Zero-Latency Resonance Suppression for Voice with a Dynamic Filter Bank: An Objective Comparison with Short-Time Fourier Processing

Resonance suppressors attenuate spectral peaks that stand out from a smoothed spectrum. Most use the short-time Fourier transform (STFT), which adds a frame of delay and can produce output ahead of a transient. We evaluate a zero-latency alternative: forty band-pass detectors at sixth-octave spacing, each compared with a weighted neighbourhood average, driving dynamic minimum-phase bell filters. The evaluation uses 582 clips of spoken and sung voice in nine languages, injected resonances of known shape, a speaker-disjoint development set, and STFT baselines of two frame lengths matched to the proposed method in collateral change. On 425 test clips the filter bank reduced a static resonance of 7.06 dB by 1.09 dB at factory depth, against 0.89 and 0.87 dB for the matched baselines; weighted by speaker it was level with the longer transform (1.10 against 1.06 dB). It was ahead on speech and broad resonances, not on the narrowest resonances, and trailed the longer transform on one singing corpus. It produced no pre-echo and did not single out harmonics, but shifted phase and cost about twice the processing time. All blind methods cut formants more than their surroundings, barely improved fidelity to the clean recording, and were far from an oracle notch. A style-conditioned neural equaliser trained on crowd-sourced settings, re-evaluated with every contributor held out, did no better than the mean curve of the style; conditioning on the input spectrum or a brightness scalar did not help. No listening test was conducted.

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-13
DOI
https://doi.org/10.5281/zenodo.24201749
Primary Topic
Speech and Audio Processing
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Zero-Latency Resonance Suppression for Voice with a Dynamic Filter Bank: An Objective Comparison with Short-Time Fourier Processing

YUSUHAO
Zenodo (CERN European Organization for Nuclear Research)
Speech and Audio Processing
preprint

Zero-Latency Resonance Suppression for Voice with a Dynamic Filter Bank: An Objective Comparison with Short-Time Fourier Processing

YUSUHAO
preprint en

Abstract

Resonance suppressors attenuate spectral peaks that stand out from a smoothed spectrum. Most use the short-time Fourier transform (STFT), which adds a frame of delay and can produce output ahead of a transient. We evaluate a zero-latency alternative: forty band-pass detectors at sixth-octave spacing, each compared with a weighted neighbourhood average, driving dynamic minimum-phase bell filters. The evaluation uses 582 clips of spoken and sung voice in nine languages, injected resonances of known shape, a speaker-disjoint development set, and STFT baselines of two frame lengths matched to the proposed method in collateral change. On 425 test clips the filter bank reduced a static resonance of 7.06 dB by 1.09 dB at factory depth, against 0.89 and 0.87 dB for the matched baselines; weighted by speaker it was level with the longer transform (1.10 against 1.06 dB). It was ahead on speech and broad resonances, not on the narrowest resonances, and trailed the longer transform on one singing corpus. It produced no pre-echo and did not single out harmonics, but shifted phase and cost about twice the processing time. All blind methods cut formants more than their surroundings, barely improved fidelity to the clean recording, and were far from an oracle notch. A style-conditioned neural equaliser trained on crowd-sourced settings, re-evaluated with every contributor held out, did no better than the mean curve of the style; conditioning on the input spectrum or a brightness scalar did not help. No listening test was conducted.

Zenodo (CERN European Organization for Nuclear Research)
Korea University (KR)
Speech and Audio Processing
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon