Paper deep dive
Cleaner Speech, Weaker Generalization: Revisiting Pitt-Derived Benchmarks for Alzheimer's Disease Detection
Luqi Sun, Shreeram Suresh Chandra, Lin Zhang, You-Jin Li, Brian MacWhinney, Yu Tsao, Emily Mower Provost, Berrak Sisman
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 9/2/2026, 4:37:43 AM
Summary
This study investigates the impact of speech preprocessing and dataset curation on the generalization of Alzheimer's Disease (AD) detection models. Using the Pitt Corpus and its derived datasets (Pitt, ADReSS, ADReSSo, ADReSS-M), the authors evaluate supervised deep learning models and Large Audio-Language Models (LALMs). Results indicate that while speech-enhanced datasets improve in-domain performance, they significantly reduce robustness in cross-domain evaluations compared to unprocessed data. This suggests that 'cleaner' speech does not necessarily lead to more reliable real-world AD detection due to distribution shifts.
Entities (24)
Relation Signals (20)
Berrak Sisman → affiliatedwith → Johns Hopkins University
confidence 95% · Berrak Sisman 1 ... 1 Center for Language and Speech Processing (CLSP), Johns Hopkins University
Luqi Sun → affiliatedwith → Johns Hopkins University
confidence 95% · Luqi Sun 1 ... 1 Center for Language and Speech Processing (CLSP), Johns Hopkins University
Pitt → derivedfrom → Pitt-origin
confidence 95% · Pitt is a denoised version of Pitt-origin (Becker et al., 1994) with the same recordings and sample size.
ADReSS → derivedfrom → Pitt-origin
confidence 95% · All speech samples are derived from Pitt-origin (Becker et al., 1994), and have undergone speech enhancement processing
ADReSSo → derivedfrom → Pitt-origin
confidence 95% · All speech samples are derived from Pitt-origin (Becker et al., 1994), and have undergone speech enhancement processing.
DNSMOS → evaluates → Speech Quality
confidence 95% · evaluate speech quality of the five datasets using DNSMOS (Reddy et al., 2021), a widely used non-intrusive speech quality assessment metric.
Audio-Flamingo-3 → istypeof → Large Audio-Language Model
confidence 95% · recent LALMs, including ... Audio Flamingo 3 (Ghosh et al., 2026)
Kimi-Audio → istypeof →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Speech-based Alzheimer's disease (AD) detection increasingly relies on speech-enhanced and curated versions of the Pitt Corpus, where speech enhancement, sample selection, and demographic balancing are often treated as beneficial preprocessing steps. However, whether these transformations improve real-world AD detection or instead affect model generalization and prediction behavior remains unclear. In this work, we revisit the role of speech preprocessing and dataset curation across widely used benchmarks for speech-based AD detection. We evaluate the speech quality of different datasets, the cross-dataset generalization of multiple deep learning models under matched and mismatched enhancement settings, and the behavior of several recent large audio-language models (LALMs). Experimental results show that across multiple supervised speech models, speech-enhanced datasets often improve in-domain performance while reducing robustness in cross-domain evaluation. Matched enhancement between training and test data alleviates, but does not eliminate, this degradation. LALMs show a similar sensitivity: enhanced datasets induce stronger class imbalance and prediction shifts than unprocessed data. These results suggest that speech preprocessing and dataset curation can substantially influence downstream AD detection behavior, indicating that ``cleaner'' speech datasets are not necessarily more reliable for real-world AD detection.
Tags
Links
- Source: https://arxiv.org/abs/2609.00276v1
- Canonical: https://arxiv.org/abs/2609.00276v1
Trouble viewing inline? Open PDF directly →
Full Text
74,105 characters extracted from source content.
Expand or collapse full text
Cleaner Speech, Weaker Generalization: Revisiting Pitt-Derived Benchmarks for Alzheimer’s Disease Detection Luqi Sun 1 , Shreeram Suresh Chandra 1† , Lin Zhang 1† , You-Jin Li 2 , Brian MacWhinney 4 , Yu Tsao 2 , Emily Mower Provost 3 , Berrak Sisman 1 1 Center for Language and Speech Processing (CLSP), Johns Hopkins University 2 Research Center for Information Technology Innovation, Academia Sinica 3 Carnegie Mellon University, 4 University of Michigan, Ann Arbor lsun59, sisman@jhu.edu Abstract Speech-based Alzheimer’s disease (AD) de- tection increasingly relies on speech-enhanced and curated versions of the Pitt Corpus, where speech enhancement, sample selection, and de- mographic balancing are often treated as ben- eficial preprocessing steps. However, whether these transformations improve real-world AD detection or instead affect model generaliza- tion and prediction behavior remains unclear. In this work, we revisit the role of speech pre- processing and dataset curation across widely used benchmarks for speech-based AD detec- tion. We evaluate the speech quality of differ- ent datasets, the cross-dataset generalization of multiple deep learning models under matched and mismatched enhancement settings, and the behavior of several recent large audio-language models (LALMs). Experimental results show that across multiple supervised speech mod- els, speech-enhanced datasets often improve in-domain performance while reducing robust- ness in cross-domain evaluation. Matched en- hancement between training and test data al- leviates, but does not eliminate, this degrada- tion. LALMs show a similar sensitivity: en- hanced datasets induce stronger class imbal- ance and prediction shifts than unprocessed data. These results suggest that speech prepro- cessing and dataset curation can substantially influence downstream AD detection behavior, indicating that “cleaner” speech datasets are not necessarily more reliable for real-world AD detection. 1 Introduction Alzheimer’s disease (AD) is the leading cause of dementia worldwide and poses a rapidly growing social and economic burden (Europe, 2019; Bet- ter, 2023). As the number of individuals living † These two authors contributed equally as second au- thors. We publish our code here:https://github.com/ Sleepwalker554/Back_to_Origin with dementia continues to rise globally (Scheltens et al., 2021), there is increasing interest in develop- ing robust and cost-effective approaches for early AD detection. Among different modalities, speech has emerged as a promising signal due to its non- invasive nature, low collection cost, and sensitivity to cognitive decline (Yang et al., 2022; Ding et al., 2024). In speech-based AD detection, the Pitt Cor- pus (Becker et al., 1994) is the most widely used English benchmark dataset. It contains four tasks, among which the “Cookie Theft” task from the Boston Diagnostic Aphasia Examination (Good- glass and Kaplan, 1983) collected spontaneous speech, and therefore has been the most widely used in research. In this task, all participants (both participants with AD and healthy controls) were asked to describe a picture depicting the “cookie theft” scene. Importantly, the Pitt Corpus exists in two versions: the original Pitt-origin recordings and the noise-reduced Pitt version (Pitt), both dis- tributed through DementiaBank (Lanzi et al., 2023). Since the two datasets are otherwise identical, prior studies often do not explicitly distinguish between them, leaving the impact of speech preprocessing on downstream AD detection models largely unex- plored. Beyond Pitt and Pitt-origin, several influential AD detection challenges have introduced additional Pitt-derived datasets, including ADReSS (Luz et al., 2020), ADReSSo (Luz et al., 2021), and ADReSS-M (Luz et al., 2023). These datasets are all constructed from the “Cookie Theft” picture de- scription task in the original Pitt-origin dataset, but apply different preprocessing and curation strate- gies, such as speech enhancement, volume nor- malization, demographic balancing, and sample selection. While these processing steps are often intended to improve data quality and evaluation fairness, they also alter the original speech distribu- tion and reduce the dataset size. arXiv:2609.00276v1 [cs.SD] 31 Aug 2026 Despite the widespread use of Pitt-derived datasets such as Pitt, ADReSS, ADReSSo, and ADReSS-M, the effects of speech preprocessing and dataset curation on downstream AD detec- tion remain insufficiently understood. Although these datasets originate from the same “Cookie Theft” picture description recordings in Pitt-origin, they apply different preprocessing and selection strategies, including denoising, volume normal- ization, demographic balancing, and sample fil- tering. While such processing can improve per- ceptual speech quality and simplify evaluation, it may also alter the original speech distribution and affect model generalization across datasets and recording conditions. In addition, the reduced dataset size in ADReSS, ADReSSo, and ADReSS- M may increase the risk of overfitting to specific data distributions. As a result, models trained on curated datasets may achieve strong in-domain per- formance while exhibiting reduced robustness in cross-domain settings. This paper revisits how speech preprocessing and dataset curation affect AD detection across multiple modeling paradigms. We evaluate Pitt- origin alongside several widely used Pitt-derived datasets under both in-domain and cross-domain settings, using matched and mismatched speech enhancement conditions and multiple state-of- the-art enhancement methods. Our experiments span diverse AD detection architectures, includ- ing acoustic-feature-based models, self-supervised speech