This is a Room Impulse Response (RIR) data collection and generation project from Gebze Technical University (GTU). Data collection and generation phases are described below. You may also find the data that we collected for our work.
- Collecting real data
- Data
- Generating RIR data using MESH, positions and a GAN model
- Generated vs Real RIR data comparison summary
We built an automated RIR collection system to collect real "room impulse response" data in a room. The system collects 2400 data points ( microphone-speaker position based sound records) in 10 hours.
The system consists of:
- Speakers (4),
- Microphones (6),
- Step motors (2) and their controllers,
- Microphone Stand (1 custom constructed),
- Speaker Stand (1 custom constructed),
- USB hubs (2) and cables,
- Computer (1) .
Following information is given in the data collection directory
- How to construct this system
- How to start recording sounds in a room.
- How to clean records and extract RIRs from that records.
- How to visualize recorded data
- Recording Position Heatmaps
- RIR signal waves
- How to compare recorded RIRs (read data) with a pre-existing model's (FAST-RIR) generated RIR data .
Speakers and microphones are mounted as shown below.
This dataset contains 15.202 RIRs, from 11 different rooms of the GTU Department of Computer Engineering .
- Download : DATA (5GB)
- You may find python file to read this pickle file in [read_data.py][02.data/data_reader/read_data.py]
- [Data properties][02.data/data_properties/README.md]
We took MESH2IR paper as a referrence point, we will modify it to find our own model.
Within this work , we tried to
- Change models
- Change inputs/outputs
We tested the fidelity of generated RIRs using out collected real data (GTU-RIR) as shown below :
We obtained comparison results as follows :
Details are found at 03.data_generation directory.