Single-channel dereverberation with Wiener filtering

Dereverberation system for monophonic musical signals, based on a regularized Wiener filter with a known room impulse response (RIR). Implements two filter variants (stationary and short-time, STFT) and an optional post-processing stage (noise gate) to reduce residual background noise in silent/decay portions of the signal.

Developed as part of a bachelor's thesis in Music Informatics(L-31).

What it does

• Dereverberates an audio file (or a folder of files) given a single reference RIR, with automatic parameter estimation based on the reverberation time ($T_{60}$) and direct-to-reverberant ratio ($DRR$) of the RIR itself.
• Provides two filter variants: stationary (a single filter for the whole signal) and short-time (STFT, a filter applied frame by frame).
• Includes an optional noise gate, to attenuate residual background noise in silent/decaying portions without affecting sections with real signal.
• Provides a separate script (conv.py) to build your own test dataset with known ground truth, by convolving an anechoic signal with one or more RIRs.

Installation

git clone <repository-url>
cd <repository-folder>
pip install -r requirements.txt

Usage

Dereverberation (basic, interactive use)

python main.py

The script asks two things: the path to the audio file or folder to process, and the path to the RIR of the environment. Everything else (parameter estimation, temporal synchronization, post-processing) happens automatically with sensible defaults.

Dereverberation (advanced, command-line use)

python main.py -i input.wav -r rir.wav --out_dir output --gate

Main flags:

Flag	Meaning
-i, --input	Audio file or folder to process
-r, --rir	RIR file to use
--out_dir	Output folder
--gate	Also apply the noise gate, in addition to the non-gated output
--no_sync	Disable automatic wet/RIR synchronization — use only if the file is already temporally coherent with the RIR (e.g. generated with conv.py)
--no_auto	Disable automatic parameter estimation from the RIR

Run python main.py --help for the full list of parameters (regularization, temporal windows, STFT parameters, gate parameters).

Building a test dataset with ground truth

python conv.py -d anechoic.wav -r rir.wav -o wet.wav

Or, to convolve with multiple RIRs at once (useful for testing several environments):

python conv.py -d anechoic.wav -r rir_folder/ -o output_folder/

The generated signal is already temporally coherent with the RIR used: process it afterwards with python main.py --no_sync.

Applying the noise gate afterwards

If you already have dereverberated files and want to apply (or re-apply with different parameters) only the noise gate, without re-running the whole dereverberation pipeline:

python apply_noise_gate.py --in_dir input_folder --out_dir output_folder

Add --plots to generate diagnostic plots (waveform, spectrograms, the gain curve applied over time).

Project structure

main.py               Entry point: dereverberation (stationary + STFT)
conv.py             Generates test signals via convolution
noise_gate.py           Noise gate / downward expander library
apply_noise_gate.py     Applies the gate in batch to already dereverberated files
io_utils.py             Support functions for audio reading/writing
rir_utils.py            RIR pre-processing, T60/DRR estimation, automatic parameters
wiener_dereverb.py      Stationary Wiener filter
wiener_stft.py          Wiener filter in the STFT domain
plot_utils.py           Diagnostic plot generation
requirements.txt        Python dependencies

How it works (in brief)

The filter is designed in the frequency domain as:

G(f) = H*(f) / (|H(f)|² + β(f))

where H(f) is the Fourier transform of the RIR and β(f) a regularization term, necessary because real RIRs are generally non-minimum phase: their exact inverse cannot be realized in a stable, causal form. The regularization parameters, RIR truncation duration, and the extent of the filter's temporal window are estimated automatically from the RIR's own estimated $T_{60}$ and $DRR$.

Known limitations

• Single-channel method: does not exploit any multichannel information available in the RIR (e.g. Ambisonics formats).
• As with any Wiener filter using a known RIR, inversion of the late reverberant tail is structurally limited by the regularization required for filter stability — it is more effective on early reflections than on the diffuse late tail.
• Performance degrades in highly reverberant environments (long T60, very negative DRR).

License