Update on the current state of speech and language testing in Alzheimer's disease

Abstract

Alzheimer's disease (AD) causes alterations in speech and language across multiple domains, including semantic content, pragmatics, prosody, morphosyntax, phonology and timing/fluency, with some of these changes detectable early in the disease course. Definitive biomarker assessments of AD are often invasive or resource-intensive; there is a need for scalable measures in screening and monitoring AD. Advances in natural language processing and automated transcription can quantify speech and language features from structured tasks and natural conversations. This review summarizes domain-specific changes across the AD continuum, evaluates evidence linking these features to biomarkers of AD pathology (amyloid and tau), and discusses opportunities and limitations for future work. We conclude by identifying priorities for the field, namely, domain-aware task design, demographic and language-sensitive norms and standardized, biomarker-validated datasets.

Keywords

acoustics Alzheimer's disease clinical marker language natural language processing screening speech

Overview

Alzheimer's disease (AD) currently affects 32 million people worldwide, and its prevalence is projected to increase markedly as populations age.¹ According to the Alzheimer's Association international consensus, AD is defined by the presence of extracellular amyloid-β (Aβ) plaques and neurofibrillary tau tangles.^2,3 It is categorized into three clinicopathological stages: a pre-symptomatic stage where individuals are biomarker-positive but cognitively unimpaired, a mild cognitive impairment (MCI) stage characterized by subtle but measurable memory and executive deficits, and a dementia stage where there is progressive decline in memory, judgement, language and independent functioning (Figure 1).³ Diagnosis of AD requires a combination of case history, neuropsychological assessment, brain imaging (positron emission tomography (PET) or magnetic resonance imaging (MRI)) and cerebrospinal fluid (CSF) analysis by lumbar puncture.^4–6 Although these methods confirm biomarker presence, they remain invasive, costly or logistically inaccessible for large-scale or repeated testing.^7,8 Furthermore, standard neuropsychological assessments can be influenced by education levels, culture and rater experience.^9,10 This demands scalable screening methods that are economical and sensitive to the earliest changes.

Figure 1.

Chronological biomarkers of AD. Amyloid plaque biomarkers are dynamic early in the disease, appearing before clinical symptoms and reaching a plateau by the time these symptoms manifest.³

Cognitive impairments resulting from AD often lead to breakdowns in conversation and general communication exchanges. Dysnomia, verbal paraphasia, word finding, and verbal fluency deficits all manifest in AD and deteriorate with disease progression.¹¹ In some cases, these features can present very early in the disease course. Subtle changes in understanding figurative language (expressions whose meaning differs from the words’ literal meanings, such as proverbs, metaphors, exaggeration, and idioms) can also be present early in AD.¹² Such speech and language changes not only alter daily communication but also reflect underlying cognitive decline. Subtle language deficits, such as understanding figurative expressions, may be noticeable before global cognition is severely affected.¹² These changes may serve as viable markers of quality of life and disease onset or progression. In this context, digitally derived speech and language features may not only enable screening and longitudinal monitoring but can also flag individuals whose speech patterns are suspicious for underlying Alzheimer's pathology. As such, they offer a non-invasive and, in many cases, pre-clinical window into brain health.

Advancements in machine learning (ML) and natural language processing (NLP) are being explored to enhance the identification of communication-specific changes in performance. Hand-crafted features describing acoustic and linguistic parameters (such as semantics, prosody, and pitch changes) can inform outcome measure development in addition to agnostic computational approaches such as deep learning.¹³ These complementary techniques have the potential to identify and characterize patterns of early cognitive impairment. In context, this review aims to synthesize evidence for domain-specific speech and language outcome measures across the AD continuum, summarize emerging associations with biomarkers, and considers how ML-enabled automation may support scalable screening and monitoring.

Communication profiles of people with Alzheimer's disease

The communication profile of AD is characterized by both speech and language deficits, manifesting as impairments across prosody, phonology, articulation, morphology, syntax, semantics, and pragmatics.¹⁴

Challenges at the phonetic (speech-motor) level can manifest as difficulties in articulation, namely, imprecise consonants or prosody.¹⁵ In this review, prosody is treated as a domain in its own right; however, it reflects both speech-motor implementation and higher-level linguistic processes.¹⁶ Morphological impairments lead to trouble in producing correct word endings. Syntax-related changes in AD involve syntactic simplification, such as shorter utterances and fewer embedded clauses.^17,18 A prominent feature of AD is anomia. Word-finding difficulties contribute to impoverished vocabulary and unpredictable word-finding failures, sometimes referred to as “expressive vocabulary drop out”.^19,20 Lexical and semantic impairments may also underlie challenges in understanding complex sentences.²¹ Difficulties in understanding social cues or maintaining coherent conversations relate to pragmatic deficits.²² Altered prosody leads to challenges in conveying or understanding emotional narratives.²³ These manifestations occur at distinct stages of the disease, culminating in complex communication deficits as AD progresses. The deficits combine to produce a communication profile that makes daily activities difficult to complete and social interactions challenging for all those involved. Table 1 illustrates changes in linguistic domains across AD stages and provides examples of how their interpretation may vary across languages.

Table 1.

Changes in linguistic domains across Alzheimer's disease stages.

