People who are blind often listen to screen readers at 450–600 words per minute, roughly three times conversational speed. Increasingly they do this through phone and earbud speakers no bigger than a fingernail. Research on fast-speech intelligibility asks how to compress speech so listeners can still follow it, but it evaluates the output on idealized playback hardware. Separately, materials chemists have found that packing the tiny air chamber behind a microspeaker with hydrothermally grown zeolites makes the chamber behave as if it were larger. Hydrothermal growth means crystallizing the material in a sealed, heated, pressurized vessel. Zeolites are porous crystals whose nanometre pores soak up and release gas molecules as the speaker cone moves. Nobody has asked whether this hardware trick changes how well listeners understand fast speech. Nobody has asked either whether any benefit could simply be reproduced by a digital equalizer.
This note proposes a controlled listening and measurement study to answer that question. The central idea is a test of what only the material can do. A digital equalizer can copy any change the zeolite makes to the speaker's linear frequency response. It cannot copy the mechanical change: because the chamber acts larger, the cone needs to move less to produce the same loudness, and that should mean less distortion. Our hypothesis is that zeolite filling beats a matched equalizer only when the speaker is driven hard and the speech is fast. Distortion products would then smear the dense, short consonant bursts that compressed speech depends on.
Core Claim
Acoustic Enhancement Performance of Hierarchical ZSM-5 Zeolites with Different Si/Al Ratios hydrothermally synthesized ZSM-5 zeolites over a range of silicon-to-aluminum ratios (SAR). It packed them into a 0.3 cm³ microspeaker back cavity and reported a resonance shift of 199.53 Hz and a sound pressure level gain of 4.74 dB at 500 Hz for its best sample (A3, SAR 614). That paper measured tones and frequency curves only. It contains no speech, no intelligibility test and no transient analysis.
We claim that the relevant question for intelligibility is not "does it boost bass." Speech intelligibility lives mostly well above 500 Hz, and extra low-frequency energy could even mask consonants. The relevant question is whether the reduced cone movement lowers distortion enough, at realistic listening levels, to lower transcription errors for 2x–3.5x speech. The claim is falsifiable in a clean way. If an empty cavity with a matched equalizer gives the same word error rate as the zeolite cavity at every speed and drive level, the chemistry offers nothing for this use. We would report that as the finding.
The intelligibility side supplies a second, smaller contribution. Fast-speech work disagrees on whether simple uniform time compression or learned non-uniform duration control is better. Intelligibility analysis of fast synthesized speech found that compressing normal-rate synthetic speech with WSOLA kept word errors under 20% at 3x, while non-linear model-based methods exceeded 40%. WSOLA, or waveform similarity overlap-add, shortens audio by splicing overlapping chunks without changing pitch. Total-Duration-Aware Duration Modeling for Text-to-Speech Systems found that a duration predictor conditioned on target total length cut 2x word error rate from 15.1% to 7.4% on LibriSpeech. Uniform compression shortens consonant bursts most severely in absolute terms. A distorting speaker should therefore hurt it more than duration-aware synthesis, which protects consonant onsets. Running both methods through the same hardware conditions tests that interaction directly.
Research Program
The first step reuses the synthesis route of the ZSM-5 study to produce a SAR series spanning its low-ratio sample (A1, SAR 87) through A3 and A4 (SAR 1233). That study reports A1's acoustic gain at only 2.86 dB, and the research file attributes low-SAR penalties to residual sodium ions blocking the pores. This gives a natural dose-response axis. We will add one deliberately disabled sample: A3 crystals pre-saturated with an adsorbate so that their pores no longer exchange gas, at matched particle size and packing mass. This ablation separates "there are particles in the chamber" from "the particles absorb and release gas."
The second step builds the stimuli. We will synthesize test sentences from the LibriTTS test split with a normal-rate neural speech synthesizer. The synthetic voice is chosen because both Lacuna-indexed rate-control studies show that accelerated normal-rate synthetic speech is clearer than natural fast speech. We will then accelerate the speech to 1.5x, 2x, 2.5x, 3x and 3.5x in two ways. One is WSOLA. The other is a total-duration-aware duration model following the regression variant of the duration paper above, because that variant carries its reported 2x number. Semantically unpredictable sentences will be added for human tests, following Large-Scale Subjective Evaluations of Speech Rate Control Methods for HMM-Based Speech Synthesizers, so that listeners cannot guess words from context.
The third step plays every stimulus through a single microspeaker model in each cavity condition, recorded in a small anechoic box with a measurement microphone. Recordings will be made at three output levels matched by A-weighted loudness: moderate, loud, and near the speaker's rated maximum. Matching loudness rather than input voltage is essential. Otherwise the zeolite condition wins simply by being louder. In parallel we will measure total harmonic distortion and two-tone intermodulation across frequency, plus impulse-response decay time. These measurements tell us whether any intelligibility change tracks distortion, ringing, or neither.
Baselines
The do-nothing reference is the same microspeaker with an empty 0.3 cm³ back cavity, driven with the same loudness-matched stimuli. Every other condition is reported as a change relative to it.