representation models based on XLS- R (Conneau et al., 2021), and Sensitive Layer Se- lection (SLS)-based models (Zhang et al., 2024) built on pretrained speech encoders. We further analyze whether recent LALMs, including Kimi- Audio (Ding et al., 2025), Qwen3-Omni (Xu et al., 2025), Qwen2-Audio (Chu et al., 2024), Audio Flamingo 3 (Ghosh et al., 2026), and ultravox-v0_5- llama-3_2-1b (Fixie.ai, 2025), exhibit similar sen- sitivity to preprocessing under both audio-only and audio + transcript settings. The observed effects re- main consistent across acoustic-feature-based mod- els, pretrained speech representation models, and LALMs, suggesting that preprocessing sensitivity is not limited to a specific architecture. Contributions. This paper makes three main contributions: (1) We provide the first systematic benchmark-level analysis of how preprocessing and dataset curation affect downstream AD detection across widely used Pitt-derived datasets. (2) We show that although speech-enhanced datasets can improve in-domain evaluation performance, they frequently reduce robustness in cross-domain eval- uation, even under matched enhancement settings. (3) We demonstrate that these effects persist across multiple modeling paradigms, including acoustic- feature-based models, pretrained speech represen- tation models, and recent LALMs, indicating that preprocessing sensitivity is a broader modeling phe- nomenon rather than an architecture-specific arti- fact. 2Benchmark Datasets and Related Work 2.1 Datasets The following are six of the most widely used En- glish datasets for speech-based Alzheimer’s disease detection. The sample sizes of these six datasets are shown in Table 1, and all of them are available through DementiaBank (Lanzi et al., 2023). More detailed information about the datasets is shown in Appendix A. 1.Pitt-origin: Pitt-origin (Becker et al., 1994) contains four tasks, including the “Cookie Theft” picture description task (Goodglass and Kaplan, 1983) for all participants, and three tasks only for participants with AD: se- mantic fluency, sentence construction, and story recall. The “Cookie Theft” picture de- scription task requires participants to orally describe the content of the picture. Since this task is based on spontaneous speech and con- tains language samples from both participants with AD and healthy controls, it has been widely used. 2.Pitt: Pitt is a denoised version of Pitt- origin (Becker et al., 1994) with the same recordings and sample size. However, the speech enhancement pipeline used to con- struct the dataset was not publicly released. 3.ADReSS: ADReSS (Luz et al., 2020) was re- leased as part of the ADReSS Challenge held at Interspeech 2020. All speech samples are derived from Pitt-origin (Becker et al., 1994), and have undergone speech enhancement pro- cessing (the enhancement method was not released). The dataset was also matched in terms of age and gender. 4.ADReSSo: ADReSSo (Luz et al., 2021) was released as part of the ADReSSo Challenge 0:000:501:402:303:204:10 Time (min) 0 512 1024 2048 4096 8192 16384 Frequency (Hz) Pitt-origin/ADReSS-M 0:000:501:402:303:204:10 Time (min) 0 512 1024 2048 4096 8192 16384 Frequency (Hz) Pitt 0:000:501:402:303:204:10 Time (min) 0 512 1024 2048 4096 8192 16384 Frequency (Hz) ADReSS 0:000:501:402:303:204:10 Time (min) 0 512 1024 2048 4096 8192 16384 Frequency (Hz) ADReSSo -80 dB -70 dB -60 dB -50 dB -40 dB -30 dB -20 dB -10 dB Figure 1: Mel spectrograms of the same participant with AD across Pitt-origin and its derived datasets. held at Interspeech 2021. All speech sam- ples are derived from Pitt-origin (Becker et al., 1994), and have undergone speech enhance- ment processing (the enhancement method was not released).The dataset was also matched for age and gender. Compared with ADReSS, ADReSSo expanded the sample size. 5.ADReSS-M: ADReSS-M (Luz et al., 2023) was released as part of the ADReSS-M Chal- lenge held at ICASSP 2023. This dataset uses the same speech samples as ADReSSo, with the same quantity. In contrast to ADReSSo, ADReSS-M did not undergo any speech pre- processing. 6.English Lu Corpus: Lu (Lanzi et al., 2023) is an English speech dataset independent of Pitt-origin, collected under different recording conditions and from different participant pop- ulations. Similar to Pitt-origin, the dataset is based on the “Cookie Theft” picture descrip- tion task and contains spontaneous speech recordings without speech preprocessing. DatasetTotal AD Control Pitt-origin / Pitt552309243 ADReSS1567878 ADReSSo / ADReSS-M237122115 Lu562729 Table 1: Sample numbers of datasets. 2.2 Datasets Speech Quality Since Pitt-origin and its four derived datasets orig- inate from the same recordings, we visualize Mel spectrograms for the same participant across all five datasets. As shown in Figure 1, background noise is substantially more prominent in the unprocessed datasets (Pitt-origin and ADReSS-M) than in the denoised datasets (Pitt, ADReSS, and ADReSSo). The denoised datasets exhibit cleaner spectral pat- terns and reduced background interference, indicat- ing that preprocessing introduces noticeable distri- butional differences across the datasets. To further quantify these differences, we eval- uate speech quality of the five datasets using DNSMOS (Reddy et al., 2021), a widely used non-intrusive speech quality assessment metric. DNSMOS can evaluate speech samples from three dimensions, including speech quality (SIG), background-noise quality (BAK), and overall qual- ity (OVRL). As shown in Table 2, the denoised datasets consistently achieve substantially higher BAK scores than Pitt-origin and ADReSS-M, while SIG scores remain relatively similar across datasets. These results confirm that speech enhancement sig- nificantly alters background-noise characteristics. DatasetOVRL↑ SIG↑ BAK↑ Pitt-origin2.30732.98192.7290 Pitt2.56063.04793.6024 ADReSS2.59233.08573.5945 ADReSSo2.52223.03733.1829 ADReSS-M2.31282.98052.6973 Table 2: DNSMOS scores of Pitt-origin and its derived datasets. (SIG: speech quality; BAK: background noise quality; OVRL: overall speech quality.) 2.3 Existing AD Detection Models Early AD detection studies primarily explored tra- ditional machine learning approaches. For example, the ADReSS (Luz et al., 2020), ADReSSo (Luz et al., 2021) and ADReSS-M (Luz et al., 2023) Challenges adopted support vector machines as a baseline, while Hason et al. (Hason and Krishnan, 2022) used random forests for speech-based AD detection. More recently, deep learning has become the dominant paradigm in AD detection. Existing stud- ies have explored hybrid architectures incorporat- ing disfluency-related paralinguistic features (Jin et al., 2023), end-to-end ASR-based frame- works (Lin et al., 2024), and transformer-based models for speech-level AD recognition (Dong et al., 2024). Pu and Zhang et al. (Pu and Zhang, 2025) further incorporated pause information into Transformer-based language models for AD detec- tion. Recent work has also begun exploring large language models and large audio-language models for AD detection. For example, Heitz et al. (Heitz et al., 2025) used GPT-4 to extract semantic fea- tures from spontaneous speech transcripts, while Park et al. (Park et al., 2025) applied Chain-of- Thought reasoning and supervised fine-tuning for AD detection. 3 Methodology 3.1 Preprocessing and Distribution Shift Speech enhancement and denoising methods in- evitably modify the acoustic characteristics of speech signals (Xia et al., 2020). Common ef- fects include enhancement artifacts (Iwamoto et al., 2022), spectral smoothing (Li et al., 2018), and other forms of signal distortion. These distortions not only affect the naturalness of speech but may also have a negative impact on downstream mod- els. For example, prior work in automatic speech recognition (ASR) has shown that speech enhance- ment does not necessarily improve downstream recognition performance and can even reduce ro- bustness under mismatched conditions (Yoshioka et al., 2015; Chen et al., 2018; Menne et al., 2019). At the same time, speech enhancement methods, whether based on traditional signal processing or deep learning approaches, can shift the distribu- tion of naturally recorded speech, thereby intro- ducing a significant domain shift (Scheibler et al., 2024; Frankel et al., 2025). This issue is partic- ularly relevant for Alzheimer’s detection bench- marks. Widely used datasets such as Pitt, ADReSS, and ADReSSo are constructed from Pitt-origin us- ing different preprocessing and curation pipelines, while the exact enhancement procedures were not publicly released. As a result, models trained on these datasets may learn representations associated with processed speech distributions rather than nat- urally recorded speech. However, real-world clini- cal speech is typically collected under uncontrolled recording conditions and does not undergo stan- dardized enhancement. This mismatch raises an important question: Do speech enhancement, preprocessing, and dataset curation improve the robustness of AD detection models, or instead introduce distribution shifts that reduce generalization across datasets and record- ing conditions? 