	Phonology	Morphology	Syntax	Semantics	Pragmatics	Prosody
Definition	The study of sound and sound patterns including rules that specify how sounds are used to form meaningful words.	Rules used to form meaningful words. Studies internal word structure and how morphemes (meaningful units) combine. ²⁴	Rules that specify word order, sentence structure, and phrases.²⁵	Rules governing the meaning of sentences, words, and phrases and their interpretation by the speaker and listener.²⁶	Concerns language use in day-to-day life. Involves the use of communication goals and context interpretation.²⁷	Intonation, stress, rhythm, and pitch variation patterns are used to convey speakers’ emotions.²⁸
Example	The word “cat” consists of, /k/, /æ/, and /t/ sounds. These sounds are called phonemes and vary based on spoken language.	‘Child’ (free lexical morpheme)+-ish (bound derivational morpheme) = ‘childish’. Languages differ in how much morphology is used. Some are more isolating (e.g., Mandarin; fewer affixes) while Turkish and Japanese are agglutinative. Many affixes combine into a word.	Any sentence following the word order subject-verb-object.	A word could change its meaning with context, “move” can mean various things like pull, push, or replace.	Changes in language and speech concerning social context. Indirect request: “Can you pass the water?” ≠ ability question.	A sentence can be delivered with various emotions- joy, sadness, anger.
AD STAGE
MCI	Subtle changes inability to process sounds can occur. However, they go unnoticed during everyday communication.¹⁴	Individuals with MCI show reduction in ability to learn new words and their meanings.¹⁴	Basic language abilities are usually preserved.²⁹To determine when and how much syntax is impacted in MCI, automated methods, greater sample sizes, and longitudinal analyses are required.¹⁴	Basic semantic knowledge tends to be intact.Some deficits were observed in lexical access.³⁰	No clear deficits are evident.	People with MCI present with difficulty in differentiating ironic or sincere stories that require mental inference capacities, compared to control subjects.³¹
Early	Speech is fluent and grammatical with no observable changes.Individuals can form coherent words and meaningful sentences, with no hesitation in speech.Individuals can still make their first language sounds and communicate as needed.	Vocabulary is intact for often used words and items while less frequently used words and objects are more challenging.The experience is often described as a “Tip of tongue” experience.In this stage, individuals show impairment in vocabulary with an observable progressive decline.³²	A limited number of studies are available in this domain.Individuals in early-stage AD show the least impairment in processing syntax.³²	PwAD demonstrate impairments in semantic processing.Early manifestations include difficulties with word recognition and naming (anomia), progressing to impaired word comprehension and reduced perception of objects.	AD is characterized by a decline in pragmatic competence, particularly in the illocutionary (speech acts of pressure and advice) and perlocutionary (acts of persuasion) functions impacting interactive and expressive verbal communication.	Speech is often described as “flattened” and lacks emotional expression.A noticeable monotonous voice is observed alongside deficits in visual and auditory emotional coding and decoding.Individuals may confuse emotions such as happiness, sadness, and neutral intonation with each other.³³
Mid	Distortions, sound substitutions, and prolonged intersegment intervals are commonly observed.³⁴	Individuals show impairment, particularly with finding the right word.Understanding phrases and text alongside production of narrative speech is impaired.Individuals at this stage find word fluency tasks more challenging than confrontation naming tasks, where individuals name pictures or tangible objects.	Complex syntactic components without syntactic cues are changeling to comprehendDifficulty with active voice constructions.	Fluency and naming tasks reveal that reading comprehension and narrative speech production are both impaired.	Comprehension deficits, and paraphrastic errors are frequently observed.Loss of “interactive and expressive linguistic features”.³⁵ is noted, impacting the ability to maintain conversations and understand social cues.Individuals find it challenging to use the right words according to context.	Individuals show a decline in emotional prosody.Changes in speech tempo and hesitation ratio are evident.
Late	An increase in hesitation is observed, leading to a gradual loss of speech.	Anomia, nominal aphasia, dysphasia, and mutism are commonly observed.	Semantic processing is more compromised than syntax processing.³⁶	Individuals commonly substitute superordinate category names (basic-level categories dog, cat, fish, etc.) with target names.	In some cases, individuals experience a complete loss of communication.	Both expression and reception abilities are impaired.Expressive tasks can result in reduced pitch variation, while receptive tasks show decline in emotional perception compared to healthy individuals.³⁷

Language skills

AD causes marked pragmatic deficits in its MCI and early stages.^15,22,38 Pragmatics encompass communication behaviors such as turn-taking in conversations, appropriate conversational deployment of prosodic cues (pitch, loudness, timing), understanding nonverbal cues (facial expression and body language), adapting to the listener's needs, and interpretation of implied meaning or context. Individuals with AD have longer and more frequent pauses, which result in the disruption of the natural flow of conversation.²⁹ Difficulty maintaining conversations, following social cues, and speaking loudly and out of turn are examples of pragmatic impairments known to occur in AD.³⁹ These pragmatic difficulties are increased by deterioration in acoustic-phonetic features, such as variability in pitch, loudness, and timing, all of which are impacted by disease.¹² The latency of responses, the maintenance and appropriateness of the topic and the ability to respond dynamically to communication partners are all influenced by language competence.

Semantic impairment is also a dominant feature of AD.³⁸ Semantic language skills involve the comprehension and expression of conceptual meaning (i.e., word concept mapping and its conceptual relationship).⁴⁰ People with AD (pwAD) can present with deficits in comprehending word meanings,⁴¹ with difficulty understanding the definitions of words and concepts,⁴² including their denotations (literal meanings) and connotations (associated meanings).⁴³ Difficulties extend to recognizing meaning-based word relationships, such as synonyms and antonyms and how words relate to each other in sentences.⁴⁴ Word relationships can change on the basis of context, and recognizing those relationships appears challenging in AD.⁴⁵ When language becomes more abstract, such as figurative language, where speakers use metaphors, similes, idioms, and other forms of nonliteral meanings, pwAD often lose track of communicative intent.⁴⁶ This relies on the semantics-pragmatics interface, which is discussed further under pragmatics. Accessing and using a lexical store can translate into profound word-finding difficulties, the use of incorrect words, and loss of verbal fluency.¹⁵ Although these features reflect lexical retrieval impairment, semantic degradation may contribute as the disease progresses. Language impairments typically begin with signs of anomia, which later progress to difficulties in understanding words and reduced object perception. In later stages, individuals with AD experience impaired fluency and naming, alongside deficits in reading and narrative speech production.³⁸ These can co-occur with slower speech and false starts.^47,48

Syntax refers to the set of rules and principles that govern the arrangement of words and phrases in a sentence to create coherent and meaningful language. Evidence shows that AD alters syntax in a selective and quantifiable manner.^49,50 A study using Cookie-Theft picture description task reported that individuals with MCI retained their basic sentence-building ability, producing an equivalent number of well-formed utterances to healthy individuals.¹⁸ However, they showed a reduction in syntactic complexity with shorter clauses and overall shrinkage in sentence length. While the bulk of grammar remains intact until later in the disease, this can be attributed to the relationship with procedural memory systems,⁵¹ as procedural memory is thought to oversee learning mental grammar rules.⁵² Automated methods, larger sample sizes, and longitudinal analyses are needed to understand the degree of impact AD has on syntax.

Morphology is a branch of linguistics that addresses the internal structure and formation of individual words or parts of words, such as prefixes, suffixes, and roots. The smallest meaning-bearing units are called morphemes, and they can be inflectional (containing grammatical information such as tense, number or agreement, e.g., distribute-distributed, dog-dogs, she walks), or derivational (forming new words that change meaning, e.g., friend-friendly, happy-happiness). In AD, studies show that morphological decomposition during word recognition remains relatively preserved even in the context of lexical impairment.^53,54 However, connected-speech and morphosyntactic analyses show that word-level grammatical marking can deteriorate in AD with errors specific to inflectional morphology.⁵⁴ This feature shows further decline as disease severity increases and task demands rise.⁵⁵ Research in Finnish-speaking population has shown changes in morphological organization during word recognition (effects of morphological family structure), changes across healthy aging, MCI and AD, showcasing the potential sensitivity of morphology-related measures depending on the task and language.²⁴ Overall, morphological changes are impaired less consistently when compared to lexico-semantic markers, but they offer a possibility of complementary information in languages that are rich in inflectional systems and in tasks that focus on grammatical encoding.

