Our Research

Flagship Research Projects

At Reality AI Lab, we're dedicated to pushing the boundaries of audio technology and artificial intelligence.


Neural Network Models for Spatial Audio Perception

This project investigates the application of neural networks to model human perception of spatial audio cues in complex acoustic environments. Our research utilizes a custom-built hemispherical speaker array with 91 individually controlled channels to study sound localization under various room acoustics conditions. Key research objectives include:

  • Developing neural network architectures capable of predicting human responses to spatial audio stimuli
  • Analyzing the impact of room reflections on sound source localization accuracy
  • Creating algorithms that integrate traditional signal processing with machine learning to compensate for acoustic artifacts
  • Evaluating the real-time performance of these algorithms in simulated and real-world environments

Our methodology combines psychoacoustic experiments, signal processing techniques, and deep learning approaches. We are currently focusing on addressing challenges related to generalization across different acoustic spaces and reducing computational complexity for real-time applications.

Machine Learning Approaches to Acoustic Echo Cancellation

This research explores novel approaches to acoustic echo cancellation in multi-channel audio systems using machine learning techniques. We aim to address limitations of traditional methods in handling non-linear echo paths and dynamic acoustic environments. Our current work includes:

  • Developing hybrid models that combine adaptive filtering with neural network architectures
  • Investigating the use of recurrent neural networks for modeling temporal dependencies in echo paths
  • Implementing and evaluating spectral processing techniques for improved echo cancellation in frequency domain
  • Analyzing methods to preserve spatial audio characteristics while suppressing echoes

The research involves extensive experimentation with various neural network architectures, signal processing algorithms, and real-world acoustic scenarios. We are particularly focused on optimizing these methods for low-latency operation in immersive audio applications, addressing challenges such as computational efficiency and robustness to acoustic variability.

Research Infrastructure - Reality AI Lab

Research Infrastructure

The Reality AI Lab comprises several specialized facilities for research in acoustic signal processing, spatial audio, and machine learning:

Immersive Audio Research Chamber (IARC)

The IARC is an anechoic room with the following specifications:

  • Dimensions: 12m x 15m x 5m
  • Spatial audio projection system: 128-channel hemispherical speaker array
  • Speaker control: Individual precision control for each element
  • Motion capture system: Real-time tracking capability for listeners and sound sources

This facility is used for spatial audio experiments and perceptual studies, allowing for controlled acoustic environments and precise sound field manipulation.

Computing Cluster

The lab's computing infrastructure includes:

  • Number of nodes: 24
  • GPU configuration: Each node equipped with NVIDIA A100 GPUs
  • Primary functions:
    • Training of neural network models on large-scale audio datasets
    • Real-time processing of multi-channel audio streams
    • Development and testing of audio AI algorithms

This cluster is optimized for audio AI workloads, supporting both offline training and real-time processing tasks.

Acoustic Measurement and Prototyping Lab

This lab is equipped with:

  • Microphone array: Dual-sphere 64-channel configuration for sound field capture
  • Laser Doppler vibrometer: For surface vibration analysis
  • Automated turntable system: Used for HRTF measurements

This facility supports detailed acoustic measurements, analysis, and the development of custom audio hardware components.