Signals

In code-switched speech between lexically related languages, the language of a word can be graded rather than categorical. We propose that disagreement between human annotators marks these positions and can serve as a calibration target. On a trilingual Kreol Morisien, French, and English broadcast corpus, a classifier trained to reproduce two annotators’ label distributions is well calibrated (expected calibration error 0.008) and, under cross-validation, is more uncertain at the 100 words they disputed than at agreed words (entropy ratio 1.68). We use this signal to gate language-specific low-rank adapters in the decoder of a fine-tuned Whisper large-v3 model, with controls that separate the effects of decoding, adaptation, and the signal itself. Standard decoding inserts an automatically detected language token the fine-tuned model was never trained with; decoding without it accounts for essentially the whole of the overall improvement over the baseline. Against a control decoded identically, language conditioning redistributes errors between the languages rather than reducing them: French word error rate (WER) falls by 1.5 to 2.0 points across three training seeds, Kreol Morisien WER rises by about 0.3 points, overall WER changes by less than 0.09 points, and switch-point error does not improve. Ablations locate the French effect in the decode-time signal: replacing it with a uniform, language-swapped, or shuffled signal returns the system to the level of the uniform control, a model trained without language information, and most of the French gain persists when the classifier’s input transcription is degraded to 56% WER. The annotators’ own labels do no better than the predicted signal, and a one-hot form performs comparably, so the recognition effect does not require the signal’s graded form. An error analysis traces the direction of the redistribution to the baseline’s dominant error class: French words written in Kreol Morisien spelling, which are five times more frequent than the reverse. An equivalence lexicon judged by two native speakers attributes 0.6 to 1.0 points of the baseline’s 11.58% WER to spelling convention rather than recognition. These results rest on one dependence that the design cannot remove: the evaluation annotations also train the classifier whose pseudo-labels train the adapters. The uniform control bounds what this dependence contributes. To our knowledge, this is the first use of annotator disagreement as a calibration target for language identification, and the first language-identity conditioning of a speech recogniser for a creole and its lexifier.

A stroke can affect postural control, balance, and walking, thereby limiting the ability to perform activities of daily living. Although conventional physical therapy supports motor recovery through active walking, researchers have increasingly investigated advanced modalities, such as integrating motor-imagery-based brain–computer interfaces (BCIs) with proprioceptive feedback (PF). This combination aims to engage both the peripheral and central nervous systems. However, this integrated approach remains underexplored, particularly in the domain of lower-limb rehabilitation. To address this gap, this study developed and evaluated a BCI with five stroke survivors. The BCI provided treadmill-based PF at three different speeds, controlled by relative power changes detected in the participants’ brain signals during kinesthetic motor imagery (KMI). Cortical activity was analyzed across three states: baseline period, active walking, and KMI + PF. Specifically, active walking exhibited lower event-related desynchronization values compared to the KMI state combined with feedback. These findings suggest that the detection of relative power changes may provide a basis for establishing a common threshold in future MI classification strategies. In terms of clinical outcomes, preliminary improvements were observed across all five participants. However, the functional ambulation category metric, which reflects the level of assistance required for walking, remained the same in two out of the five participants. In addition, the lack of a control group limits our analysis to a pilot study rather than determining clinical improvement. These results support the potential of integrating brain signals with PF and provide a basis for exploring classification strategies that could reduce calibration time and improve the applicability of BCI systems in clinical settings.

Ripeness recognition is a critical step in postharvest fruit handling. However, publicly available ripeness datasets are typically small and severely imbalanced, with a single dominant class accounting for more than 90% of the samples. Under such conditions, overall accuracy is dominated by the majority class and masks the deficiencies of the model on the minority classes. Accordingly, Macro-F1 is adopted as the primary evaluation metric. A spatial convolutional branch is combined with a wavelet-domain branch through feature-level fusion. Rather than introducing new architectural components, this study focuses on an evaluation protocol that tests whether the gain originates from frequency information. The protocol comprises a capacity-matched dual-spatial control, repetition over five random seeds, and bootstrap confidence intervals. Data-level image synthesis and algorithm-level Focal Loss are further compared for handling class imbalance. On the ripeness dataset, the fusion model achieved a slightly higher mean Macro-F1 (84.50%) than both the capacity-matched control and the single spatial branch. The same advantage was observed on a second mango dataset of surface defects. Data-level image synthesis improved Macro-F1 more than Focal Loss did, and the gains came mainly from backbones with weaker baselines. Because both datasets contain only mangoes, whether the framework generalizes to other fruits or domains remains to be tested.