Phonology refers to the systematic organization and rules governing the sounds (phonemes) used and how these sounds function to convey meaning and distinguish words. In the context of AD, phonological errors may occur when individuals substitute one sound for another, giving the word a new meaning or rendering it unintelligible. For example, they may substitute “cat” with “bat” or “dog” with “log.” These errors can often be confused for articulation deficits or distortions; however, their origin lies in language-based impairments affecting phonological encoding. Phonological paraphasia increases as AD progresses.⁵⁶ In MCI, phonological changes can be subtle and often go unnoticed.¹⁴ In moderate AD, word-level sound substitutions may help distinguish pwAD from cognitively healthy individuals.⁵⁷ Phonological abilities are markedly compromised in late AD, where speech may involve only a few mumbled words, echolalia or repetitive syllables.⁵⁸

Phonetic changes reflect alterations in the speech signal, i.e., show how speech is physically produced and sounds. This includes timing, fluency, phonation, voice quality and acoustic characteristics. In MCI cohorts, temporal variables (pause metrics and speech rate) can be extracted as “acoustic markers”.⁵⁷ This supports the idea that changes in phonetics manifest as increased pausing and hesitation rather than a simple slowed speech rate. Articulatory distortions, prolonged intersegment duration and vowel prolongations are examples of phonetic alterations that occur in the mild stage of AD and deteriorate as the disease progresses.⁵⁹ Therefore, individuals with AD are reported to speak at a slower rate and with frequent hesitations.¹⁵ Phonetic fluency is further affected by increased pause frequency as individuals search for words.¹⁵ In late AD stages, individuals often lose their ability to produce coherent speech due to the combined impact of cognitive impairment and motor decline.¹⁴

With progression to mid-AD pragmatics becomes compromised. Individuals with AD are reported to disproportionately rely on formulaic expressions, such as greetings, discourse markers, idioms, and routine fillers. In a conversational study, AD subgroups produced significantly more words in formulaic expressions than age-matched healthy individuals.⁶⁰ Lexical analyses have also identified increased reliance on high-frequency word co-occurrences correlating with post-symptom onset, showcasing reduced lexical diversity with disease progression.⁶¹

Concurrently, prosodic features are also affected. Changes in speech tempo and hesitation ratios are often observed, and speech is characterized as flattened and without emotion.¹⁵ These deficits are associated with impaired visual and auditory emotional coding and decoding. Happiness, sadness, and neutral intonation are often confused with one another.³³ In advanced stages, the expression and reception of prosody are further diminished. The use of expressive tasks for individuals in late AD results in less pitch variation with decreased perception of emotion in receptive tasks.³⁷

Figure 2 presents a hypothetical trajectory of speech and language functions in AD based on existing published literature.

Figure 2.

Theoretical progression of speech and language functions across the Alzheimer's disease continuum. The graph plots relative functional integrity on the y-axis (1 = intact, 0 = lost) for seven linguistic domains: semantics, acoustic and temporal measures, pragmatics, prosody, syntax, morphology, and phonology. The x-axis marks six clinical stages: normal aging, preclinical Alzheimer's disease (pcAD), mild cognitive impairment (MCI), early AD, mid AD, and late AD. Curves are schematic, based on published cross-sectional and longitudinal work, and illustrate an asynchronous pattern of decline. Semantics changes begin in pcAD and accelerate sharply in early AD. Acoustic and temporal measures, pragmatic and prosodic skills start to deteriorate during MCI and degrade thereafter; syntax and morphology are relatively preserved until early AD but decline thereafter; phonological processing is spared compared to other functions until the late stage of the disease. This model underscores the importance of selecting speech biomarkers that are both domain-specific and stage-sensitive for screening and monitoring in AD research and care.

Motor speech

Motor speech disorders are not frequently observed in AD.⁶² When present, articulatory changes in AD can be the result of subtle variations in spectral characteristics such as rhythm, intonation, and frequency.⁶³ The speech of pwAD is characterized by a slower rate or rhythm of glottal pulses. This can lead to a monotone voice, a greater proportion and number of voice breaks, and greater fluctuations in the amplitude of sound.⁶³ Consequently, changes in pause length and phonation time have been useful in discriminating between participants with MCI and healthy individuals.⁶⁴

In the initial stages of AD, there can be ideational, ideomotor, dressing, and constructional apraxia. This can transition to constructional, buccopharyngeal, and gait apraxia in the later stages of the disease.⁶⁵ The ability to sequence muscle movements to accurately produce speech sounds can also be impaired in apraxia of speech.⁵⁷ Impairment in the nonverbal movement of the cheeks, lips, tongue, and pharynx and difficulty in motor plans is known as orofacial apraxia.⁶⁶ There are reports of subtle or mild apraxia of speech across disease stages, with the symptoms worsening as AD progresses.⁶⁷ Apraxia of speech causes significant disturbances in articulation and prosody, which worsen as complexity and word length increase.^68,69

Relationships between biomarkers and speech/language markers

The severity of cognitive impairment in AD is associated with neurofibrillary tangle pathology,⁷⁰ a key feature of AD histopathology.⁷¹ Subjective cognitive complaints from pwAD tend to be associated with greater Aβ burden and more rapid functional decline.⁷² However, evidence for subjective cognition‒amyloid associations in healthy older individuals has not been observed consistently across multiple subjective cognitive self-report measures.⁷³ Recent studies have highlighted the promise of using speech and language measures to detect AD at different stages of the disease.^14,74–76 However, the links among speech and language markers, brain imaging, and CSF biomarkers require further investigation.

Approximately 30% of adults over the age of fifty have altered vocal function. This can arise from both disease onset and age-related physiological changes.^77,78 Despite this prevalence, there is limited evidence to support pre-symptomatic diagnosis in everyday practice.⁷⁹ Accurate identification of speech and language features that distinguish preclinical and prodromal AD from typical aging could enhance a clinician's ability to identify individuals at risk earlier in the disease course.^80,87 However, evidence in preclinical AD remains limited and uneven across linguistic domains, with prosody and morphology particularly understudied. Moreover, studies report overlapping semantic changes between preclinical AD and cognitively “normal” cohorts, underscoring the need for more sensitive and domain-specific markers.^81,82

To substantiate the value of behavioral markers such as communication in clinical practice, we need to establish how they are related to validated AD biomarkers. Currently, the field lacks standardized datasets on which different acquisition and analysis approaches can be systematically compared.⁸³ Mapping speech measures onto validated AD biomarkers could allow early detection and longitudinal monitoring of AD and improve prophylactic measures against disease progression.