3.2 AD Detection Models Traditional Deep Learning Models. We evaluate enhancement effects across three AD detection ar- chitectures with increasing modeling complexity: an acoustic-feature-based eGeMAPS model (Ey- ben et al., 2015), a self-supervised speech repre- sentation model based on XLS-R (Conneau et al., 2021), and a Sensitive Layer Selection (SLS)-based model (Zhang et al., 2024) built on pretrained speech encoders. Model architectures are illus- trated in Appendix D. Among these models, we use the SLS-based model as the primary architec- ture due to its strong performance in recent speech classification tasks. The model extracts frame- level representations from a frozen pretrained XLS- R encoder and applies Sensitive Layer Selection (SLS) (Zhang et al., 2024) to learn weighted combi- nations of Transformer-layer representations, and then performs downstream classification through a classification head. During the model training phase, we use an 80%/20% train-validation split for each dataset with no speaker overlap. Models are trained only on training data under both in-domain and cross- domain settings. For the SLS-based model, we use a learning rate of3× 10 −4 , AdamW optimization, and cross-entropy loss. Training is performed for up to 40 epochs with early stopping patience of 5. All experiments are repeated across five random seeds. All models are trained on NVIDIA RTX 5090 GPUs. Additional implementation details are provided in our released code. Large Audio-Language Models. To investigate the impact of speech enhancement on large audio- language models (LALMs) in AD detection, we additionally evaluate five recent and widely used LALMs: Kimi-Audio (Ding et al., 2025), Qwen3- Omni (Xu et al., 2025), Qwen2-Audio (Chu et al., 2024), Audio Flamingo 3 (Ghosh et al., 2026), and ultravox-v0_5-llama-3_2-1b (Fixie.ai, 2025). We first evaluate all models under the zero-shot setting using audio-only input. Subsequently, we further evaluate the best-performing Kimi-Audio under zero-shot and few-shot settings using audio-only input and audio + transcript input. Prompts are provided in Appendix B. 3.3 Speech Enhancement Methods To evaluate whether all different speech enhance- ment methods affect AD detection models, we ap- ply five representative speech enhancement mod- els to the downstream AD detection architec- tures introduced in Section 3.2. These models span multiple enhancement paradigms, including waveform-based enhancement (Denoiser (Défos- sez et al., 2020)), complex-domain enhancement (FRCRN (Zhao et al., 2022)), Transformer-based enhancement (MossFormer (Zhao and Ma, 2023)), semantic-conditioned enhancement (Resemble (Liu et al., 2024)), and a recent Mamba-based enhance- ment model (MAP-SEMamba (Li et al., 2025)). 4 Speech Enhancement Impact on Deep Learning Models We use the three AD detection architectures men- tioned in Section 3.2 to evaluate the impact of preprocessing on AD classification, including an acoustic-feature-based eGeMAPS model, an XLSR-based self-supervised speech model, and an SLS-based model. This section focuses on the SLS-based model, while results for the eGeMAPS and XLS-R models are provided in Appendix G and Appendix F. Across all architectures, the ob- served preprocessing effects remain broadly con- sistent, indicating that the phenomenon observed in this paper does not depend on a particular model architecture, but is broadly present. 4.1 Cross-Dataset Generalization To evaluate the impact of speech enhancement on model generalization, we conduct both in-domain and cross-domain experiments on Pitt-origin and its four derived datasets introduced in Section 2. For in-domain evaluation, each dataset is split into 80% training and 20% validation sets with no speaker overlap. We use the English Lu Corpus as an inde- pendently collected cross-domain test set to evalu- ate the model’s cross-domain generalization ability. As shown in Figure 2, Pitt-origin and its de- noised counterpart Pitt achieve comparable in- domain performance, with Pitt performing slightly better. In contrast, ADReSS and ADReSSo further improve in-domain Macro-F1 scores. A possible reason is that speech enhancement, normalization, and dataset curation reduce distributional complex- ity of the ADReSS and ADReSSo datasets and simplify the learning problem, thereby facilitating in-domain overfitting. Pitt-originPittADReSSADReSSoADReSS-M Dataset 0.00 0.20 0.40 0.60 0.80 Macro-F1 0.7910 * 0.7449 * 0.8120 * 0.6300 * 0.8745 * 0.4166 0.8964 * 0.6754 * 0.7713 * 0.5470 Same domain Test Test on Lu SLS-based Model Figure 2: SLS-based model generalization test on raw data. (Same Domain Test: evaluation on the test split of the same dataset used for training; Test on Lu: evaluation on the Lu dataset. Asterisks ∗ : performance significantly different from the binomial majority-class baseline,p < 0.05.) However, this trend changes under cross-domain evaluation. Models trained on the original Pitt- origin data achieve the strongest generalization performance on the Lu corpus, whereas models trained on denoised datasets exhibit clear perfor- mance degradation. These results suggest that speech-enhancement-induced distribution shifts re- duce cross-dataset generalization.Meanwhile, we found that although ADReSSo achieves rel- atively strong cross-domain performance on the SLS-based model, this performance is not consis- tently observed across the eGeMAPS and XLS-R models (Appendix G.1, Appendix F.1), indicating that the effect is architecture-dependent rather than a universal advantage of the dataset itself. Overall, although speech enhancement and dataset curation may improve in-domain evalua- tion performance, these gains do not consistently translate to cross-dataset robustness. In contrast, models trained on unprocessed Pitt-origin data ex- hibit stronger cross-dataset generalization. 4.2 Speech-Enhanced Test Sets Using the speech enhancement methods introduced in Section 3.3, we construct five speech-enhanced versions of the Lu corpus. We train models on Pitt- origin and its four derived datasets (using the 80% training split of each dataset) and perform cross- domain evaluation on both the original Lu corpus and its enhanced variants. The experimental results are shown in Figure 3. The model trained on Pitt-origin achieves the best overall performance on the original unprocessed Lu corpus (Macro-F1 = 0.7449), indicating stronger Same Domain Test LuLu -Denoiser Lu -FRCRN Lu -MossFormer Lu -Resemble Lu -MAP-SEMamba Test Set Pitt-origin Pitt ADReSS ADReSSo ADReSS-M Training Set 0.7910 * 0.7449 * 0.6552 * 0.6840 * 0.6620 * 0.7004 * 0.6811 * 0.8120 * 0.6300 * 0.5791 * 0.6103 * 0.6300 * 0.6103 * 0.6247 * 0.8745 * 0.4166 0.6621 * 0.6718 * 0.6296 * 0.6476 * 0.6707 * 0.8964 * 0.6754 * 0.7150 * 0.6822 * 0.7280 * 0.7432 * 0.7306 * 0.7713 * 0.54700.53750.50510.48250.4349 0.6284 * 0.5 0.6 0.7 0.8 Macro-F1 SLS-based Model Figure 3: SLS-based model generalization test on speech-enhanced data. (Same Domain Test: evaluation on the test split of the same dataset used for training. Asterisks ∗ : performance significantly different from the binomial majority-class baseline, p < 0.05.) robustness to unprocessed cross-domain speech. Similar trends are observed for the eGeMAPS and XLS-R models (Appendix 15, Appendix 19). How- ever, performance decreases in most cases when the same models are evaluated on speech-enhanced versions of Lu. In contrast, models trained on ADReSS and ADReSSo show clear performance improvements on speech-enhanced Lu test sets. For example, the model trained on ADReSSo achieves a Macro-F1 score of 0.7432 on Lu-Resemble, ap- proaching the best overall result. This suggests that enhanced training data generalize more effectively to similarly enhanced test conditions. Neverthe- less, these results still do not surpass the perfor- mance of Pitt-origin on the original Lu corpus, and performance varies substantially across different enhancement methods. Overall, models trained on unprocessed speech perform best on unenhanced test data, whereas models trained on speech-enhanced datasets con- sistently perform better on similarly enhanced test sets. This suggests that alignment between training and testing enhancement conditions plays an im- portant role in cross-dataset generalization. How- ever, because the preprocessing pipelines used in Pitt, ADReSS, and ADReSSo were not publicly released, it is impossible to reproduce their exact enhancement conditions during evaluation, lead- ing to substantial performance variability across different enhanced test sets. These findings further raise an important ques- tion: If the same speech enhancement method is applied to both training and test data, can stronger generalization performance be achieved? To an- swer this question, we conduct matched enhance- ment experiments in Section 4.3, where the same speech enhancement methods are applied to both training and testing data. 4.3 Matched Enhancement Settings To determine whether the observed performance degradation is caused primarily by enhancement mismatch between training and test data, we con- struct a matched enhancement setting. Specifically, we apply five speech enhancement methods intro- duced in Section 3.3 to the Pitt-origin training data and evaluate models on two Lu test conditions: (1) unprocessed Lu speech (mismatched) and (2) Lu speech enhanced using the same enhancement method as the training data (matched). We addi- tionally retain the model trained on unprocessed Pitt-origin data as a control (raw-trained), with evaluation on the original Lu corpus serving as the baseline. The experimental results are shown in Figure 4. Across all five enhancement methods, the matched setting consistently outperforms the correspond- ing mismatched setting. These results suggest that alignment between training and testing enhance- ment conditions partially alleviates enhancement- induced distribution mismatch. However, even un- der the matched setting, models trained on unpro- cessed Pitt-origin data (raw-trained) still achieve higher Macro-F1 scores in most cases. Moreover, all enhancement-related