Home

Magnolia

We will focus on three independent problems. These are:

1. Removing a source (whose location is static) that is dynamic and loud. Source Removal Details
2. Multiple moving sources (i.e., the channel changes with time) Moving Sources Details
3. Standoff distance acoustic enhancement (dynamic gain adjustment in low SNR) Standoff Distance

The resulting algorithms of interest will be a mix of ICA and simulated beamforming methods.

Dataset and Collection

Sampling rate at 48kHz, but we can apply anti-aliasing and subsampling to simulate lower rate capture devices. Microphones are omnidirectional.

Necessary equipment:

• Microphone
  • LAV microphones (lapel)
  • Omni-directional microphones
  • Vesper microphones and microphone arrays
• Audio interface board, XR-18

Collaboration Efforts

• Microphone Equipment
  • Lab41 POC: Abhinav Ganesh
  • Vesper
  • GELB Music (Redwood City) Riley Bradley, 650-365-8878
  • Notes from Meetings
• Acoustic Modeling
  • Lab41 POC: Patrick Callier
  • Doug James at Stanford.
  • Notes from Meetings
• Algorithm Help
  • Lab41 POC: Karl Ni
  • Arlo Faria at Re-Meeting.
  • Notes from Meetings

Algorithms and Software

This project will be primarily an analysis of cost functions (in supervised and unsupervised settings) that can be used in order to denoise and isolate signals.

Approaches and Exercises

Each teammate will focus on independent research, though work will depend intimately on techniques that are jointly derived. At the end of the challenge, each teammate will write a technical report to include review materials gleaned from the below subsection.

Educational Materials

In addition to independent research, each week, we'll be entertaining implementations from scratch in order to have better control over our software and have a better understanding of the approaches. Each week, we'll have educational sessions focused on the following topics in the context of acoustic signal processing and acoustic machine learning.

Week Number	General Approach	Specific Implementations
1, 2, and 3	Supervised Methods Cost Functions	Inverse Problem Gradient Descent Backpropagation>/li>
4 and 5	Unsupervised Cost Functions	ICA and variants (e.g., RICA) Beamforming
6 and 7		Deep ICA GANs with ICA Cost
8 and 9	Recurrent Methods	CLDNNs (Sainath et al) LSTM+RICA Cost