Plasma biomarkers such as Aβ and p-tau are efficient indicators of AD pathology.⁸⁴ Compared to imaging- and CSF-based measures, blood-based biomarkers offer advantages in scalability, accessibility and suitability for repeated sampling. Examining how these biomarkers relate to speech and language changes may therefore provide complementary insights into early disease detection and longitudinal monitoring.

Several studies converge on semantic vulnerability, particularly in connected speech. Aβ-positive participants presented with a notable decline in specific word content in spontaneous speech tasks,⁸¹ whereas lexical diversity may remain unchanged, suggesting informativeness and semantic richness can be more sensitive than broad lexical counts.⁸⁵ Aβ load and linguistic deficits have also been correlated.

Domain-specific speech and language associations with amyloid and tau burden

Individuals with higher Aβ loads produce fewer concrete nouns and fewer concrete words overall.⁸² This is consistent with subtle semantic changes in preclinical AD.⁸² A correlation with education level was also found; individuals with higher education levels use more abstract nouns despite Aβ levels, suggesting an influence on cognitive reserve.⁸² The influence of factors such as education on the variability of speech and language performance indicates that standardized assessments may not capture the full picture of every individual. However, this highlights the need for personalized approaches to testing, where cognitive reserve and individual differences can be accounted for. Longitudinal work also supports this clinical relevance. A study using the ADNI dataset showed that the rate of change in semantic fluency was significantly associated with Aβ burden.⁸⁶ Importantly, the deterioration in semantic fluency from MCI to AD was partly correlated with Aβ levels but not with tau or neurodegeneration, marking it an important risk factor for conversion of MCI into AD. As reported in Figure 2, in contrast to semantics, syntax and morphology are relatively preserved until late AD. This is reflected in the lack of strong biomarker associations, where Aβ-positive individuals did not differ from Aβ-negative individuals in syntactic complexity.^86,87 Additionally, research using social conversation recordings and biomarker data show correlation of increased amyloid pathology with simpler speech. Low CSF Aβ levels resulted in lower syntactic complexity.⁸⁸ Although these links were found to be modest, they suggest a meaningful correlation.

Acoustic temporal measures have shown associations with tau pathology. A multimodal study of cognitively unimpaired individuals found that speech patterns in a delayed recall on a story memory task correlated with tau PET signal across adulthood. Longer and greater between-utterance pause time, and a slowed speech rate were associated with high tau burden.⁸⁹ The percentage of silence duration has also shown correlations with p-Tau217 and Aβ deposition.^87,89 Wang et al. showed that the silence duration increased in a two-stage pattern alongside an increase in Aβ and p-Tau. In early Aβ accumulation stages, silence duration increased in parallel with p-Tau217.⁸⁷ However, when cortical Aβ levels increase further, associations with changes in silence duration are relatively small. This suggests that frequent pausing in connected speech may be a sensitive marker of early AD pathology that plateaus after substantial Aβ load.

In the Aβ-positive, cognitively impaired group, a reduction in speech fluency and speech rate is reported, resulting from extended pausing and intervals between words. Articulatory precision showed the highest sensitivity in differentiating between Aβ-positive and Aβ-negative groups.⁹⁰

Fine-grained phonetic features represented by Mel-frequency cepstral coefficients (MFCCs) reflect the voice quality/timbre and articulatory properties. Siddiqui et al. found that greater variance in MFCC significantly predicted higher p-tau217 levels. Individuals with higher tau had measurably different voice acoustic patterns. This suggests that micro-level phonetic changes (possibly related to vocal tract control or clarity of speech sounds) accompany tau accumulation. In the same study, the MCI cohort, with a slower articulation rate (fewer syllables per second), was also associated with higher p-tau217 values. These findings position phonological production markers, such as articulation rate and voice spectral characteristics, as sensitive markers of p-tau217.⁹¹

Relative gaps in morphosyntax, prosody and pragmatics

In contrast to semantic content and pausing, biomarker-linked evidence for syntax and morphology remains limited and inconsistent. Although pragmatic language testing involving figurative language interpretation has been validated as a sensitive tool for early cognitive decline,¹² pragmatic competence and discourse coherence are understudied with respect to biomarker positivity, particularly across the full AD continuum. There is a critical need for studies that test multi-domain feature sets within the same biomarker-characterized cohorts and show how different biomarkers (Aβ, p-Tau181, p-Tau217 and neurodegeneration markers) associate with semantic, temporal, and pragmatic level features.

Machine learning for automated speech and language assessment in Alzheimer's disease: current capabilities and limitations

Building on domain-specific speech and language changes shown above, recent work has applied NLP and ML to use these features for automated classification, primarily in symptomatic AD and MCI cohorts. This section focuses on what ML pipelines measure, how they are evaluated and why current models do not yet include validated tools for preclinical detection.

Nyongesa et al. automatically processed speech transcripts with NLP to derive fifty-three linguistic features across lexical, syntactic and discourse domains in MCI, AD and cognitively unimpaired groups.⁹² Compared with cognitively unimpaired and MCI participants, the AD group resulted in higher pronoun use and reduced syntactic complexity, consistent with impaired informativeness and simplification of connected speech. Higher pronoun rates and lower complexity were associated with lower MMSE scores, showing that automated transcript features can track clinical severity.

Beyond word and sentence level counts, artificial intelligence-based (AI) approaches can quantify discourse-level properties such as idea density and coherence in narrative speech.⁹³ Research has demonstrated that automated discourse coherence analysis captures discourse disruption in AD, with weaker semantic continuity between utterances and reduced logical alignment to the narrative theme.^94,95 These discourse-focused AI approaches may complement vocabulary-based markers by examining higher-level organization of discourse that is not well captured by word counts alone.

Most ML studies in this area follow a similar pipeline, which begins with the collection of speech in structured tasks (e.g., picture description, story recall, reading), extraction of acoustic and/or linguistic features and training models to classify diagnostic groups or predict symptom severity.