settings remain below the baseline. These results suggest that the observed performance degradation is not caused solely by train-test enhancement mismatch. Instead, speech DenoiserFRCRNMossFormerResembleMAP-SEMamba Speech Enhancement Method 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Macro-F1 Baseline=0.7449 * 0.5406 0.6793 * 0.6171 * 0.6917 * 0.6616 * 0.6433 * 0.7072 * 0.6476 * 0.6913 * 0.7364 * 0.6552 * 0.6840 * 0.6620 * 0.7004 * 0.6811 * Mismatched: Enhanced Pitt-origin Lu Matched: Enhanced Pitt-origin Enhanced Lu Raw-trained: Pitt-origin Enhanced Lu Baseline: Pitt-origin Lu = 0.7449 * SLS-based Model Figure 4: SLS-based model performance comparison with matched and mismatched speech enhancement methods. (Mismatched: enhanced training set with raw test set; Matched: the same enhancement for training and test set; Raw-trained: raw training set with enhanced test set; Baseline: raw training and raw test set. Asterisks ∗ : performance significantly different from the binomial majority-class baseline, p < 0.05.) enhancement itself alters the underlying speech distribution and reduces cross-dataset robustness. Similar trends are observed for the eGeMAPS and XLS-R models (Appendix 20 and Appendix 16), in- dicating that the effect is not limited to a particular model architecture. Overall, although matched enhancement condi- tions partially reduce distribution mismatch, they do not fully recover the generalization performance achieved by models trained on unprocessed speech. These findings suggest that speech enhancement itself can substantially affect downstream model robustness in cross-dataset evaluation. 5 Speech Enhancement Impact on Large Audio-Language Models To evaluate the impact of speech enhancement on large audio-language models (LALMs) in Alzheimer’s disease detection, we evaluate five re- cent and widely used LALMs: Kimi-Audio (Ding et al., 2025), Qwen3-Omni (Xu et al., 2025), Qwen2-Audio (Chu et al., 2024), Audio Flamingo 3 (Ghosh et al., 2026), and ultravox-v0_5-llama- 3_2-1b (Fixie.ai, 2025) under multiple inference settings. 5.1 Audio-Only Settings We first evaluate multiple LALMs on all datasets using audio-only input to analyze model behavior when relying solely on speech signals. The ex- perimental results are shown in Appendix C. The results show that most models struggle to reliably distinguish AD from Control under zero-shot audio- only evaluation, while only Kimi-Audio consis- tently achieves statistically significant performance across most datasets. For example, Qwen2-Audio tends to classify nearly all speech samples as Con- trol, whereas Audio Flamingo 3 tends to classify nearly all samples as AD. In contrast, Kimi-Audio achieves the strongest overall performance and re- mains statistically significant on all datasets except the Lu corpus. Based on the above results, we select Kimi- Audio as the main model for further experiments and analysis. We evaluate Kimi-Audio under two inference settings, zero-shot and two-shot. Here, two-shot means that two labeled examples are pro- vided at the inference stage, one from the AD class and one from the control class, without any param- eter update or fine-tuning of the model. Test SetMacro-F1 F1 Score F1 Score (AD + ) (Control + ) Zero-Shot Pitt-origin0.6666 ∗ 0.6726 ∗ 0.6605 ∗ Pitt0.5924 ∗ 0.7300 ∗ 0.4548 ∗ ADReSS0.5629 ∗ 0.6834 ∗ 0.4425 ∗ ADReSSo0.6348 ∗ 0.7273 ∗ 0.5424 ∗ ADReSS-M0.7000 ∗ 0.6899 ∗ 0.7102 ∗ Two-Shot Pitt-origin0.6251 ∗ 0.7198 ∗ 0.5305 ∗ Pitt0.5878 ∗ 0.7409 ∗ 0.4348 ∗ ADReSS0.5879 ∗ 0.6842 ∗ 0.4915 ∗ ADReSSo0.6503 ∗ 0.7291 ∗ 0.5714 ∗ ADReSS-M0.6441 ∗ 0.5941 ∗ 0.6941 ∗ Table 3: Kimi-Audio zero-shot and two-shot perfor- mance in the audio-only scenario. (AD + / Control + : F1 score with AD/Control as positive. Asterisks ∗ : performance significantly different from the binomial majority-class baseline, p < 0.05.) The experimental results of Kimi-Audio are shown in Table 3. On the unprocessed Pitt-origin and ADReSS-M datasets, the model produces rela- tively balanced predictions across AD and Control classes. This suggests that when the input speech is closer to the distribution of unprocessed speech, the model is more likely to rely on relatively sta- ble acoustic patterns. In contrast, on the speech- enhanced datasets (Pitt, ADReSS, and ADReSSo), the model exhibits clear class imbalance, with sub- stantially higher F1 scores with AD as positive than F1 scores with control as positive. These results suggest that speech enhancement substantially al- ters LALM prediction behavior under audio-only evaluation. This trend remains consistent under both zero-shot and two-shot settings, suggesting that the observed bias is not resolved by providing a small number of labeled examples. Overall, the audio-only experiments suggest that speech enhancement does not consistently improve the discriminative ability of LALMs and may in- stead introduce systematic prediction bias, thereby reducing robustness across datasets. 5.2 Audio + Transcript Settings The previous results raise a natural question: Can textual information mitigate the preprocessing sen- sitivity observed in audio-only evaluation? To in- vestigate this, we further evaluate the audio + tran- script input format on Pitt-origin and Pitt using Kimi-Audio under both zero-shot and two-shot in- ference settings. The results are shown in Table 4. When using audio + transcript input, a trend consistent with the audio-only setting can still be observed. On the denoised Pitt dataset, the model exhibits sub- stantial class imbalance, which becomes even more pronounced under some settings. In contrast, pre- dictions on the unprocessed Pitt-origin dataset re- main more balanced across AD and Control classes. These results suggest that adding transcript infor- mation does not eliminate enhancement-induced prediction bias. In addition, the two-shot setting also fails to substantially improve this issue. On Pitt, the model continues to exhibit a strong bias toward the AD class. This suggests that simply pro- viding a small number of labeled examples cannot effectively correct the prediction shift caused by changes in data distribution. Overall, the multimodal experiments further suggest that speech enhancement substantially af- fects downstream LALM prediction behavior, even when textual information is available. Test SetMacro-F1 F1 Score F1 Score (AD + ) (Control + ) Zero-Shot Pitt-origin0.6514 ∗ 0.6350 ∗ 0.6678 ∗ Pitt0.47560.71640.2348 Two-Shot Pitt-origin0.6651 ∗ 0.7313 ∗ 0.5988 ∗ Pitt0.5498 ∗ 0.7194 ∗ 0.3801 ∗ Table 4: Kimi-Audio zero-shot and two-shot perfor- mance in audio + transcript scenario. (AD + / Control + : F1 score with AD/Control as positive. Asterisks ∗ : performance significantly different from the binomial majority-class baseline, p < 0.05.) 6 Conclusion This paper revisits the role of speech preprocessing and dataset curation in speech-based Alzheimer’s disease (AD) detection across widely used Pitt- derived benchmarks. We systematically evaluate the effects of speech enhancement using multi- ple experimental settings, including in-domain and cross-domain evaluation, speech-enhanced test con- ditions, matched and mismatched enhancement set- tings, and multiple speech enhancement pipelines. We further analyze whether these effects remain consistent across different modeling paradigms, in- cluding acoustic-feature-based models, pretrained speech representation models, and recent large audio-language models. Across both supervised AD detection models and LALMs, the results con- sistently show that speech-enhanced datasets often improve in-domain evaluation performance while reducing cross-dataset robustness. Even when matched enhancement conditions are applied to both training and test data, models trained on unpro- cessed Pitt-origin speech generally achieve stronger generalization performance. In addition, LALMs exhibit systematic prediction bias under speech- enhanced conditions, suggesting that preprocessing sensitivity extends beyond conventional supervised architectures to modern foundation audio models. Overall, our findings show that speech enhance- ment and dataset curation are not merely prepro- cessing choices, but factors that can substantially influence downstream AD detection behavior and cross-dataset generalization. These results high- light the importance of carefully considering pre- processing pipelines when constructing and evalu- ating speech-based AD detection benchmarks, and suggest that cleaner speech does not necessarily lead to more robust AD detection systems. 