🎧 Spectral Correction Adapticve Deconvolution Web Demo

Pipeline: Multi-resolution window selection → conditioning-aware consensus inverse model → global Wiener filtering.
Execution: Runs fully locally in the browser.
Note: The web demo is relatively slow. For production use, run the MATLAB implementation locally for ~10× performance improvement.

📂 Drag & drop any audio file

🎵 WAV, MP3, OGG, M4A, AAC, etc.

Click here too if you prefer the file picker

No file loaded yet.

⚡ Loading Pyodide...

Visible settings

n_iter

lambda_h

smooth_mu

noise_floor

maxInverseGain

seg_seconds_list

hop_fraction

pool_mult

n_outer

min_rms

cfg.Lh: Length of estimated impulse response (FIR model order). Higher values make the system model more expressive, but increase risk of overfitting.
cfg.n_iter: Number of alternating optimization iterations (x ↔ h updates). Controls convergence vs computation time.
cfg.lambda_h: Ridge regularization on h estimation. Stabilizes inversion and prevents spiky solutions.
cfg.smooth_mu: Smoothness penalty (second derivative). Suppresses ringing and enforces decay behavior.
cfg.noise_floor: Minimum noise floor for stability. Prevents division-by-near-zero amplification.
cfg.maxInverseGain: Caps frequency-domain inverse gain to avoid spectral blow-up.
cfg.seg_seconds_list: Multi-resolution window sizes for segment selection.
cfg.K: Number of selected training segments.
cfg.hop_fraction: Overlap between sliding windows.
cfg.pool_mult: Pre-selection multiplier for candidate pooling.
cfg.n_outer: Outer refinement iterations.
cfg.min_rms: Silence threshold (RMS floor).