Across languages and elicitation tasks (e.g., picture description, reading and story recall), models using acoustic-temporal measures and transcript-derived features consistently captured slower speech rate, increased pausing, syntactic simplification and reduction in semantic informativeness, alongside lowered idea density. Collectively, these features correlate with cognitive decline across studies.^96–99

While multimodal models combining speech with clinical and neuropathological data may improve discrimination between disease severity, speech-only pipelines remain attractive for their low-burden, scalable and remote collection.¹⁰⁰ By employing automatic ML techniques for analysis of spontaneous speech, studies have suggested that computational approaches alone can achieve up to 95% accuracy in detecting dementia compared with healthy individuals.⁵⁰ These findings suggest that automatic techniques could be a viable option for effective triage and severity estimation within clinical workflows.

Across classical ML models, support vector machines (SVMs) have shown that structured connected speech features and paralinguistic markers can detect AD related impairment.¹⁰¹ Using this method, discourse and prosodic markers, namely, filled-pause rate, post-pause latency and post-pause lexical frequency are well documented in MCI.¹⁰² Tree-based models such as random forests (RFs) have emerged as the most effective methods for differentiating individuals with subjective cognitive decline from those with MCI and AD.¹⁰³ Automated extraction of tempo, pausing and utterance length is particularly useful in remote settings and aligns with evidence that discourse and prosodic markers are sensitive to cognitive load.⁵⁹

Deep learning models such as Densenset121 have been used to analyze time-frequency-based metrics of speech.¹⁰⁴ Using MFCC-based spectrogram images provided a pipeline for studying acoustic patterns from speech without hand-selecting individual timing or prosodic features. The reported performance of this dataset reached a sensitivity of 95.5% and specificity of 83.3% for AD detection, showcasing the vast majority of correctly identified AD cases while keeping false positives relatively low (precision of 88.9% in cross-validation). While these results show the potential of speech-based classification in standardized datasets, performance largely depends on cohort composition, task design and requires external validation before clinical deployment.

By contrast, efforts such as ADReSS show that performance is achievable under standardized conditions, but sensitivity-specificity trade-offs and dataset shift remain limitations for clinical deployment of such models. In the ADReSS challenge: a 2020 public benchmark for Alzheimer's speech recognition,¹⁰⁵ a model integrated acoustic prosodic features with disfluency patterns and reached ∼85.4% accuracy (versus 77% baseline) in classifying AD versus healthy speech.¹⁰⁵ However, there was variability in sensitivity versus precision: one ADReSS submission achieved 91.7% specificity and 88.9% precision, but a more moderate 72.7% sensitivity.¹⁰⁶ This means that while false positives were very low, the model missed ∼27% of true AD cases, a trade-off sometimes seen when optimizing for overall accuracy. In contrast, other approaches prioritized sensitivity to catch nearly all AD cases at the cost of more false alarms. For clinical triage, sensitivity is often prioritized to minimize missed cases, while for population screening, false positives can present as a substantial burden. Therefore, model thresholds should be aligned to the intended use case.

Beyond their limitations these benchmark studies show promise in methodological comparison; however, they largely evaluate symptom classification rather than early detection. Importantly, it should be noted that classifying AD vs healthy participants is not equivalent to detecting AD. Diagnostic models learn patterns of established impairment, while early clinical detection requires identifying subtle changes and predicting biomarker positivity or future decline. Accordingly, the existing ML models are best positioned as tools for automated feature extraction and triage (flagging likely impairment and estimating severity), rather than stand-alone screening tests for AD. However, differentiation in cognitively unimpaired, MCI and AD is useful for mining speech and language features that are domain-specific.

Taking these features into account, early detection would require identifying subtle changes and distinguishing them from risk signals of normal aging, education, and comorbidities. As Lindsay et al. note, spontaneous speech is a compound cognitive task engaging memory, attention, and executive function in addition to language.³⁹ An individual might speak slowly not only because of language formulation issues but also due to difficulty in recollecting words or even hearing loss. Future ML models should aim to disentangle these factors, or at least incorporate them (for instance, by also assessing memory through the speech content).

Overall, ML work is currently saturated in acoustic-temporal measures and coarse lexico-syntactic features. Automated assessment of morphology, prosody, and pragmatic competence is less explored and less frequently validated. Integration of automated assessment of morphology and discourse-pragmatic competence would further improve screening strategies. Ideally, future models could integrate multi-level language measures with speech acoustics and test across languages and recording contexts while linking outputs to the AD biomarker status.

Advantages and disadvantages of clinical speech markers

Speech provides valuable insight into both neurological and motor functions because of the complex coordination of the cognitive, linguistic, and motor processes involved.⁹⁵ There has been increased interest in the use of speech analytics and mobile health technologies for tracking and evaluating health because of their efficiency, accessibility, and affordability.^107,108 Mobile health technologies have the potential to offer greater convenience, personalization, and accessibility. Using speech analytics for early disease identification is particularly promising because speech and voice alterations, such as slowed speech rates or diminished loudness, can be early symptoms of brain health.¹⁰⁹ This method also allows easy monitoring of disease progression; convenient technology enables data to be collected frequently and remotely. However, most remote speech protocols currently include acoustic-temporal features (rate, pauses, duration). App-based tasks could provide high-level pragmatics and morphosyntax by incorporating tests that require richer interactions and longer samples with production of more reliable transcripts.

Remote testing is attractive to health researchers, trialists and clinicians because it can reduce accessibility limitations. Visiting clinical sites in person places an additional burden on patients and families, particularly elderly populations with limited mobility.¹¹⁰ Another advantage of using speech in AD is its persistent relevance throughout the disease progression process. Speech is sensitive to cognitive decline at the preclinical stage, and changes in speech manifest in later stages of AD.^108,109 Analysis of a single speech sample can reveal various aspects of an individual's cognitive and motor functions, and repeated measurements provide data on longitudinal changes. Throughout literature, changes in semantics and lexical fluency measures are well documented, but we still lack comparable repeated measures data for prosody, pragmatic coherence and turn-taking across the full AD continuum, which can be fulfilled through remote monitoring of day-to-day conversations.

The collection of speech data is a noninvasive method that poses minimal risk of inducing anxiety , unlike cerebrospinal fluid or blood-based analyses.^110–112 Linguistic features, such as syntactic complexity and information content, can be automatically transcribed from low-quality audio recordings. Furthermore, certain acoustic features, such as the fundamental frequency, are robust and can be recorded via affordable equipment such as a smartphone microphone.¹⁰⁹ Speech clinical markers have demonstrated usefulness in remote testing and active or passive data collection contexts.¹¹² It is feasible that data collected from clinical interviews could provide communication information, such as confrontation naming, single-word production, and word generation in context. However, in some linguistic contexts (e.g., structured interviews with limited opportunity for contemporaneous connected speech), the amount and type of speech is not sufficient for accurate descriptions of impairment in pwAD, who score poorly and do not capture early symptoms observed by family members during normal conversation.¹¹³ To truly capture communicative ability, communication should be examined in environments where it is most affected, such as home-based conversations. This limitation is highly relevant for pragmatic and prosodic functions that appear in natural interactions. In contrast, isolated tasks can still be informative for lexical-semantic and phonetic markers.