7 Limitations Our research focuses on Alzheimer’s disease speech detection from speech in English contexts. In this field, the vast majority of studies have used the Pitt Corpus (Becker et al., 1994) and its derived datasets, and the number of publicly available En- glish AD speech datasets independent of the Pitt Corpus is very limited. We note that although there are some larger English AD speech datasets, these datasets have not yet been made publicly avail- able and therefore cannot be used for experimental validation in this study. Other publicly accessi- ble English dementia datasets, such as the VAS Corpus (Liang et al., 2022) and the English WLS Corpus (Herd et al., 2014), are not restricted to AD patients, as they include individuals with MCI or other forms of cognitive decline without distin- guishing these diagnostic groups. To ensure the rigor of the research, we did not use these datasets in our experiments. Therefore, due to the limited number of publicly available English AD speech datasets independent of the Pitt Corpus, this paper selects the English Lu Corpus (Lanzi et al., 2023) as the cross-domain test set to evaluate the model’s generalization ability. Our research primarily focuses on the impact of speech enhancement on speech-based AD detec- tion models, with particular attention to the rela- tionship between speech enhancement processing and model generalization ability. Therefore, we only use text as an auxiliary diagnostic tool for LALMs and do not conduct an in-depth study on using text to detect AD. Although our experimental results consis- tently show that, compared with Pitt, ADReSS, ADReSSo, and ADReSS-M, models trained on the original Pitt-origin generally have stronger gen- eralization ability, it should be particularly noted that Pitt-origin itself also contains substantial en- vironmental noise. How to improve the robust- ness of Alzheimer’s disease detection models to noise remains an important and unresolved prob- lem. Future work needs to explore more robust model architectures, enabling models to stably ex- tract Alzheimer’s disease-related pathological fea- tures from noisy speech. Use of AI Assistants We used GPT-5.5 and Claude Opus 4.7 for code as- sistance, and the same models purely for language- related assistance in writing the paper. Licensing We use the DementiaBank dataset, a shared mul- timedia database for studying communication in dementia.DementiaBank is part of Talk- Bank.Therefore, use of the data is gener- ally governed by the Creative Commons Attribu- tion–NonCommercial–ShareAlike 3.0 license (C BY-NC-SA 3.0), unless otherwise specified. This li- cense requires attribution, restricts commercial use, and requires share-alike for derivatives. Access to DementiaBank is password-protected and re- stricted to approved members of the DementiaBank consortium. We comply with the DementiaBank data use requirements, including restrictions on re- distributing or posting password-protected data on external websites or servers. References James T Becker, François Boiler, Oscar L Lopez, Ju- dith Saxton, and Karen L McGonigle. 1994. The natural history of Alzheimer’s disease: description of study cohort and accuracy of diagnosis. Archives of neurology, 51(6):585–594. MAPPING A Better. 2023. Alzheimer’s disease facts and figures. Alzheimers Dement, 19(4):1598–1695. Szu-Jui Chen, Aswin Shanmugam Subramanian, Hainan Xu, and Shinji Watanabe. 2018. Building State-of-the-art Distant Speech Recognition Using the CHiME-4 Challenge with a Setup of Speech En- hancement Baseline. In Interspeech 2018, pages 1571–1575. Yunfei Chu, Jin Xu, Qian Yang, Haojie Wei, Xipin Wei, Zhifang Guo, Yichong Leng, Yuanjun Lv, Jinzheng He, Junyang Lin, and 1 others. 2024. Qwen2-audio technical report. arXiv preprint arXiv:2407.10759. Alexis Conneau, Alexei Baevski, Ronan Collobert, Ab- delrahman Mohamed, and Michael Auli. 2021. Un- supervised Cross-Lingual Representation Learning for Speech Recognition. In Interspeech, pages 2426– 2430. Ding Ding, Zeqian Ju, Yichong Leng, Songxiang Liu, Tong Liu, Zeyu Shang, Kai Shen, Wei Song, Xu Tan, Heyi Tang, and 1 others. 2025. Kimi-audio technical report. arXiv preprint arXiv:2504.18425. Kewen Ding, Madhu Chetty, Azadeh Noori Hoshyar, Tanusri Bhattacharya, and Britt Klein. 2024. Speech based detection of Alzheimer’s disease: a survey of AI techniques, datasets and challenges. Artificial Intelligence Review, 57(12):325. Zhongren Dong, Zixing Zhang, Weixiang Xu, Jing Han, Jianjun Ou, and Björn W Schuller. 2024. HAF- Former: A hierarchical attention-free framework for Alzheimer’s disease detection from spontaneous speech.In International Conference on Acous- tics, Speech and Signal Processing (ICASSP), pages 11246–11250. IEEE. Alexandre Défossez, Gabriel Synnaeve, and Yossi Adi. 2020. Real Time Speech Enhancement in the Wave- form Domain. In Interspeech 2020, pages 3291– 3295. Alzheimer Europe. 2019. Dementia in Europe Year- book 2019: Estimating the prevalence of dementia in Europe. Alzheimer Europe, 180. Florian Eyben, Klaus R Scherer, Björn W Schuller, Johan Sundberg, Elisabeth André, Carlos Busso, Laurence Y Devillers, Julien Epps, Petri Laukka, Shrikanth S Narayanan, and 1 others. 2015. The Geneva minimalistic acoustic parameter set (GeMAPS) for voice research and affective com- puting. IEEE transactions on affective computing, 7(2):190–202. Florian Eyben, Martin Wöllmer, and Björn Schuller. 2010. Opensmile: the munich versatile and fast open- source audio feature extractor. In Proceedings of the 18th ACM international conference on Multimedia, pages 1459–1462. ACM. Fixie.ai.2025.Ultravoxv0.5llama3.2 1b.https://huggingface.co/fixie-ai/ ultravox-v0_5-llama-3_2-1b.Hugging Face model card. Lior Frankel, Shlomo E Chazan, and Jacob Goldberger. 2025. Automatic detection of domain shifts in speech enhancement systems using confidence-based met- rics. In ICASSP 2025-2025 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE. Sreyan Ghosh, Arushi Goel, Jaehyeon Kim, Sonal Ku- mar, Zhifeng Kong, Sang-gil Lee, Chao-Han Yang, Ramani Duraiswami, Dinesh Manocha, Rafael Valle, and 1 others. 2026. Audio flamingo 3: Advancing au- dio intelligence with fully open large audio language models. Advances in Neural Information Processing Systems, 38:41819–41886. Harold Goodglass and Edith Kaplan. 1983. Boston diag- nostic aphasia examination booklet. Lea & Febiger. Lior Hason and Sri Krishnan. 2022.Spontaneous speech feature analysis for alzheimer’s disease screening using a random forest classifier. Frontiers in Digital Health, 4:901419. Jonathan Heitz, Gerold Schneider, and Nicolas Langer. 2025. Linguistic features extracted by gpt-4 improve alzheimer’s disease detection based on spontaneous speech. In Proceedings of the 31st International Con- ference on Computational Linguistics, pages 1850– 1864. Pamela Herd, Deborah Carr, and Carol Roan. 2014. Cohort profile: Wisconsin longitudinal study (wls). International journal of epidemiology, 43(1):34–41. Kazuma Iwamoto, Tsubasa Ochiai, Marc Delcroix, Rin- taro Ikeshita, Hiroshi Sato, Shoko Araki, and Shigeru Katagiri. 2022. How bad are artifacts?: Analyzing the impact of speech enhancement errors on ASR. In Interspeech 2022, pages 5418–5422. Longbin Jin, Yealim Oh, Hyunseo Kim, Hyuntaek Jung, Hyo Jin Jon, Jung Eun Shin, and Eun Yi Kim. 2023. Consen: Complementary and simultaneous ensemble for alzheimer’s disease detection and mmse score prediction. In International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–2. IEEE. Alyssa M Lanzi, Anna K Saylor, Davida Fromm, Hou- jun Liu, Brian MacWhinney, and Matthew L Cohen. 2023. DementiaBank: Theoretical rationale, proto- col, and illustrative analyses. American Journal of Speech-Language Pathology, 32(2):426–438. You-Jin Li, Rong Chao, Borching Su, and Yu Tsao. 2025. Speech enhancement with map-based training for robust asr. In ICASSP 2025-2025 IEEE Interna- tional Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE. Zeng-Xi Li, Li-Rong Dai, Yan Song, and Ian McLough- lin. 2018. A conditional generative model for speech enhancement. Circuits, Systems, and Signal Process- ing, 37(11):5005–5022. Xiaohui Liang, John A Batsis, Youxiang Zhu, Tiffany M Driesse, Robert M Roth, David Kotz, and Brian MacWhinney. 2022. Evaluating voice-assistant com- mands for dementia detection. Computer Speech & Language, 72:101297. Zih-Jyun Lin, Yi-Ju Chen, Po-Chih Kuo, Likai Huang, Chaur-Jong Hu, and Cheng-Yu Chen. 2024. De- mentia Assessment Using Mandarin Speech with an Attention-Based Speech Recognition Encoder. In International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 12461–12465. IEEE. Xubo Liu, Qiuqiang Kong, Yan Zhao, Haohe Liu, Yi Yuan, Yuzhuo Liu, Rui Xia, Yuxuan Wang, Mark D Plumbley, and Wenwu Wang. 2024. Separate anything you describe. IEEE Transactions on Audio, Speech and Language Processing, 33:458–471. Saturnino Luz, Fasih Haider, Sofia de la Fuente, Davida Fromm, and Brian MacWhinney. 2020. Alzheimer’s Dementia Recognition Through Spontaneous Speech: The ADReSS Challenge. In Interspeech, pages 2172– 2176. Saturnino Luz, Fasih Haider, Davida Fromm, Ioulietta Lazarou, Ioannis Kompatsiaris, and Brian MacWhin- ney. 2023. Multilingual alzheimer’s dementia recog- nition through spontaneous speech: a signal process- ing grand challenge. In International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1–2. IEEE. Saturnino Luz, Fasih Haider, Sofia de la Fuente, Davida Fromm, and Brian MacWhinney. 2021. Detecting cognitive decline using speech only: The adresso challenge. In Interspeech, pages 3780–3784. Tobias Menne, Ralf Schlüter, and Hermann Ney. 2019. Investigation into joint optimization of single chan- nel speech enhancement and acoustic modeling for robust asr. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Process- ing (ICASSP), pages 6660–6664. IEEE. Chanwoo Park, Anna Seo Gyeong Choi, Sunghye Cho, and Chanwoo Kim. 2025. Reasoning-Based Ap- proach with Chain-of-Thought for Alzheimer’s De- tection Using Speech and Large Language Models. In Interspeech 2025, pages 2185–2189. Yu Pu and Wei-Qiang Zhang. 2025. Integrating pause in- formation with word embeddings in language models for alzheimer’s disease detection from spontaneous speech. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Process- ing (ICASSP), pages 1–5. IEEE. Chandan KA Reddy, Vishak Gopal, and Ross Cutler. 2021. Dnsmos: A non-intrusive perceptual objective speech quality metric to evaluate noise suppressors. In ICASSP 2021-2021 IEEE International Confer- ence on Acoustics, Speech and Signal Processing (ICASSP), pages 6493–6497. IEEE. Robin Scheibler, Yusuke Fujita, Yuma Shirahata, and Tatsuya Komatsu. 