The strongest competitor is the empty cavity plus a digital equalizer. We will design an inverse filter so that the empty cavity's measured small-signal frequency response matches the zeolite-A3 cavity's response to within 1 dB from 100 Hz to 8 kHz. Its gain will be limited to what the amplifier can deliver without clipping. This baseline gets the same measurement procedure, the same loudness matching, and the same tuning effort: filter order and target fit are tuned on held-out tones, never on speech results. It is the strongest competitor because it is free in hardware and reproduces every linear effect of the material. The zeolite condition beats it only if A3's human transcription error is lower by a margin outside the 95% confidence interval of a listener-and-sentence mixed model at matched loudness, at one or more drive levels.
A conventional porous fill, polyester fiber or acoustic foam at the same packed volume, is the traditional-materials baseline. The ZSM-5 paper describes such fills as dissipative rather than compliance-adding, and this condition checks that claim on speech. A physically larger empty cavity, sized to give the same resonance shift as A3, serves as an upper-bound reference. If the zeolite matches it, the "apparent volume" story holds for speech too. The pore-saturated A3 sample is the ablation that isolates gas sorption as the active ingredient.
There is no reported intelligibility number for any of these hardware conditions. No reported result was found on Lacuna for speech intelligibility or word error rate through zeolite-filled or equalized microspeakers. The strongest prior numbers we can anchor to come from elsewhere. On the acoustic side, the ZSM-5 paper reports Δf₀ of 199.53 Hz and ΔSPL@500 Hz of 4.74 dB. On the speech side, the duration paper reports 2x LibriSpeech word error rate of 7.4% for its regression model against 15.1% for its baseline, and the fast-synthesis intelligibility paper reports WSOLA under 20% word errors at 3x. We will first reproduce the speech-side ordering on idealized headphones before moving to the device.
Why not just…?
Why not just equalize the speaker digitally? This is the real threat, and the experiment is built around it. If intelligibility depends only on the linear frequency response, the matched equalizer will tie the zeolite at every level, and we will say so. The only mechanism by which the material could win is lower cone excursion at a given loudness, and therefore lower distortion. That mechanism predicts three things that can be checked: no advantage at moderate level, a growing advantage at high drive, and an advantage that correlates with measured intermodulation rather than with frequency response. If the equalizer instead pushes the empty speaker into clipping at loud levels, that counts as a fair limitation of equalization, not a trick, because real devices face the same ceiling.
Why not just fill the cavity with foam? The ZSM-5 paper states that traditional porous media absorb sound rather than add compliance. The research file reports no foam numbers on speech, though, so this remains an assertion. The foam condition at matched fill volume tests it directly. If foam's added damping shortens ringing and matches the zeolite on fast speech, then damping, not sorption, is what matters. The pore-saturated ablation would show the same result from the other side.
Why not just fix it in software with better duration control, ignoring hardware? Duration-aware synthesis clearly helps on clean playback, halving 2x word error rate in the duration paper. Hardware and synthesis are not exclusive, though. The interesting question is whether device distortion erases or preserves that software gain. If duration-aware speech is robust on the empty cavity and the zeolite adds nothing on top, the practical recommendation becomes "fix it in software." The factorial design of synthesis method × cavity × drive level answers this directly.
Proposed Experiments
The primary measurement is human transcription word error rate on semantically unpredictable sentences. It will be collected from sighted listeners and, separately, from a smaller panel of experienced screen-reader users, because Intelligibility analysis of fast synthesized speech found that rankings of compression methods differ between these groups. Automatic speech recognition word error rate on the full LibriTTS-derived set serves as a cheaper secondary metric. Its use is supported by Utterance Selection for Optimizing Intelligibility of TTS Voices Trained on ASR Data, which found that commercial recognizers correlate with human ratings at r = 0.728–0.876. We will verify that correlation on our own recordings before relying on it.
The decisive check is the A3-versus-matched-equalizer contrast across drive levels and speeds. A level-by-condition interaction in which the zeolite gap grows with drive and with speed would support the excursion-and-distortion hypothesis. The pore-saturated sample should then fall back to the equalizer's error rate, and the SAR series should order A1 < A4 ≈ A3 in benefit, as it does in the paper's acoustic gains. Distortion measurements should predict the word error rate gap across conditions better than frequency-response differences do. A flat result, with the equalizer tying the zeolite everywhere, is equally publishable: it would tell device designers that, for intelligibility, the chemistry buys nothing that software cannot. We will also record humidity and repeat the key contrast after several days of exposure to room air. Zeolites take up water vapor, and an effect that disappears in a pocket is not useful to a screen-reader user.
References
- Acoustic Enhancement Performance of Hierarchical ZSM-5 Zeolites with Different Si/Al Ratios
- Intelligibility analysis of fast synthesized speech
- Total-Duration-Aware Duration Modeling for Text-to-Speech Systems
- LibriTTS: A Corpus Derived from LibriSpeech for Text-to-Speech
- Large-Scale Subjective Evaluations of Speech Rate Control Methods for HMM-Based Speech Synthesizers
- Utterance Selection for Optimizing Intelligibility of TTS Voices Trained on ASR Data