Preprocessing and quality assurance of speech samples remains a challenge. Some models used in automatic speech recognition have yet to be optimized for disordered speech, as observed in AD. As such, manual transcription may still be necessary for some spontaneous speech tasks, which can be labor-intensive.⁷⁵ Additionally, many acoustic features are sensitive to changes in software and hardware configurations, sampling rates, and recording device quality, which can increase the cost of speech data collection.^50,114 Currently, most studies on speech clinical markers use nonstandardized methods that vary between groups, which restricts the translation of these studies into clinical practice.¹¹⁵ However, these technical issues do not affect all linguistic domains equally. ASR errors could distort morphology and syntax measures, low-quality recordings can blunt prosodic features, and a lack of interaction-based conversations limits pragmatic interpretation. Therefore, standardization also requires domain-specific benchmarks.

Speech as a screening tool can be used to detect signs of AD-related impairment in asymptomatic or mildly symptomatic populations, prompting further evaluation. An example would be a speech screening app for at-risk individuals which can identify speech patterns suggestive of cognitive decline. As a monitoring tool, speech analysis could be used in individuals already diagnosed with MCI or AD to track their progression and treatment effects.

Some speech features might be more relevant for screening than monitoring. Lexical richness might serve as a screening metric,¹¹⁶ while change in pause duration over six months serves as a monitoring metric.¹¹⁷ In practice, the same underlying speech analysis can provide both outputs- flagging for detection and quantifying changes in diagnosed cases.¹¹⁸

Currently, speech-based AD assessments are largely in the research stage. On a clinical front, speech and language evaluation has been qualitative. This includes noting if individuals have word-finding difficulties during the assessment through tasks, namely, the Cookie Theft picture description from the Boston Diagnostic Aphasia Exam.¹⁸ Objective speech analysis with calculated features has not yet been adopted in routine clinical assessments. Augmenting tasks used in the clinic AI analysis could instantly compute speech metrics and compare them to AD norms. In clinical research and trials, speech analysis is already gaining traction. Pharmaceutical trials for early AD have started to include digital biomarkers such as speech alongside traditional scales to observe changes in speech features with drug treatment.¹¹⁹ Research has shown high Aβ burden to be associated with less specific words and semantic fluency.¹²⁰ Frequent pauses and low speech rate are observed in individuals with higher tau burden.⁸⁹ Thus, validation of speech markers against established biomarkers of AD shows the potential for clinical application.

Despite this potential, several challenges must be addressed before speech analysis can be employed in clinical care settings. Speech patterns vary widely between populations due to factors such as education levels, native language and dialect.⁹² Vergaillie et al. showed that, with similar levels of Aβ, people with higher education levels used more diverse and abstract vocabulary than those with less education.⁸² Thus, the development of scales using speech should be adjusted for demographic factors to avoid misleading diagnoses. However, demographic adjustment must extend beyond vocabulary metrics to account for accent and dialect-sensitive phonetics/prosody and language-specific morphology/syntax patterns to avoid misattribution of changes in pathology to sociolinguistic variation. Age-related changes in speech are another confound; healthy older adults may have a slower speech rate and long pauses.¹²¹ Solutions must be developed to distinguish normal aging from pathological aging. This could include databases stratified by age, education or creating algorithms that use study populations as their baseline.

Conclusion

Changes in speech and language function can occur early in AD. Communication difficulties arise from a combination of cognitive decline and, in some cases, motor speech changes. These cognitive-communication challenges can manifest as deficits in speech signal (rate, pausing, duration) and extend to linguistic levels. At a lexico-semantic level, speech may become less informative, with fewer nouns and greater reliance on vague terms or pronouns. Changes in syntactic planning manifest as shorter utterances and simple sentence structures. Reduced and incorrect use of tense is related to morphological deficits. Phonological changes include sound substitutions or segment omissions. Prosodic expressivity can flatten with reduced emotional prosody, and pragmatic abilities may decline with poorer topic maintenance and reduced narrative coherence. Accordingly, connected speech points towards functionally meaningful assessments that may bridge the gap between behavioral testing and biomarkers. Speech and language testing advancements show promise for detecting subtle changes in speech and language patterns to diagnose individuals more efficiently. Speech data could improve screening protocols by including outcomes that are relevant to everyday communication and quality of life. Existing clinical interviews and standardized elicitation batteries can be used to collect speech data. The use of speech screening measures before traditional AD diagnosis methods may be cost-effective and noninvasive. Future work should include standardized protocols for speech data collection, demographic and language-aware norms, and validation of multi-level linguistic markers against AD biomarkers supported by ML and NLP methods.

Footnotes

Acknowledgements

The authors have no acknowledgments to report.

ORCID iDs

Shloka Dhareshwar

Ruby Mineur

Peter J. Snyder

Jessica Alber

Yen Ying Lim

Adam P. Vogel

Author contribution(s)

Shloka Dhareshwar: Formal analysis; Investigation; Methodology; Visualization; Writing – original draft.

Melissa Kane: Investigation; Writing – original draft.

Courtney Lewis: Investigation; Validation; Writing – original draft.

Ruby Mineur: Investigation; Writing – original draft.

Kylie Yao: Writing – review & editing.

Peter J. Snyder: Writing – review & editing.

Jessica Alber: Writing – review & editing.

Catherine Robb: Supervision; Writing – review & editing.

Yen Ying Lim: Supervision; Writing – review & editing.

Adam P. Vogel: Conceptualization; Project administration; Supervision; Validation; Writing – review & editing.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: Adam P. Vogel is funded by an Australian Research Council Fellowship (grant number #220100253).

Declaration of conflicting interests

The authors declared the following potential conflicts of interest with respect to the research, authorship, and/or publication of this article: Adam P. Vogel is an employee of Redenlab Ltd.

References

Gustavsson

Norton

Fast

, et al. Global estimates on the number of persons across the Alzheimer’s disease continuum. Alzheimers Dement 2023; 19: 658–670.

Rajmohan

Reddy

. Amyloid-beta and phosphorylated tau accumulations cause abnormalities at synapses of Alzheimer’s disease neurons. J Alzheimers Dis 2017; 57: 975–999.

