Audio Denoising using Spectral Analysis and FIR Filter Design
Overview
This project focuses on audio denoising using digital signal processing techniques.
A sinusoidal interference at 15.2 kHz was added to an audio signal and removed using a linear-phase FIR filter designed under strict real-time DSP constraints.
The workflow includes spectral analysis using Welch PSD estimation, evaluation of windowing parameters, and comparison between multiple FIR filter design methods.
Highlights
- Spectral estimation using Welch’s method (FFT = 1024, 50% overlap)
- Analysis of window types: Kaiser, Rectangular, Hanning, Hamming
- FIR filter design under latency and computational constraints
- Comparison of Kaiser, Equiripple, Blackman, and Least Squares designs
- Real-time audio denoising with strong stopband attenuation (> 60 dB)
Methodology
Spectral Analysis
- Extracted a 3-second audio segment
- Added a 15.2 kHz sinusoidal interference
- Computed single-sided PSD using Welch estimation
- Studied effects of:
- FFT size
- Window type
- Window length
- Overlap percentage
- Sampling frequency
Filter Design
The digital filtering problem required removing a narrowband interference while preserving audio quality.
Constraints:
- Latency < 3.2 ms
- Filter length: 100–200
- Stopband attenuation ≥ 50 dB
- Passband ripple ≤ 0.4 dB
Chosen Solution: A Kaiser-window FIR filter achieved:
- ~75 dB stopband attenuation
- Minimal passband ripple
- Real-time computational feasibility
Code
- The code is available on GitHub