Does the Basal Ganglia Help Listeners Escape Noise-Induced Word Illusions?

When people hear words in background noise, they do not just miss sounds; they often hear a specific wrong word, and many listeners hear the same wrong word. These "consistent confusions" are reproducible. In one study, 82% of confusions recurred on retest. Most of them cannot be explained by noise simply covering up speech energy. Instead, listeners borrow fragments of the noise and bind them into a false word. A separate literature on brain damage suggests that the basal ganglia, a set of deep brain structures best known for movement control, help track the timing and intensity of prosody: the rhythm, stress and pitch movement of speech. This proposal asks whether that timing system is part of what lets healthy listeners use syllable-level prosody to keep target speech separate from a competing masker.

The test uses Parkinson's disease, which depletes dopamine in the basal ganglia, as a reversible manipulation. Patients are tested on and off their dopamine medication. They hear English tokens that are already known to produce consistent confusions, and each token is "repaired" in one of several ways. The repairs either add energy (raise the signal-to-noise ratio) or add local prosodic prominence (a pitch accent and lengthening on the vulnerable syllable) at matched energy. The core prediction is a selective interaction. Taking patients off dopamine should shrink the benefit from prosodic repair much more than the benefit from extra energy. If it does, that would be the first evidence linking subcortical prosodic timing to the microscopic structure of speech misperception in noise.

Core Claim and Why It Is Open

Three strands of evidence point to this experiment.

The first strand concerns how much energy explains. In Discovering consistent word confusions in noise, energetic masking (noise overlapping speech in time and frequency) accounted for only 24% of consistent confusions. Misperceptions Arising from Speech-in-Babble Interactions used a decoder over "glimpses" (time-frequency regions where speech dominates noise) and found that nearly every babble-induced misperception recruited masker material. It also found that listeners ignored audible target evidence in over 80% of tokens. Together, these papers motivate the claim that confusions are failures of grouping and selection, not merely of audibility.

The second strand concerns pitch. Undoing Misperceptions found that shifting the whole token's pitch by −1 to +3 semitones did nothing to reverse confusions, whereas +3 dB of signal-to-noise ratio reversed 25.2%. That paper constrains the design: a static pitch offset is the wrong prosodic lever. It also leaves open whether pitch and timing changes placed on a specific syllable would work. How prosody improves word recognition supports the idea that local prosody matters. It showed that modeling syllable-level pitch accents and phrase boundaries jointly with acoustics raised clean-speech word recognition accuracy by 2.5%, but it never tested noise. Language Effects in Noise-Induced Word Misperceptions shows that English confusions cluster at word beginnings, where unstressed initial syllables are reduced. This gives us a principled place to put prosodic repair.

The third strand concerns the basal ganglia. Effects of cortical and subcortical brain damage on the processing of emotional prosody found that Parkinson's patients are impaired at identifying emotion from pure prosody. Their impairment was milder than that of right-hemisphere stroke patients and was qualitatively different: they showed blunted intensity calibration. Pell interpreted this as a basal ganglia role in tracking the temporal unfolding of prosodic cues. Computational work describes the basal ganglia as selectively amplifying dominant inputs while suppressing competitors, in Selective amplification using a contracting model of the basal ganglia. It also describes basal ganglia neurons as encoding information through response durations, in Encoding by Response Duration in the Basal Ganglia.

None of this work used noise. The research file reports no study of basal ganglia dysfunction and speech perception under energetic or informational masking. Informational masking is interference from intelligible competing speech that persists even when energy is held fixed.

The claim is therefore narrow and testable. Syllable-local prosodic prominence helps listeners reject masker fragments, and this benefit depends on dopaminergic basal ganglia function in a way that the benefit from added energy does not.

Research Program

Stage one: healthy listeners. We first establish whether prosodic repair works at all in healthy listeners. Stimuli come from the English Consistent Confusion Corpus. It contains 3,207 robust misperceptions under three maskers, each with 15 listeners' responses: speech-shaped noise, three-talker modulated noise, and four-talker babble. We assume, as in the Spanish signal-modification study, that separate target and masker signals can be obtained for remixing; this needs confirming with the corpus authors. We select multisyllabic tokens whose confusion involves the initial or another unstressed syllable. Each token receives one of five treatments, and every manipulated target is renormalized to the original root-mean-square level:

  • the unmodified token;
  • +3 dB signal-to-noise ratio, the energetic repair;
  • a whole-token +2 semitone pitch shift, a replication of the null result;
  • a local prosodic repair: a pitch accent plus about 30% lengthening on the confusion-bearing syllable, made with a vocoder of the STRAIGHT type;
  • the same local repair placed on a different, non-confused syllable.

Stage one has a decision gate. If local prosodic repair does not reverse more confusions than the static shift, the subcortical question loses its lever. We would then report that negative result and stop.

Stage two: Parkinson's patients. Patients are tested twice, on medication and after an overnight withdrawal, in counterbalanced order with different token subsets per session. Age- and audiogram-matched controls are tested twice with the same timing. The key quantity is a selective-benefit index:

$$\Delta = \left(R^{\text{pros}}_{\text{ON}} - R^{\text{pros}}_{\text{OFF}}\right) - \left(R^{\text{snr}}_{\text{ON}} - R^{\text{snr}}_{\text{OFF}}\right)$$

Here $R^{\text{pros}}$ is the reversion rate under local prosodic repair, meaning the fraction of tokens on which the listener reports the target rather than the confusion. $R^{\text{snr}}$ is the reversion rate under +3 dB, and ON and OFF denote medication state. The basal ganglia hypothesis predicts $\Delta > 0$. Generic slowing, fatigue or attention loss predicts $\Delta \approx 0$, because such a deficit would cut both benefits roughly equally. Pell's emotional-prosody identification task, embedded in the same babble, is run as a bridge to the clean-speech literature.