2024.Universal Score-based Speech Enhancement with High Content Preserva- tion. In Interspeech 2024, pages 1165–1169. Philip Scheltens, Bart De Strooper, Miia Kivipelto, Henne Holstege, Gael Chételat, Charlotte E Teu- nissen, Jeffrey Cummings, and Wiesje M van der Flier. 2021.Alzheimer’s disease.The Lancet, 397(10284):1577–1590. Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30. Yangyang Xia, Sebastian Braun, Chandan KA Reddy, Harishchandra Dubey, Ross Cutler, and Ivan Ta- shev. 2020. Weighted speech distortion losses for neural-network-based real-time speech enhancement. In ICASSP 2020-2020 IEEE International Confer- ence on Acoustics, Speech and Signal Processing (ICASSP), pages 871–875. IEEE. Jin Xu, Zhifang Guo, Hangrui Hu, Yunfei Chu, Xiong Wang, Jinzheng He, Yuxuan Wang, Xian Shi, Ting He, Xinfa Zhu, and 1 others. 2025. Qwen3-omni technical report. arXiv preprint arXiv:2509.17765. Qin Yang, Xin Li, Xinyun Ding, Feiyang Xu, and Zhen- hua Ling. 2022. Deep learning-based speech analysis for Alzheimer’s disease detection: a literature review. Alzheimer’s Research & Therapy, 14(1):186. Takuya Yoshioka, Nobutaka Ito, Marc Delcroix, At- sunori Ogawa, Keisuke Kinoshita, Masakiyo Fuji- moto, Chengzhu Yu, Wojciech J Fabian, Miquel Espi, Takuya Higuchi, and 1 others. 2015. The ntt chime-3 system: Advances in speech enhancement and recog- nition for mobile multi-microphone devices. In 2015 IEEE Workshop on Automatic Speech Recognition and Understanding (ASRU), pages 436–443. IEEE. Qishan Zhang, Shuangbing Wen, and Tao Hu. 2024. Audio deepfake detection with self-supervised xls-r and sls classifier. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6765– 6773. Shengkui Zhao and Bin Ma. 2023. Mossformer: Push- ing the performance limit of monaural speech sep- aration using gated single-head transformer with convolution-augmented joint self-attentions.In ICASSP 2023-2023 IEEE International Confer- ence on Acoustics, Speech and Signal Processing (ICASSP), pages 1–5. IEEE. Shengkui Zhao, Bin Ma, Karn N Watcharasupat, and Woon-Seng Gan. 2022. Frcrn: Boosting feature rep- resentation using frequency recurrence for monaural speech enhancement. In ICASSP 2022-2022 IEEE in- ternational conference on acoustics, speech and sig- nal processing (ICASSP), pages 9281–9285. IEEE. A Dataset Information A.1 Dataset Details Non-Denoising Denoised Original SizePitt-originPitt Reduced SizeADReSS-M ADReSSo, ADReSS Table 5:Relationship of Pitt-origin and its de- rived datasets. (Pitt-origin and Pitt have the same size. ADReSS-M and ADReSSo have the same size. ADReSS is smaller than ADReSS-M and ADReSSo.) 50100150200250 Duration (S) 0 10 20 30 40 50 60 70 80 Sample Number Pitt-origin/Pitt 50100150200250 Duration (S) 0 3 6 9 12 15 18 21 Sample Number ADReSS 50100150200250 Duration (S) 0 4 8 12 16 20 24 28 32 Sample Number ADReSSo 50100150200250 Duration (S) 0 4 8 12 16 20 24 28 32 Sample Number ADReSS-M Figure 5: Duration distribution of Pitt-origin and its derived datasets. DatasetMin(s) Max(s) Mean(s) Total(min) Pitt-origin / Pitt17.89268.4969.97643.72 ADReSS26.06268.4975.30195.78 ADReSSo22.35268.4976.89303.72 ADReSS-M22.35268.4976.89303.72 Table 6: Duration distribution of Pitt-origin and its de- rived datasets. Age ADControl MaleFemaleMaleFemale < 500106 [50, 55)5189 [55, 60)10111728 [60, 65)18181325 [65, 70)26332440 [70, 75)23432033 [75, 80)234569 [80, 90]133714 unknown2000 Total12018989154 Table 7: Pitt-origin / Pitt dataset age and gender meta- data. Age ADControl MaleFemaleMaleFemale [50, 55)2020 [55, 60)7676 [60, 65)4949 [65, 70)914914 [70, 75)911911 [75, 80)4343 Total35433543 Table 8: ADReSS dataset age and gender metadata. Age ADControl MaleFemaleMaleFemale [50, 55)1011 [55, 60)76616 [60, 65)78713 [65, 70)7221123 [70, 75)9211018 [75, 80]122254 Total43794075 Table 9: ADReSSo / ADReSS-M dataset age and gender metadata. DatasetOVRL↑SIG↑BAK↑ Pitt-origin2.30732.98192.7290 Pitt-origin-Denoiser2.96873.29464.2052 Pitt-origin-FRCRN2.87043.32603.9332 Pitt-origin-MossFormer3.02323.41364.1258 Pitt-origin-Resemble2.88563.31513.8913 Pitt-origin-MAP-SEMamba2.87993.34343.8376 Table 10: DNSMOS scores of Pitt-origin under different speech enhancement methods. (SIG: speech quality; BAK: background noise quality; OVRL: overall speech quality.) B Prompt B.1 Audio Prompt You are a clinical speech-language pathologist specialized in detecting Alzheimer’s disease and dementia from spontaneous speech. You analyze speech patterns including: word-finding difficulties, semantic paraphasias, empty speech, reduced syntactic complexity, repetitions, incomplete utterances, and pragmatic impairments. Based on the audio, classify the speaker. Listen to this speech sample carefully. Based on the speech characteristics, is this speaker showing signs of dementia or is this a healthy control? Answer with exactly one word: ’Dementia’ or ’Control’. B.2 Audio + Transcript Prompt You are a clinical speech-language pathologist specialized in detecting Alzheimer’s disease and dementia from spontaneous speech and transcription. You analyze speech patterns and transcribed text including: word-finding difficulties, semantic paraphasias, empty speech, reduced syntactic complexity, repetitions, incomplete utterances, and pragmatic impairments. Based on the audio and transcription, classify the speaker. Listen to this speech sample and read the transcription carefully. Based on both the speech characteristics and the transcription content, is this speaker showing signs of dementia or is this a healthy control? Answer with exactly one word: ’Dementia’ or ’Control’. C Large Audio-Language Models C.1 LALMs Audio-only Zero-Shot Setting Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Pitt-origin Pitt ADReSS ADReSSo ADReSS-M Lu Test Set 0.6666 * 0.6726 * 0.6605 * 0.5924 * 0.7300 * 0.4548 * 0.5629 * 0.6834 * 0.4425 * 0.6348 * 0.7273 * 0.5424 * 0.7000 * 0.6899 * 0.7102 * 0.47690.40620.5476 Kimi-Audio Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Pitt-origin Pitt ADReSS ADReSSo ADReSS-M Lu Test Set 0.48780.72900.2465 0.43270.71890.1464 0.40230.66660.1379 0.39930.68570.1130 0.43040.68030.1804 0.39080.67890.1026 Qwen3 Omni Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Pitt-origin Pitt ADReSS ADReSSo ADReSS-M Lu Test Set 0.30570.00000.6113 0.30520.00000.6103 0.33330.00000.6667 0.32670.00000.6534 0.32670.00000.6534 0.32730.00000.6545 Qwen2-Audio Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Pitt-origin Pitt ADReSS ADReSSo ADReSS-M Lu Test Set 0.35890.71780.0000 0.35930.71860.0000 0.33330.66670.0000 0.33980.67970.0000 0.33980.67970.0000 0.33930.67860.0000 Audio Flamingo 3 Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Pitt-origin Pitt ADReSS ADReSSo ADReSS-M Lu Test Set 0.44410.26420.6241 0.44840.27520.6216 0.47560.29410.6571 0.46530.28930.6413 0.50310.33960.6666 0.5679 * 0.4255 * 0.7103 * ultravox-v0_5-llama-3_2-1b 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 Large Audio-Language Models (Audio-only 0-Shot) Figure 6: LALMs performance in the audio-only zero-shot setting. (AD + / Control + : F1 score with AD/Control as positive. Asterisks ∗ : performance significantly different from the binomial majority-class baseline, p < 0.05.) C.2 Kimi-Audio Audio-only Setting Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Pitt-origin Pitt ADReSS ADReSSo ADReSS-M Test Set 0.6666 * 0.6726 * 0.6605 * 0.5924 * 0.7300 * 0.4548 * 0.5629 * 0.6834 * 0.4425 * 0.6348 * 0.7273 * 0.5424 * 0.7000 * 0.6899 * 0.7102 * 0-Shot Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Pitt-origin Pitt ADReSS ADReSSo ADReSS-M Test Set 0.6251 * 0.7198 * 0.5305 * 0.5878 * 0.7409 * 0.4348 * 0.5879 * 0.6842 * 0.4915 * 0.6503 * 0.7291 * 0.5714 * 0.6441 * 0.5941 * 0.6941 * 2-Shot 0.0 0.2 0.4 0.6 0.8 1.0 Kimi-Audio (Audio-only) Figure 7: Kimi-Audio in the audio-only setting. (AD + / Control + : F1 score with AD/Control as positive. Asterisks ∗ : performance significantly different from the binomial majority-class baseline, p < 0.05.) C.3 Kimi-Audio Audio + Transcript Setting Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Pitt-origin Pitt Test Set 0.6514 * 0.6350 * 0.6678 * 0.47560.71640.2348 0-Shot Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Pitt-origin Pitt Test Set 0.6651 * 0.7313 * 0.5988 * 0.5498 * 0.7194 * 0.3801 * 2-Shot 0.0 0.2 0.4 0.6 0.8 1.0 Kimi-Audio (Audio+Transcript) Figure 8: Kimi-Audio in the audio + transcript setting. (AD + / Control + : F1 score with AD/Control as positive. Asterisks ∗ : performance significantly different from the binomial majority-class baseline, p < 0.05.) D Model Architectures Input Audio OpenSMILE Toolkit eGeMAPS Features Attention Pooling Prediction Linear Layer Feature Projection eGeMAPS-based Model Frozen XLS-R Input Audio Embedding CNN Layer Attention Pooling Prediction Linear Layer XLSR-based Model Attention Pooling Prediction Linear Layer CNN Layers Input Audio Frozen XLS-R Transformer Layers ... Temporal Feature Embedding Average Pooling CNN Layer SLS-based Model Figure 9: Model Architectures. SLS-based Model. The model takes as input the frame-level representations from all Transformer layers of the parameter-frozen pretrained XLS-R model (Conneau et al., 2021), and obtains a weighted fused representation through the Sensitive Layer Selection (SLS) (Zhang et al., 2024) mechanism. Specifically, the model first applies mean pooling with the padding mask (Vaswani et al., 2017) along the temporal dimension to the representations of each Transformer layer of XLSR. Subsequently, the weights of different layers are obtained through a linear layer and a sigmoid function. Finally, the model performs a weighted sum of the representations from different layers using the learned layer weights to obtain the weighted fused