Noise Cancellation

Computational mathematics project for denoising local audio datasets with pluggable methods, then evaluating results with SNR and spectrogram correlation.

Repository Layout

• noise_cancellation/methods/: denoising methods. base.py defines the common method interface.
• noise_cancellation/data_prep/: local data preparation helpers for clean piano, speech, instrumental music, and noise mixtures.
• noise_cancellation/audio.py, metrics.py, and plots.py: shared audio IO, evaluation metrics, and SVG plotting utilities.
• experiments/prepare_data.py: refreshes generated local data from data/raw/.
• experiments/run_experiments.py: evaluates a selected method against prepared local data.
• .vscode/launch.json: VS Code launch configs for data prep, method runs, and parameter sweeps.
• data/: tracked local dataset folder.
• outputs/: local generated results folder, ignored by Git.

Methods

Currently registered methods:

least_squares
fft_threshold
l1_norm

Each method inherits from DenoisingMethod in noise_cancellation/methods/base.py and is registered in noise_cancellation/methods/__init__.py.

Data

The experiment runner assumes prepared local data already exists.

data/raw/speech/mini_librispeech/
data/raw/instrumental_music/reverie.ogg
data/raw/noise/subway/
data/clean/<dataset>/
data/noise/<noise_type>/
data/noisy/<dataset>/<noise_type>/

Datasets:

piano_notes_chords
speech
instrumental_music

Noise types:

gaussian: ratio 0.20
distant_speech: ratio 0.40
subway: ratio 0.60

Refresh generated local data only when needed:

python experiments/prepare_data.py --dataset all

Running Experiments

Run a method on all datasets:

python experiments/run_experiments.py --method least_squares --dataset all
python experiments/run_experiments.py --method fft_threshold --dataset all
python experiments/run_experiments.py --method l1_norm --dataset all

Run one dataset:

python experiments/run_experiments.py --method least_squares --dataset speech
python experiments/run_experiments.py --method least_squares --dataset instrumental_music
python experiments/run_experiments.py --method least_squares --dataset piano_notes_chords

For quicker runs, add --limit <N>:

python experiments/run_experiments.py --method least_squares --dataset speech --limit 5

The VS Code Run ... launch configs include commented --limit arguments you can uncomment and edit.

Parameter sweep for least squares:

python experiments/run_experiments.py --method least_squares --dataset all --parameter-sweep

Tune FFT, L1, or least-squares parameters on a smaller clip subset:

python experiments/tune_parameters.py --method fft_threshold --dataset all --clean-clip-limit 2
python experiments/tune_parameters.py --method l1_norm --dataset all --clean-clip-limit 1 --profile focused
python experiments/tune_parameters.py --method least_squares --dataset all --clean-clip-limit 2 --profile focused

L1 tuning is much slower than FFT because each candidate solves many sparse reconstruction windows.

Outputs are organized by method:

outputs/experiments/<method>/<dataset>/denoised/
outputs/experiments/<method>/<dataset>/metrics/
outputs/experiments/<method>/<dataset>/plots/
outputs/experiments/<method>/parameter_sweep/

Tuned Parameters

FFT thresholding was changed from a single full-clip FFT to overlapping window reconstruction. The tuning pass tried window_size values 512, 1024, and 2048; hop_size values at one-half and one-quarter of the window; keep_ratio values from 0.05 through 1.0; f_min = 50; and dataset-specific f_max values. The chosen presets are:

speech: keep_ratio 1.0, window_size 1024, hop_size 256, f_min 50, f_max 8000
instrumental_music: keep_ratio 1.0, window_size 1024, hop_size 256, f_min 50, f_max 4000
piano_notes_chords: keep_ratio 1.0, window_size 2048, hop_size 512, f_min 50, f_max 4000

L1 tuning focused on the highest-impact parameters: lambda_reg, n_freqs, and window_size, while keeping grid_type = linear, f_min = 50, tol = 0.0001, and max_iter = 80. The search tried window_size values 512, 1024, and 2048; n_freqs values 80, 120, and 200; and lambda_reg values from 0.005 to 0.08. The chosen presets are:

speech: n_freqs 200, f_max 8000, lambda_reg 0.08, window_size 512, hop_size 256
instrumental_music: n_freqs 200, f_max 5000, lambda_reg 0.08, window_size 512, hop_size 256
piano_notes_chords: n_freqs 200, f_max 5000, lambda_reg 0.08, window_size 512, hop_size 256

Least-squares tuning focused on K, n_freqs, window_size, hop_size, and f_max, while keeping grid_type = linear and f_min = 50. The focused search tried window_size values 1024 and 2048, n_freqs values 200 and 500, dataset-specific K ratios, and f_max values up to 7000. Numerically unstable high-frequency candidates were skipped if the pseudoinverse SVD did not converge. The chosen presets are:

speech: n_freqs 500, f_max 7000, K 500, window_size 1024, hop_size 256
instrumental_music: n_freqs 500, f_max 4000, K 100, window_size 2048, hop_size 512
piano_notes_chords: n_freqs 500, f_max 4000, K 25, window_size 2048, hop_size 512

Ranked tuning outputs are written under outputs/experiments/<method>/parameter_sweep/.

Adding Methods

New denoising methods should inherit from DenoisingMethod, implement denoise(audio, sample_rate), and register in noise_cancellation/methods/__init__.py.

Method-specific parameters belong under methods.<method_name> in configs/experiment_parameters.json.

Environment

Install dependencies:

pip install -r requirements.txt

The scripts expect Python 3.10+, NumPy, SciPy, and SoundFile. VS Code launch configs set cwd to the workspace root and PYTHONPATH to the workspace root.