An effective solution for stationary audio noise reduction. Created by Cooper Barth ([email protected]), Andrew Finke, and Jack Wiig for EECS 352 - Machine Perception of Music and Audio @ Northwestern University with Professor Bryan Pardo.
Our goal was to implement a useful feature from audio editing software. Additionaly, we sought to better understand how noise reduction was implemented and how to isolate sources from a single sound. We also wanted to explore potential alternatives to current implementations of noise reduction.
Currently, different algorithms are needed depending on the type of sound to be eliminated, such as wind, rain, and background conversation. Stationary noise, or noise that does not change in nature, is commonly removed using Wiener Filtering or Nonnegative Matrix Factorization. One approach to perform this task is to use noise estimate created from a portion of the signal containing only background. The effectiveness of an algorithm that attempts to reduce noise is based on how it maximizes a sample-to-noise ratio, SNR, where the signal is compared to the noise estimate after the signal has been noise reduced.
Removing noise without the original source files presents numerous obstacles. Firstly, preserving the original signal quality so speech is easily understandable is one such challenge. Additionally, properly identifying the background noise is key as an improper classification could lead to the source signal being removed or not a sufficiently level of noise reduced. Because of this, implementing an algorithm that works for most examples of the selected noise type is a difficult undertaking as there are many different types of noise. Using an algorithm designed for a different type of noise than the one present in the input source file could lead to sporadic results. Finally, after completing noise reduction, noise artifacts may still be left over in the output track.
To filter the noisy audio, we choose to use an adaptive Wiener filtering method to reduce the noise in our source audio tracks. The filter utilizes a two-step noise removal and signal repair method, as outlined in the following figure:
The algorithm was tested on dozens of audio files. In each case, random noise was added to a clean audio file and then removed successfully as perceived by the human ear. The algorithm has two major issues. At low initial SNR values, musical noise is still present in the signal. We have taken approaches to reduce this, but more steps must be taken to completely eliminate it. In certain signals, there is some slight distortion caused by the high-pass filter and TSNR that HRNR doesn’t fix. This could be improved by updating the algorithm to focus on this issue.
To quantitatively measure the algorithm, a signal-to-noise ratio (SNR) was used. A higher SNR in a signal indicates a lower magnitude of noise presence. To measure algorithm performance, we used average segment SNR, which measures the average SNR per sound segment. The below table shows that the denoised segment SNR for each audio sample is significantly higher than its corresponding noisy segment SNR.
| Audio Sample | Initial SNR - Noisy | Average Segment SNR - Noisy | Average Segment SNR - Denoised |
|---|---|---|---|
| Buble | 15 | 1.878 | 60.297 |
| Pure Speech | 15 | 1.163 | 31.557 |
| Toto | 15 | 1.461 | 77.042 |
| Boss | 15 | 1.242 | 72.392 |
| Flute | 15 | 1.386 | 16.236 |
- Audio Signal with Stationary Noise
- Denoised Audio Signal
- git clone
https://github.com/cooperbarth/Joint-Audio-Correction-Kit - cd into
/Website/Python Server/ - install the required packages
conda create -n eecs352p --file requirements.txt - to start the local server run
FLASK_APP=server.py FLASK_DEBUG=1 python -m flask run - cd into
/Website/JS Frontend/ - open
index.html
Want to integrate the filter into your own program?:
python3 -m pip install jack-audio