AirUI

Team Members: Aman Bhargava, Alice Zhou, Adam Carnaffan

Goal

We aim to create an audio classification system for various gestures (scratches, taps, swipes, etc.) made on a wooden surface by a user. This will form the basis for a type of natural user interface known as ‘reality user interface’ (hence the name Artificially Intelligent Reality User Interface, or AirUI).

Utility

This user interface method would offer an extremely low cost alternative for touch input in computer systems and could help to make technology more accessible, particularly for individuals who have difficulty using conventional user interfaces.

Rationale for Neural Network Model

Convolutional Neural Networks are a commonly used state-of-the-art method for audio classification, particularly when combined with time-frequency methods such as the short-time Fourier transform (STFT).

Todo List

• [AMAN] Send out ‘pushback’ -- Tuesday 11:59PM
• [AMAN] Data collection video -- Wednesday 11:59PM
• [ALL] Sending out data collection video & request -- Thursday 11:59PM
• [ALICE] Model creation -- Saturday 11:59PM
• [AMAN] Comment Model -- Sunday 11:59PM
• [ADAM] Grid search hyperparameter hunt -- Wednesday November 18 11:59PM
• [ALICE] Comment + Proofread hyperparameter hunt -- Friday November 20 11:59PM
• [ADAM] Final model testing 💕 -- Monday November 23 11:59PM
• [ALICE] Final results presentation -- Monday November 23 11:59PM
• [ALL] Complete slides + presentation rehearsal

Overall Project Architecture

• The data will be pre-processed using audio production software Ableton Live (cropping), Scipy.io.wavfile (loading), Numpy (energy graph), and Scipy (spectrogram).
• Processed data will then be passed into one or multiple conventional ML models such as KNN, SVM, and/or Logistic Regression as a benchmark in the baseline stage.
• MLP will be used in the development stage.
• CNN (convolutional layers followed by fully connected layers) will be used in the final stage.

Results

Baseline Model

Our baseline models included a multilayer perceptron (ReLU activation, 100 hidden neurons, adam solver) for classifying downsampled energy distributions (20 samples per feature vector). Classes included scratches, taps, and 'noise' (in this case, the audio was gathered from lecture to simulate 'real world' noise).

The results are as follows:

=================================================
=== RESULTS FOR INITIAL DATASET BASLINE MODEL ===
=================================================


Total Dataset Size: 		258 Samples
Total Training Set Size: 	193 Samples
Total Validation Set Size: 	65 Samples

Test Set Confusion Matrix: ['scratch', 'tap', 'silence/rose lecture']
[[23  1  0]
 [ 1 16  0]
 [ 0  1 23]]

Training Set Confusion Matrix: ['scratch', 'tap', 'silence/rose lecture']
[[61  1  0]
 [ 0 69  0]
 [ 0  0 62]]

OVERALL SCORE OF MLP on TEST SET: 	93.84615384615384%
OVERALL SCORE OF MLP on TRAIN SET: 	99.48186528497409%