Jack

Andrews

Beach

, et al. Revised criteria for diagnosis and staging of Alzheimer’s disease: alzheimer’s Association Workgroup. Alzheimers Dement 2024; 20: 5143–5169.

DeTure

Dickson

. The neuropathological diagnosis of Alzheimer’s disease. Mol Neurodegener 2019; 14: 32.

Van de Pol

Hensel

Barkhof

, et al. Hippocampal atrophy in Alzheimer disease: age matters. Neurology 2006; 66: 236–238.

McKhann

Drachman

Folstein

, et al. Clinical diagnosis of Alzheimer’s disease: report of the NINCDS-ADRDA Work Group* under the auspices of Department of Health and Human Services Task Force on Alzheimer’s Disease. Neurology 1984; 34: 939–939.

Juganavar

Joshi

Shegekar

. Navigating early Alzheimer’s diagnosis: a comprehensive review of diagnostic innovations. Cureus 2023; 15: e44937.

Zhu

Ploessl

Kung

. PET/SPECT imaging agents for neurodegenerative diseases. Chem Soc Rev 2014; 43: 6683–6691.

Rosselli

Uribe

Ahne

, et al. Culture, ethnicity, and level of education in Alzheimer’s disease. Neurotherapeutics 2022; 19: 26.

10.

Alzola

Carnero

Bermejo-Pareja

, et al. Neuropsychological assessment for early detection and diagnosis of dementia: current knowledge and new insights. J Clin Med 2024; 13: 3442.

11.

Georgiou

Prapiadou

Thomopoulos

, et al. Naming ability assessment in neurocognitive disorders: a clinician’s perspective. BMC Psychiatry 2022; 22: 837.

12.

Chakrabarty

Klooster

Biswas

, et al. The scope of using pragmatic language tests for early detection of dementia: a systematic review of investigations using figurative language. Alzheimers Dement 2023; 19: 4705–4728.

13.

Mahajan

Baths

. Acoustic and language based deep learning approaches for Alzheimer’s dementia detection from spontaneous speech. Front Aging Neurosci 2021; 13: 623607.

14.

Mueller

Hermann

Mecollari

, et al. Connected speech and language in mild cognitive impairment and Alzheimer’s disease: a review of picture description tasks. J Clin Exp Neuropsychol 2018; 40: 917–939.

15.

Szatloczki

Hoffmann

Vincze

, et al. Speaking in Alzheimer’s disease, is that an early sign? Importance of changes in language abilities in Alzheimer’s disease. Front Aging Neurosci 2015; 7: 195.

16.

Wagner

Watson

. Experimental and theoretical advances in prosody: a review. Lang Cogn Process 2010; 25: 905–945.

17.

Lian

Wang

. Dependency grammar approach to the syntactic complexity in the discourse of Alzheimer patients. Behav Sci 2025; 15: 1334.

18.

Ivanova

Martínez-Nicolás

García-Piñuela

, et al.

Defying syntactic preservation in Alzheimer’s disease: what type of impairment predicts syntactic change in dementia (if it does) and why?

Front Lang Sci 2023; 2: 1199107.

19.

Reilly

Flurie

Ungrady

. Eyetracking during picture naming predicts future vocabulary dropout in progressive anomia. Neuropsychol Rehabil 2022; 32: 560–578.

20.

Venneri

Forbes-Mckay

Shanks

. Impoverishment of spontaneous language and the prediction of Alzheimer’s disease. Brain 2005; 128: E27.

21.

Klimova

Maresova

Valis

, et al. Alzheimer’s disease and language impairments: social intervention and medical treatment. Clin Interv Aging 2015; 10: 1401–1408.

22.

Huang

Che

. Pragmatic impairment and multimodal compensation in older adults with dementia. Lang Health 2023; 1: 44–57.

23.

Themistocleous

. Linguistic and emotional prosody: a systematic review and ALE meta-analysis. Neurosci Biobehav Rev 2025; 175: 106210.

24.

Nikolaev

Ashaie

Hallikainen

, et al. Effects of morphological family on word recognition in normal aging, mild cognitive impairment, and Alzheimer’s disease. Cortex 2019; 116: 91–103.

25.

Zaccarella

Meyer

Makuuchi

, et al. Building by syntax: the neural basis of minimal linguistic structures. Cereb Cortex 2017; 27: 411–421.

26.

Hendriks

. The acquisition of compositional meaning. Philos Trans R Soc B Biol Sci 2020; 375: 20190312.

27.

Martin

Lee

Bicknell

, et al. A longitudinal investigation of pragmatic language across contexts in autism and related neurodevelopmental conditions. Front Neurol 2023; 14: 1155691.

28.

Sinagra

Wiener

. The perception of intonational and emotional speech prosody produced with and without a face mask: an exploratory individual differences study. Cogn Res Princ Implic 2022; 7: 89.

29.

Nasreen

Rohanian

Hough

, et al. Alzheimer’s dementia recognition from spontaneous speech using disfluency and interactional features. Front Comput Sci 2021; 3: 640669.

30.

Taler

Monetta

Sheppard

, et al. Semantic function in mild cognitive impairment. Front Psychol 2020; 10: 3041.

31.

Gaudreau

Monetta

Macoir

, et al. Mental state inferences abilities contribution to verbal irony comprehension in older adults with mild cognitive impairment. Behav Neurol 2015; 2015: 685613.

32.

Emery

. Language impairment in dementia of the Alzheimer type: a hierarchical decline? Int J Psychiatry Med 2000; 30: 145–164.

33.

Tosto

Gasparini

Lenzi

, et al. Prosodic impairment in Alzheimer’s disease: assessment and clinical relevance. J Neuropsychiatry Clin Neurosci 2011; 23: E21–E23.

34.

Cera

Ortiz

Bertolucci

PHF

, et al. Phonetic and phonological aspects of speech in Alzheimer’s disease. Aphasiology 2017; 32: 88–102.

35.

Gutiérrez-Rexach

Schatz

. Cognitive impairment and pragmatics. SpringerPlus 2016; 5: 127.

36.

Kempler

Curtiss

Jackson

. Syntactic and symbolic abilities in Alzheimer’s disease. J Speech Hear Res 1987; 30: 343–350.

37.

Horley

Reid

Burnham

. Emotional prosody perception and production in dementia of the Alzheimer’s type. J Speech Lang Hear Res 2010; 53: 1132–1146.

38.

Ferris

Farlow

. Language impairment in Alzheimer’s disease and benefits of acetylcholinesterase inhibitors. Clin Interv Aging 2013; 8: 1007.

39.

