Soundar - Map the World with Sound

We use MLP, Polynomial Regression Model to predict DOA (Direction of Arrival) of binaural audio tracks, and use Lasso Regression Model to predict the distance of audio source.

Features

1. Binaural localization mode, replacing traditional multi-microphone array localization
2. Easier to apply to wearable devices (headphones, hearing aids, etc.)
3. High accuracy in angle prediction ​

Application

1. Assist hearing-impaired individuals in noticing potential sudden threats
2. Aid in detecting mechanical failures in automated production lines
3. Auditory systems for bionic robots/animal robots
4. Enhance surveillance capabilities in security systems ​

Dataset

& The dataset consists of six types of sound sources: sine wave of 130.81, 261.63, 1046.5, ambulance noise, gunshot, fart

• Each dataset format:
  • $R=1\sim 30$, with a tolerance of $0.5$
  • $degree=0,5,10,15,\cdots,175,180$ ​

ML Models

DOA Prediction

• Features: ITD, ILD
• Ouput: DOA
• Selected Models: MLP, Polynomial Regression, GMM (for validation)

Distance Prediction

• Feature: DOA, ITD, ILD, RMS Energy
• Output: R (distance)
• Selected Model：Lasso Regression Model

ITD and ILD data distribution, different colors represent different angles

Results

MLP prediction results for DOA

Polynomial Regression prediction results for DOA

Cartesian coordinate prediction results