Split and filtered: Intelligibility effects of band-dependent modulation cutoffs

Prior studies have shown that low modulation cutoffs severely degrade overall speech intelligibility when a single modulation cutoff is applied to the amplitude envelope. This study investigated how a series of amplitude modulations affect speech intelligibility when each cutoff frequency across 14 frequency bands were modulation-filtered independently through subgroups using the TIMIT corpus. Speech signals were first lowpass filtered below 6000 Hz and decomposed into 14 frequency bands based on a 1/3-octave filter bank. After extracting the amplitude envelope for each frequency band using the Hilbert transform, these were downsampled by a factor of 64, lowpass filtered at 7 cutoff frequencies (1, 2, 4, 8, 16, 32, and 64 Hz), then upsampled. Bands were rearranged into 3 broader subbands, corresponding to first, second, and higher formant frequencies. Different modulation frequencies are assigned to each group, yielding a total of 15 435 combinations. Results show that the average Short-Time Objective Intelligibility (STOI) score, computed from 45 sentences, was severely degraded in F1 and F2 regions under low cutoff conditions (especially for 1 and 2 Hz). This outcome may contribute to future privacy-oriented speech processing strategies by identifying the modulation-sensitive area, enabling the suppression of redundant signals while preserving clarity.