representation. This representation then sequentially passes through batch normalization, temporal average pooling, a one-dimensional convolutional layer, and attention pooling. Finally, an AD prediction is output through a linear classification layer. The SLS-based model contains 169K trainable parameters. All models are trained on NVIDIA RTX 5090 GPUs. Additional implementation details are provided in our released code. XLSR-based Model. The audio input is fed into a parameter-frozen pretrained XLSR model (Conneau et al., 2021) to extract frame-level speech embeddings. The representation then sequentially passes through batch normalization and is aggregated by an attention pooling layer to obtain a temporal feature representation. A padding mask (Vaswani et al., 2017) is also applied to ensure that padded time steps do not contribute to the pooled representation. Finally, a linear classification layer outputs the AD prediction. The XLSR-based model contains 168K trainable parameters. All models are trained on NVIDIA RTX 5090 GPUs. Additional implementation details are provided in our released code. eGeMAPS-based Model. The model takes 25-dimensional eGeMAPS (Eyben et al., 2015) acoustic features extracted using the OpenSMILE toolkit (Eyben et al., 2010) as input. Each input utterance is first segmented into 10 equal segments, from which 25-dimensional eGeMAPS (Eyben et al., 2015) acoustic features are extracted using the OpenSMILE toolkit (Eyben et al., 2010). Two linear layers are then applied to remap the feature dimensions, followed by an attention pooling layer to aggregate segment-level representations. Finally, a linear classification layer outputs the AD prediction. The eGeMAPS-based model contains 6.2K trainable parameters. All models are trained on NVIDIA RTX 5090 GPUs. Additional implementation details are provided in our released code. E SLS-based Model E.1 Generalization Test On Raw Data Pitt-originPittADReSSADReSSoADReSS-M Dataset 0.00 0.20 0.40 0.60 0.80 Macro-F1 0.7910 * 0.7449 * 0.8120 * 0.6300 * 0.8745 * 0.4166 0.8964 * 0.6754 * 0.7713 * 0.5470 Same domain Test Test on Lu SLS-based Model Figure 10: SLS-based model generalization test on raw data. (Same Domain Test: evaluation on the test split of the same dataset used for training; Test on Lu: evaluation on the Lu dataset. Asterisks ∗ : performance significantly different from the binomial majority-class baseline, p < 0.05.) E.2 Generalization Test On Speech-enhanced Data Same Domain Test LuLu -Denoiser Lu -FRCRN Lu -MossFormer Lu -Resemble Lu -MAP-SEMamba Test Set Pitt-origin Pitt ADReSS ADReSSo ADReSS-M Training Set 0.7910 * 0.7449 * 0.6552 * 0.6840 * 0.6620 * 0.7004 * 0.6811 * 0.8120 * 0.6300 * 0.5791 * 0.6103 * 0.6300 * 0.6103 * 0.6247 * 0.8745 * 0.4166 0.6621 * 0.6718 * 0.6296 * 0.6476 * 0.6707 * 0.8964 * 0.6754 * 0.7150 * 0.6822 * 0.7280 * 0.7432 * 0.7306 * 0.7713 * 0.54700.53750.50510.48250.4349 0.6284 * 0.5 0.6 0.7 0.8 Macro-F1 SLS-based Model Figure 11: SLS-based model generalization test on speech-enhanced data. (Same Domain Test: evaluation on the test split of the same dataset used for training. Asterisks ∗ : performance significantly different from the binomial majority-class baseline, p < 0.05.) E.3 Performance Comparison with Matched Speech Enhancement Methods DenoiserFRCRNMossFormerResembleMAP-SEMamba Speech Enhancement Method 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Macro-F1 Baseline=0.7449 * 0.5406 0.6793 * 0.6171 * 0.6917 * 0.6616 * 0.6433 * 0.7072 * 0.6476 * 0.6913 * 0.7364 * 0.6552 * 0.6840 * 0.6620 * 0.7004 * 0.6811 * Mismatched: Enhanced Pitt-origin Lu Matched: Enhanced Pitt-origin Enhanced Lu Raw-trained: Pitt-origin Enhanced Lu Baseline: Pitt-origin Lu = 0.7449 * SLS-based Model Figure 12: SLS-based model performance comparison with matched and mismatched speech enhancement methods. (Mismatched: enhanced training set with raw test set; Matched: the same enhancement for training and test set; Raw-trained: raw training set with enhanced test set; Baseline: raw training and raw test set. Asterisks ∗ : performance significantly different from the binomial majority-class baseline, p < 0.05.) E.4 SLS-based Model All Results Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Same Domain Lu Lu-Denoiser Lu-FRCRN Lu-MossFormer Lu-Resemble Lu-MAP-SEMamba Test Set 0.7910 * 0.8102 * 0.7718 * 0.7449 * 0.8000 * 0.6897 * 0.6552 * 0.7391 * 0.5714 * 0.6840 * 0.7609 * 0.6071 * 0.6620 * 0.7579 * 0.5661 * 0.7004 * 0.7693 * 0.6316 * 0.6811 * 0.7668 * 0.5954 * Training Set: Pitt-origin Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Same Domain Lu Lu-Denoiser Lu-FRCRN Lu-MossFormer Lu-Resemble Lu-MAP-SEMamba Test Set 0.8120 * 0.8516 * 0.7724 * 0.6300 * 0.7600 * 0.5000 * 0.5791 * 0.7327 * 0.4256 * 0.6103 * 0.7525 * 0.4681 * 0.6300 * 0.7600 * 0.5000 * 0.6103 * 0.7525 * 0.4681 * 0.6247 * 0.7499 * 0.4995 * Training Set: Pitt Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Same Domain Lu Lu-Denoiser Lu-FRCRN Lu-MossFormer Lu-Resemble Lu-MAP-SEMamba Test Set 0.8745 * 0.8667 * 0.8824 * 0.41660.23530.5980 0.6621 * 0.6667 * 0.6576 * 0.6718 * 0.7073 * 0.6364 * 0.6296 * 0.6747 * 0.5846 * 0.6476 * 0.6667 * 0.6286 * 0.6707 * 0.6723 * 0.6690 * Training Set: ADReSS Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Same Domain Lu Lu-Denoiser Lu-FRCRN Lu-MossFormer Lu-Resemble Lu-MAP-SEMamba Test Set 0.8964 * 0.8931 * 0.8997 * 0.6754 * 0.6842 * 0.6667 * 0.7150 * 0.7342 * 0.6957 * 0.6822 * 0.7294 * 0.6350 * 0.7280 * 0.7500 * 0.7059 * 0.7432 * 0.7466 * 0.7397 * 0.7306 * 0.7856 * 0.6756 * Training Set: ADReSSo Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Same Domain Lu Lu-Denoiser Lu-FRCRN Lu-MossFormer Lu-Resemble Lu-MAP-SEMamba Test Set 0.7713 * 0.7644 * 0.7782 * 0.54700.40000.6939 0.53750.37490.7000 0.50510.31110.6990 0.48250.27270.6923 0.43490.19050.6793 0.6284 * 0.5372 * 0.7196 * Training Set: ADReSS-M Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Same Domain Lu Lu-Denoiser Lu-FRCRN Lu-MossFormer Lu-Resemble Lu-MAP-SEMamba Test Set 0.7884 * 0.8194 * 0.7572 * 0.54060.54050.5406 0.6433 * 0.7253 * 0.5614 * 0.6148 * 0.7033 * 0.5263 * 0.6326 * 0.7369 * 0.5283 * 0.6392 * 0.6977 * 0.5807 * 0.7306 * 0.7856 * 0.6756 * Training Set: Pitt-origin-Denoiser Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Same Domain Lu Lu-Denoiser Lu-FRCRN Lu-MossFormer Lu-Resemble Lu-MAP-SEMamba Test Set 0.7602 * 0.7974 * 0.7230 * 0.6793 * 0.7660 * 0.5926 * 0.6260 * 0.7423 * 0.5098 * 0.7072 * 0.7586 * 0.6558 * 0.6881 * 0.7555 * 0.6207 * 0.7502 * 0.7907 * 0.7097 * 0.6541 * 0.7697 * 0.5384 * Training Set: Pitt-origin-FRCRN_SE Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Same Domain Lu Lu-Denoiser Lu-FRCRN Lu-MossFormer Lu-Resemble Lu-MAP-SEMamba Test Set 0.8185 * 0.8336 * 0.8035 * 0.6171 * 0.5758 * 0.6585 * 0.6571 * 0.6988 * 0.6154 * 0.58360.65110.5161 0.6476 * 0.6667 * 0.6286 * 0.6445 * 0.6829 * 0.6061 * 0.7598 * 0.8027 * 0.7169 * Training Set: Pitt-origin-MossFormer Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Same Domain Lu Lu-Denoiser Lu-FRCRN Lu-MossFormer Lu-Resemble Lu-MAP-SEMamba Test Set 0.7859 * 0.8249 * 0.7470 * 0.6917 * 0.7500 * 0.6333 * 0.6443 * 0.7500 * 0.5385 * 0.5684 * 0.7200 * 0.4167 * 0.6376 * 0.7551 * 0.5200 * 0.6913 * 0.7789 * 0.6037 * 0.6247 * 0.7499 * 0.4995 * Training Set: Pitt-origin-Resemble Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Same Domain Lu Lu-Denoiser Lu-FRCRN Lu-MossFormer Lu-Resemble Lu-MAP-SEMamba Test Set 0.8374 * 0.8450 * 0.8297 * 0.6616 * 0.6753 * 0.6479 * 0.6514 * 0.7126 * 0.5902 * 0.6759 * 0.7416 * 0.6102 * 0.7072 * 0.7586 * 0.6558 * 0.7280 * 0.7500 * 0.7059 * 0.7364 * 0.7625 * 0.7103 * Training Set: Pitt-origin-MAP-SEMamba 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 SLS-based Model Figure 13: SLS-based model all results. (AD + / Control + : F1 score with AD/Control as positive. Asterisks ∗ : performance significantly different from the binomial majority-class baseline, p < 0.05.) F XLSR-based Model F.1 Generalization Test On Raw Data Pitt-originPittADReSSADReSSoADReSS-M Dataset 0.00 0.10 0.20 0.30 0.40 0.50 0.60 0.70 0.80 Macro-F1 0.6163 0.6735 * 0.7238 * 0.5624 0.8436 * 0.3608 0.7502 * 0.3280 0.6851 * 0.4563 Same domain Test Test on Lu XLSR-based Model Figure 14: XLSR-based model generalization test on raw data. (Same Domain Test: evaluation on the test split of the same dataset used for training; Test on Lu: evaluation on the Lu dataset. Asterisks ∗ : performance significantly different from the binomial majority-class baseline, p < 0.05.) F.2 Generalization Test On Speech-enhanced Data Same Domain Test LuLu -Denoiser Lu -FRCRN_SE Lu -MossFormer Lu -Resemble Lu -MAP-SEMamba Test Set Pitt-origin Pitt ADReSS ADReSSo ADReSS-M Training Set 0.6163 0.6735 * 0.6080 0.6463 * 0.6334 * 0.5871 0.6887 * 0.7238 * 0.56240.4855 0.6094 * 0.45020.4764 0.6079 * 0.8436 * 0.36080.52350.50040.46820.55010.4422 0.7502 * 0.32800.5515 0.6205 * 0.6318 * 0.6606 * 0.5887 0.6851 * 0.45630.46470.36080.32800.36830.4337 0.4 0.5 0.6 0.7 0.8 Macro-F1 XLSR-based Model Figure 15: XLSR-based model generalization test on speech-enhanced data. (Same Domain Test: evaluation on the test split of the same dataset used for training. Asterisks ∗ : performance significantly different from the binomial majority-class baseline, p < 0.05.) F.3 Performance Comparison with Matched Speech Enhancement Methods DenoiserFRCRNMossFormerResembleMAP-SEMamba Speech Enhancement Method 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Macro-F1 Baseline=0.6735 * 0.3658 0.6484 * 0.4530 0.5163 0.4394 0.6022 0.6046 0.5918 * 0.6793 * 0.4473 0.6080 0.6463 * 0.6334 * 0.5871 0.6887 * Mismatched: Enhanced Pitt-origin Lu Matched: Enhanced Pitt-origin