Electroencephalogram enabled music reccomendation in Android.

Ashwin Nair, Harsh Makwana, Michael Pascale, and Ashwin Jagadeesha.

Final project for COMP.4600 Special Topics in Mobile Sensor Computing at the University of Massachusetts Lowell.

MFCC -> Genre Classification:

• Looked into Spotify API; it does not give us audio to use. There are precomputed features but we will be extracting our own. Made a test app to explore using TarsosDSP to extract Mel Frequency Cepstral Coefficient features from audio files on Android. MFCC processes a Tarsos AudioEvent buffer signalled by AudioDispatcher. Unfortunately the Android version of the library can only read data from the microphone and not from files. Attempted to convert .mp3 to PCM using Android MediaCodec, however this is still not a usable format. Attempted the same with .wav but again Tarsos does not have AudioDispatcherFactory on Android.
• Switched to JavaMP3. Successfully opens loads audio file, decodes to raw array and runs MFCC with Tarsos.
• Spent some time with pandas to curate the music features dataset found at https://github.com/mdeff/fma into a list by genre and MFCC features for genre classification. Random forest in Weka achieves 86% accuracy.
• Can't use the FMA dataset becuase the MFCC algorithm used to generate it is incompatible with Tarsos. Converted files at http://opihi.cs.uvic.ca/sound/mini-genres.tar.bz2 to MP3 and then processed features to make a very tiny dataset we could use as an example.

MUSE: Successfully connected Muse with bluetooth to get stable muse values TP9,AF7,AF8,TP10. Saved these values into CSV.Currently pursuing classification of these values into different emotions. Created a new dataset by watching two different sces 1 comedy and 1 action which was then divided to positive and negative emotion respectively. Added 1 dataset as neutral while meditating. Used Naive Bayes, SVM and Random forest. As we got the heighest accuracy with random forest of 89% we took that into account to random forest to further detect real time emotion.

LIBROSA: Have used librosa in python to get the zero crossing rate; spectral centroid, which indicates where the mass of the sound is located; spectral roll-off, which is frequency below which a specified percentage of the total spectral energy; and calculated MFCCs of the audio signal. Have used PyAudio to get the audio files.