a directory and writes the key estimation result, including the file names, estimated keys, and tonalness values, into a text file.
The implementation is based on this great paper:
Gómez, Emilia. "Tonal description of polyphonic audio for music content processing." INFORMS Journal on Computing 18, no. 3 (2006): 294-304
For more details, please consult the report here on my website.