Stage three: modeling. A computational model asks what the patients' errors look like. We fine-tune Whisper on healthy response distributions, following Transfer Learning from Whisper for Microscopic Intelligibility Prediction. We then score patient responses with the hierarchical log-likelihood measures of A framework for the evaluation of microscopic intelligibility models. These measures score, in turn, where an error occurs, what kind of error it is, and which exact word is reported.

The mechanistic test compares two versions of the model:

  • The input-degradation version receives the target with its syllable-rate prosodic contour smoothed, to mimic poor temporal tracking. If off-medication responses are better predicted by this version than by the intact model, the prosodic-timing account is supported.
  • The noise-degradation version receives a generically noisier input at matched loss. If off-medication responses are better predicted by this version instead, the deficit is nonspecific.

Baselines

Unmodified token. The do-nothing reference is the confusion rate on unmodified tokens for each group and session. Every reversion rate is measured against it. Reversion requires the listener to report the target, not merely to abandon the confusion.

Energetic repair. The +3 dB treatment is the strongest existing method for undoing confusions. Undoing Misperceptions reports 25.2% reversion, on Spanish tokens rather than English, so our healthy English figure must be measured fresh. Prosodic repair counts as a real phenomenon in stage one if its reversion rate clearly exceeds the static-shift rate. The paired-token difference must be credible across listeners and tokens in a mixed-effects model. We do not require it to beat +3 dB.

Glimpse-proportion metrics. These are the strongest physical account of why any treatment helps. We compute high-energy glimpse proportion for every manipulated token, following Glimpse-Based Metrics for Predicting Speech Intelligibility in Additive Noise Conditions, which reports correlations of 0.81–0.92 with human scores. We include its change as a covariate. Prosodic benefit must survive that adjustment.

Fine-tuned Whisper. This is the strongest existing predictor of individual confusions. It reports an average log-likelihood of −46 on the English corpus, against −140 zero-shot and −160 random, in Transfer Learning from Whisper for Microscopic Intelligibility Prediction. Both model variants in stage three share its architecture, data and training budget, so only the input degradation differs.

Misplaced prosody. The ablation that isolates the key ingredient is the same prosodic repair placed on the wrong syllable. It carries identical added duration and pitch movement. If it helps as much as the correctly placed repair, the benefit is a general salience or lengthening effect, not syllable-specific tracking.

No reported result was found on Lacuna for any confusion or reversion measure in Parkinson's patients or under noise with basal ganglia dysfunction.

Why not just…?

Why not just insert a basal-ganglia-inspired gating layer into Whisper and see whether microscopic prediction improves? This is the obvious engineering version, and it would be uninformative. Any learned gate or attention layer can absorb the same gains. The Whisper study already found that over 40% of the fine-tuning benefit came from retuning the audio front end, so extra capacity there will help regardless of biology. A better prediction score would not show that the brain does anything similar. Our design reverses the roles: the human basal ganglia is the manipulated variable, and Whisper is the measuring instrument.

Why not just attribute any patient deficit to age, hearing loss or general cognitive slowing? This is the most serious threat. The design answers it in three ways. First, testing patients on and off medication holds the person's age, audiogram and education fixed. Second, the index $\Delta$ subtracts each listener's energetic benefit, which a general deficit should also reduce. Third, matched controls are tested twice to absorb session effects. If off-medication patients lose the energetic and prosodic benefits equally, we will report that the deficit is nonspecific, which is itself a useful negative result.

Why not just explain prosodic benefit with glimpses, since lengthening a syllable creates more of them? This is a genuine confound. Lengthening a syllable adds target time-frequency regions above the noise. Three measures address it: the glimpse-proportion covariate, root-mean-square renormalization, and the misplaced-prosody control, which adds the same glimpse area in the wrong place. If benefit tracks change in glimpse proportion and ignores placement, the prosodic account fails.

Why not just rerun Pell's emotion-identification study in babble? We do include it, but it cannot answer the main question. Five-way emotion accuracy says nothing about which wrong word a listener hears or where in the word the error falls, and those are exactly the measures on which the noise literature has shown that energy accounts break down.

Decisive Checks

The program succeeds if three conditions hold:

  • In healthy listeners, correctly placed local prosodic repair reverses reliably more confusions than both the static pitch shift and misplaced prosody after glimpse adjustment.
  • In patients, $\Delta$ is positive. Off-medication sessions show reduced prosodic benefit with energetic benefit roughly preserved, and controls show no session effect.
  • Off-medication responses are better predicted by the prosody-smoothed Whisper variant than by the generically degraded one, on both the error-location and exact-word log-likelihoods.

The plan specifies in advance what each failure would mean. Failure of the first condition ends the program. Failure of the second with the first intact points to cortical sufficiency for prosodic grouping in noise. Failure of the third alone suggests the deficit is real but not specifically temporal.