Mathematical Foundations of Real-
Time Noise Reduction in
DeepFilterNet
Prepared by :
Bhumi Choudhary(24BIT0586)
Kaushal More (24BIT0283)
November 8, 2025
Contents
1
1 Introduction
This document explains the mathematical principles and signal-processing methods used
in noise reduction algorithms such as noisereduce and DeepFilterNet. These methods rely
on the Fourier Transform, Short-Time Fourier Transform (STFT), and concepts from
discrete mathematics to analyze and suppress unwanted noise in real time.
2 Digital Audio and Discretization
A continuous sound wave x(t) is sampled into discrete points at a fixed rate known as the
sampling rate fs (in Hz). Mathematically:
where Ts is the sampling interval. For example, at fs = 48,000 Hz, we capture 48,000
amplitude values per second.
Each audio block processed by the model (e.g., blocksize = 1024) represents:
Time window = 3 ms
This short frame allows near real-time response.
3 Fourier Transform and Frequency Analysis
Noise reduction works in the frequency domain. The Fourier Transform (FT)
decomposes a time signal into sinusoidal frequency components:
The inverse transform reconstructs the time-domain signal:
For discrete signals, we use the Discrete Fourier Transform (DFT):
and its inverse:
2
4 Short-Time Fourier Transform (STFT)
The STFT applies the Fourier transform on overlapping time windows, providing both
time and frequency resolution:
where w[n] is a window function (e.g., Hamming window). This yields a spectrogram —
a 2D time-frequency representation used by most noise suppression systems.
Each column of the spectrogram represents one short time frame (e.g., 20 ms), and
each row represents a frequency band.
5 Spectral Subtraction and Noise Estimation
Traditional noise reduction uses spectral subtraction. Assuming the noisy signal is:
y(t) = s(t) + n(t)
its STFT is:
Y (f) = S(f) + N(f)
We estimate noise power |N(f)|2 and subtract it:
|Sˆ(f)| = max(|Y (f)| − α|N(f)|,0)
where α is an over-subtraction factor controlling aggressiveness. Finally, we reconstruct
the enhanced signal using the inverse STFT.
6 DeepFilterNet Approach
DeepFilterNet extends this idea using deep learning. It replaces the manual noise estimate
with a trained neural network Fθ:
Sˆ(f) = Fθ(|Y (f)|,ϕ(f))
where ϕ(f) are phase features and |Y (f)| is the magnitude spectrum. The network learns
a nonlinear mapping from noisy to clean spectral frames using convolutional and
recurrent layers trained on large datasets.
Internally, the model still relies on:
3
• STFT to extract spectral frames.
• Mask prediction M(f) such that Sˆ(f) = M(f) · Y (f).
• Inverse STFT for time-domain reconstruction.
7 Laplace Transform in Continuous Analysis
While the Laplace Transform is not directly used in real-time digital filtering, it provides
theoretical insight into stability and system response. For a continuous signal:
with s = σ + jω. It generalizes the Fourier Transform (when σ = 0) and helps analyze
continuous-time system poles and zeros before discretization.
8 Discrete Mathematics in Signal Processing
Several discrete math principles underlie noise reduction systems:
• Matrix operations: STFT, filtering, and neural networks rely on matrix
multiplications.
• Graph theory: Some models use spectral graph convolutions for structured signal
features.
• Probability: Noise modeling assumes random variables and stochastic processes.
• Discrete convolution:
M−1 y[n] = X
x[k]h[n − k]
k=0
defines how FIR filters work in digital systems.
9 STFT and Deep Learning Integration
DeepFilterNet uses the STFT magnitude |Y (f)| as input features. The model estimates an
ideal ratio mask M(f) such that:
4
Sˆ(f) = M(f) × Y (f)
After mask application, inverse STFT reconstructs the clean signal:
sˆ(t) = ISTFT(Sˆ(f))
This hybrid approach allows the system to retain harmonic content (speech, music) while
suppressing background noise.
10 Block Processing and Real-Time Execution
Real-time filtering divides the audio into short, overlapping blocks:
xi[n] = x[n + iH], i = 0,1,2,...
where H is the hop length. Each block undergoes:
1. STFT transform
2. Neural network enhancement
3. Inverse STFT synthesis
The output stream is reconstructed using overlap-add.
11 Conclusion
Noise reduction in modern systems like DeepFilterNet combines classical DSP (STFT,
spectral subtraction) with deep neural estimation. The Fourier and Laplace transforms
form the mathematical backbone, while discrete mathematics governs computational
implementation and optimization.