tmartin293 / Musical_Instrument_Classification

Musical Instrument Classification Project Using Beaglebone Black

Musical_Instrument_Classification

Musical Instrument Classification Project Using Beaglebone Black (CSUEB CMPE 344 Class)

• Download and install Debian 9.5 2018-10-07 4GB SD LXQT OS image from https://beagleboard.org/latest-images and move it to a MicroSD card (16 GB minimum recommended size). *

• Follow instructions on http://www.ofitselfso.com/BeagleNotes/Disabling_The_EMMC_Memory_On_The_Beaglebone_Black.php to disable booting from the Beaglebone's EMMC and force booting from the MicroSD card by default. *

• To extend filesystem on MicroSD card:

1. cd /opt/scripts/tools
2. ./grow_partition.sh

• Instructions to install required software on Beaglebone Black to Conda virtual environment:

1. Follow instructions at https://jamwheeler.com/college-productivity/using-a-raspberry-pi-for-instrumentation-software-part-3/ to setup and install:

   • Miniconda3
   • Add rpi channel to conda
   • Create & activate a python 3.6 conda environment
     • conda create -n py36 python=3.6
     • source activate py36

2. conda install -c rpi scipy

3. conda install -c rpi scikit-learn

4. conda install --channel=numba llvmlite

5. conda install -c rpi openblas

6. pip install numpy

7. pip install librosa==0.6.2

8. pip install pyaudio

9. pip install Adafruit_BBIO

10. pip install Adafruit-CharLCD

• Process for Developing Machine Learning Model: Stage 1 – Preprocessing

1. Convert files from .mp3 to .wav

Stage 2 – Feature Extraction

1. Remove leading and trailing silence
2. Extract MFCC values and append instrument label
3. Save all data to .csv file

Stage 3 – Training (Scikit-Learn SVM)

1. C-Support Vector Classification
   1. C = 50
   2. kernel = 'rbf'
   3. gamma = 0.001
   4. decision_function_shape = 'ovr'

Stage 4 – Testing

1. 75/25 train-test split to prevent overfitting and determine model statistics (accuracy, precision, recall, and F1)

Stage 5 – Model Persistence

1. Serialize SVM model with pickle and deserialize when needed for consistent instrument classifications