Enhanced Lu Raw-trained: Pitt-origin Enhanced Lu Baseline: Pitt-origin Lu = 0.6735 * XLSR-based Model Figure 16: XLSR-based model performance comparison with matched and mismatched speech enhancement methods. (Mismatched: enhanced training set with raw test set; Matched: the same enhancement for training and test set; Raw-trained: raw training set with enhanced test set; Baseline: raw training and raw test set. Asterisks ∗ : performance significantly different from the binomial majority-class baseline, p < 0.05.) F.4 XLSR-based Model All Results Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Same Domain Lu Lu-Denoiser Lu-FRCRN_SE Lu-MossFormer Lu-Resemble Lu-MAP-SEMamba Test Set 0.61630.66180.5708 0.6735 * 0.7000 * 0.6470 * 0.60800.60270.6133 0.6463 * 0.6750 * 0.6176 * 0.6334 * 0.6582 * 0.6087 * 0.58710.53130.6429 0.6887 * 0.7848 * 0.5926 * Training Set: Pitt-origin Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Same Domain Lu Lu-Denoiser Lu-FRCRN_SE Lu-MossFormer Lu-Resemble Lu-MAP-SEMamba Test Set 0.7238 * 0.7894 * 0.6583 * 0.56240.60980.5151 0.48550.67320.2978 0.6094 * 0.7097 * 0.5091 * 0.45020.67310.2273 0.47640.63920.3137 0.6079 * 0.7178 * 0.4980 * Training Set: Pitt Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Same Domain Lu Lu-Denoiser Lu-FRCRN_SE Lu-MossFormer Lu-Resemble Lu-MAP-SEMamba Test Set 0.8436 * 0.8387 * 0.8485 * 0.36080.09300.6286 0.52350.61360.4333 0.50040.41930.5814 0.46820.41790.5185 0.55010.50750.5926 0.44220.22280.6616 Training Set: ADReSS Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Same Domain Lu Lu-Denoiser Lu-FRCRN_SE Lu-MossFormer Lu-Resemble Lu-MAP-SEMamba Test Set 0.7502 * 0.7376 * 0.7629 * 0.32800.04650.6095 0.55150.46670.6364 0.6205 * 0.6000 * 0.6410 * 0.6318 * 0.5970 * 0.6667 * 0.6606 * 0.6376 * 0.6835 * 0.58870.51970.6578 Training Set: ADReSSo Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Same Domain Lu Lu-Denoiser Lu-FRCRN_SE Lu-MossFormer Lu-Resemble Lu-MAP-SEMamba Test Set 0.6851 * 0.6497 * 0.7206 * 0.45630.28000.6326 0.46470.32720.6022 0.36080.09300.6286 0.32800.04650.6095 0.36830.09520.6415 0.43370.21820.6492 Training Set: ADReSS-M Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Same Domain Lu Lu-Denoiser Lu-FRCRN_SE Lu-MossFormer Lu-Resemble Lu-MAP-SEMamba Test Set 0.7293 * 0.7371 * 0.7215 * 0.36580.16010.5714 0.60220.65060.5538 0.53200.59520.4688 0.59130.67420.5085 0.6864 * 0.7161 * 0.6567 * 0.43040.54220.3186 Training Set: Pitt-origin-Denoiser Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Same Domain Lu Lu-Denoiser Lu-FRCRN_SE Lu-MossFormer Lu-Resemble Lu-MAP-SEMamba Test Set 0.7024 * 0.7083 * 0.6966 * 0.6484 * 0.6579 * 0.6389 * 0.5772 * 0.7143 * 0.4401 * 0.60460.64200.5672 0.6326 * 0.7369 * 0.5283 * 0.6421 * 0.6905 * 0.5938 * 0.6246 * 0.7250 * 0.5242 * Training Set: Pitt-origin-FRCRN Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Same Domain Lu Lu-Denoiser Lu-FRCRN_SE Lu-MossFormer Lu-Resemble Lu-MAP-SEMamba Test Set 0.7295 * 0.7324 * 0.7266 * 0.45300.39390.5122 0.5472 * 0.7308 * 0.3636 * 0.59130.67420.5085 0.5918 * 0.7021 * 0.4815 * 0.5978 * 0.6957 * 0.5000 * 0.56330.61020.5163 Training Set: Pitt-origin-MossFormer Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Same Domain Lu Lu-Denoiser Lu-FRCRN_SE Lu-MossFormer Lu-Resemble Lu-MAP-SEMamba Test Set 0.7356 * 0.7647 * 0.7066 * 0.51630.44440.5882 0.51670.69310.3404 0.51800.62220.4137 0.56260.68080.4444 0.6793 * 0.7356 * 0.6230 * 0.47360.48040.4669 Training Set: Pitt-origin-Resemble Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Same Domain Lu Lu-Denoiser Lu-FRCRN_SE Lu-MossFormer Lu-Resemble Lu-MAP-SEMamba Test Set 0.7077 * 0.7424 * 0.6729 * 0.43940.33340.5455 0.48260.69230.2728 0.58040.68820.4727 0.48650.65310.3200 0.56260.68080.4444 0.44730.46850.4261 Training Set: Pitt-origin-MAP-SEMamba 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 XLSR-based Model Figure 17: XLSR-based model all results. (AD + / Control + : F1 score with AD/Control as positive. Asterisks ∗ : performance significantly different from the binomial majority-class baseline, p < 0.05.) G eGeMAPS-based Model G.1 Generalization Test On Raw Data Pitt-originPittADReSSADReSSoADReSS-M Dataset 0.00 0.10 0.20 0.30 0.40 0.50 0.60 Macro-F1 0.5482 0.5978 * 0.5997 0.5351 0.6190 0.4977 0.6505 * 0.4394 0.6057 0.5677 Same domain Test Test on Lu eGeMAPS-based Model Figure 18: eGeMAPS-based model generalization test on raw data. (Same Domain Test: evaluation on the test split of the same dataset used for training; Test on Lu: evaluation on the Lu dataset. Asterisks ∗ : performance significantly different from the binomial majority-class baseline, p < 0.05.) G.2 Generalization Test On Speech-enhanced Data Same Domain Test LuLu -Denoiser Lu -FRCRN Lu -MossFormer Lu -Resemble Lu -MAP-SEMamba Test Set Pitt-origin Pitt ADReSS ADReSSo ADReSS-M Training Set 0.5482 0.5978 * 0.51200.48040.49250.5520 0.6146 * 0.59970.53510.40710.38930.46740.42220.5118 0.61900.49770.53500.42830.41690.42570.4957 0.6505 * 0.43940.52620.49770.53510.42570.4576 0.60570.56770.45790.49920.49550.47060.6022 0.40 0.45 0.50 0.55 0.60 0.65 Macro-F1 eGeMAPS-based Model Figure 19: eGeMAPS-based model generalization test on speech-enhanced data. (Same Domain Test: evaluation on the test split of the same dataset used for training. Asterisks ∗ : performance significantly different from the binomial majority-class baseline, p < 0.05.) G.3 Performance Comparison with Matched Speech Enhancement Methods DenoiserFRCRNMossFormerResembleMAP-SEMamba Speech Enhancement Method 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Macro-F1 Baseline=0.5978 * 0.3921 0.5672 0.3766 0.4856 0.4007 0.5228 0.4283 0.5398 0.5150 0.5078 0.5120 0.4804 0.4925 0.5520 0.6146 * Mismatched: Enhanced Pitt-origin Lu Matched: Enhanced Pitt-origin Enhanced Lu Raw-trained: Pitt-origin Enhanced Lu Baseline: Pitt-origin Lu = 0.5978 * eGeMAPS-based Model Figure 20: eGeMAPS-based model performance comparison with matched and mismatched speech enhancement methods. (Mismatched: enhanced training set with raw test set; Matched: the same enhancement for training and test set; Raw-trained: raw training set with enhanced test set; Baseline: raw training and raw test set. Asterisks ∗ : performance significantly different from the binomial majority-class baseline, p < 0.05.) G.4 eGeMAPS-based Model All Results Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Same Domain Lu Lu-Denoiser Lu-FRCRN Lu-MossFormer Lu-Resemble Lu-MAP-SEMamba Test Set 0.54820.71980.3766 0.5978 * 0.6957 * 0.5000 * 0.51200.59770.4262 0.48040.53660.4242 0.49250.55420.4308 0.55200.58230.5217 0.6146 * 0.6666 * 0.5625 * Training Set: Pitt-origin Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Same Domain Lu Lu-Denoiser Lu-FRCRN Lu-MossFormer Lu-Resemble Lu-MAP-SEMamba Test Set 0.59970.67290.5265 0.53510.58540.4848 0.40710.23080.5833 0.38930.29030.4883 0.46740.36670.5682 0.42220.30500.5393 0.51180.49930.5243 Training Set: Pitt Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Same Domain Lu Lu-Denoiser Lu-FRCRN Lu-MossFormer Lu-Resemble Lu-MAP-SEMamba Test Set 0.61900.66670.5714 0.49770.46380.5316 0.53500.44070.6292 0.42830.32790.5288 0.41690.25920.5745 0.42570.36370.4878 0.49570.46270.5287 Training Set: ADReSS Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Same Domain Lu Lu-Denoiser Lu-FRCRN Lu-MossFormer Lu-Resemble Lu-MAP-SEMamba Test Set 0.6505 * 0.7205 * 0.5804 * 0.43940.54550.3334 0.52620.54540.5070 0.49770.53160.4637 0.53510.58540.4848 0.42570.36370.4878 0.45760.56450.3508 Training Set: ADReSSo Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Same Domain Lu Lu-Denoiser Lu-FRCRN Lu-MossFormer Lu-Resemble Lu-MAP-SEMamba Test Set 0.60570.72590.4855 0.56770.64370.4918 0.45790.48720.4286 0.49920.51950.4789 0.49550.54320.4478 0.47060.43480.5064 0.60220.65060.5538 Training Set: ADReSS-M Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Same Domain Lu Lu-Denoiser Lu-FRCRN Lu-MossFormer Lu-Resemble Lu-MAP-SEMamba Test Set 0.61900.65050.5875 0.39210.17030.6139 0.52280.47770.5679 0.36910.13040.6078 0.44500.25000.6400 0.42540.28570.5652 0.50850.35430.6627 Training Set: Pitt-origin-Denoiser Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Same Domain Lu Lu-Denoiser Lu-FRCRN Lu-MossFormer Lu-Resemble Lu-MAP-SEMamba Test Set 0.61780.65630.5794 0.56720.55550.5789 0.44940.37500.5238 0.42830.32790.5288 0.50450.57140.4375 0.44500.25000.6400 0.6171 * 0.6084 * 0.6259 * Training Set: Pitt-origin-FRCRN_SE Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Same Domain Lu Lu-Denoiser Lu-FRCRN Lu-MossFormer Lu-Resemble Lu-MAP-SEMamba Test Set 0.52720.70670.3476 0.37660.41030.3428 0.53750.70000.3750 0.53980.62070.4590 0.53980.62070.4590 0.46130.34490.5778 0.58980.63410.5455 Training Set: Pitt-origin-MossFormer Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Same Domain Lu Lu-Denoiser Lu-FRCRN Lu-MossFormer Lu-Resemble Lu-MAP-SEMamba Test Set 0.56130.74770.3749 0.48560.63160.3396 0.6235 * 0.6897 * 0.5574 * 0.53490.62920.4407 0.57920.60760.5507 0.51500.65260.3774 0.6057 * 0.6901 * 0.5213 * Training Set: Pitt-origin-Resemble Macro-F1F1 Score (Dementia + ) F1 Score (Control + ) Same Domain Lu Lu-Denoiser Lu-FRCRN Lu-MossFormer Lu-Resemble Lu-MAP-SEMamba Test Set 0.7002 * 0.7277 * 0.6727 * 0.40070.62740.1739 0.40960.67290.1463 0.52710.68690.3673 0.5849 * 0.7083 * 0.4615 * 0.6076 * 0.6818 * 0.5333 * 0.50780.69020.3254 Training Set: Pitt-origin-MAP-SEMamba 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0 eGeMAPS-based Model Figure 21: eGeMAPS-based model all results. (AD + / Control + : F1 score with AD/Control as positive. Asterisks ∗ : performance significantly different from the binomial majority-class baseline, p < 0.05.)