Lindsay

Tröger

König

. Language impairment in Alzheimer’s disease—robust and explainable evidence for ad-related deterioration of spontaneous speech through multilingual machine learning. Front Aging Neurosci 2021; 13: 642033.

40.

Martin

. The critical role of semantic working memory in language comprehension and production. Curr Dir Psychol Sci 2021; 30: 283–291.

41.

Kempler

Goral

. Language and dementia: neuropsychological aspects. Annu Rev Appl Linguist 2008; 28: 73–90.

42.

Macoir

Laforce

Monetta

, et al. Language deficits in major forms of dementia and primary progressive aphasias: an update according to new diagnostic criteria. Gériatrie Psychol Neuropsychiatr Viellissement 2014; 12: 199–208.

43.

Laatu

Portin

Revonsuo

, et al. Knowledge of concept meanings in Alzheimer’s disease. Cortex 1997; 33: 27–45.

44.

Nebes

. Chapter 6: semantic-memory function and dysfunction in Alzheimer’s disease. Adv Psychol 1990; 71: 265–296.

45.

Morris

Becker

. Cognitive neuropsychology of Alzheimer’s disease. 2nd ed. New York: Oxford University Press, 2004.

46.

Roncero

De Almeida

. The importance of being apt: metaphor comprehension in Alzheimer’s disease. Front Hum Neurosci 2014; 8: 973.

47.

Ash

Evans

O’Shea

, et al. Differentiating primary progressive aphasias in a brief sample of connected speech. Neurology 2013; 81: 329–336.

48.

Wilson

Henry

Besbris

, et al. Connected speech production in three variants of primary progressive aphasia. Brain 2010; 133: 2069–2088.

49.

Ahmed

Haigh

A-MF

De Jager

, et al. Connected speech as a marker of disease progression in autopsy-proven Alzheimer’s disease. Brain 2013; 136: 3727–3737.

50.

Vogel

Reece

. Recording speech. In: Ball

(ed) Manual of clinical phonetics. London: Routledge, 2021, pp.217–227.

51.

Keith

McCuddy

Lindberg

, et al. Procedural learning and retention relative to explicit learning and retention in mild cognitive impairment and Alzheimer’s disease using a modification of the trail making test. Neuropsychol Dev Cogn B Aging Neuropsychol Cogn 2023; 30: 669–686.

52.

Faretta-Stutenberg

Brill-Schuetz

Morgan-Short

. Declarative and procedural memory as individual differences in second language acquisition. Biling Lang Cogn 2014; 17: 56–72.

53.

Kavé

Levy

. Preserved morphological decomposition in persons with Alzheimer’s disease. J Speech Lang Hear Res 2004; 47: 835–847.

54.

Altmann

LJP

Kempler

Andersen

. Speech errors in Alzheimer’s disease. J Speech Lang Hear Res 2001; 44: 1069–1082.

55.

Fyndanis

Arfani

Varlokosta

, et al. Morphosyntactic production in Greek- and Italian-speaking individuals with probable Alzheimer’s disease: evidence from subject–verb agreement, tense/time reference, and mood. Aphasiology 2018; 32: 61–87.

56.

Forbes-McKay

Shanks

Venneri

. Profiling spontaneous speech decline in Alzheimer’s disease: a longitudinal study. Acta Neuropsychiatr 2013; 25: 320–327.

57.

Cera

Ortiz

Bertolucci

PHF

, et al. Speech and phonological impairment across Alzheimer’s disease severity. J Commun Disord 2023; 105: 106364.

58.

Boschi

Catricalà

Consonni

, et al. Connected speech in neurodegenerative language disorders: a review. Front Psychol 2017; 8: 269.

59.

Toth

Hoffmann

Gosztolya

, et al. A speech recognition-based solution for the automatic detection of mild cognitive impairment from spontaneous speech. Curr Alzheimer Res 2018; 15: 130–138.

60.

Bridges

Van Lancker Sidtis

. Formulaic language in Alzheimer’s disease. Aphasiology 2013; 27: 799–810.

61.

Zimmerer

Wibrow

Varley

. Formulaic language in people with probable Alzheimer’s disease: a frequency-based approach. J Alzheimers Dis 2016; 53: 1145–1160.

62.

Poole

Vogel

. Linking motor speech function and dementia. In: Martin

Preedy

(eds) Genetics, neurology, behavior, and diet in dementia: the neuroscience of dementia. London: Elsevier, 2020, pp.665–676.

63.

Weekes

BSH

. Aphasia in Alzheimer’s disease and other dementias (ADOD): evidence from Chinese. Am J Alzheimers Dis Demen 2020; 35: 153331752094970.

64.

Meilán

JJG

Martínez-Sánchez

Carro

, et al.

Speech in Alzheimer’s disease: can temporal and acoustic parameters discriminate dementia?

Dement Geriatr Cogn Disord 2014; 37: 327–334.

65.

Roark

Mitchell

Hosom

J-P

, et al. Spoken language derived measures for detecting mild cognitive impairment. IEEE Trans Audio Speech Lang Process 2011; 19: 2081–2090.

66.

Broussolle

Bakchine

Tommasi

, et al. Slowly progressive anarthria with late anterior opercular syndrome: a variant form of frontal cortical atrophy syndromes. J Neurol Sci 1996; 144: 44–58.

67.

Cera

Ortiz

Bertolucci

PHF

, et al. Speech and orofacial apraxias in Alzheimer’s disease. Int Psychogeriatr 2013; 25: 1679–1685.

68.

Utianski

Duffy

Clark

, et al. Prosodic and phonetic subtypes of primary progressive apraxia of speech. Brain Lang 2018; 184: 54–65.

69.

Josephs

Duffy

Strand

, et al. Syndromes dominated by apraxia of speech show distinct characteristics from agrammatic PPA. Neurology 2013; 81: 337–345.

70.

Mostafavi

Gaiteri

Sullivan

, et al. A molecular network of the aging human brain provides insights into the pathology and cognitive decline of Alzheimer’s disease. Nat Neurosci 2018; 21: 811–819.

71.

Nelson

Alafuzoff

Bigio

, et al. Correlation of Alzheimer disease neuropathologic changes with cognitive status: a review of the literature. J Neuropathol Exp Neurol 2012; 71: 362–381.

72.

Buckley

Maruff

Ames

, et al. Subjective memory decline predicts greater rates of clinical progression in preclinical Alzheimer’s disease. Alzheimers Dement 2016; 12: 796–804.

73.

Amariglio

Becker

Carmasin

, et al. Subjective cognitive complaints and amyloid burden in cognitively normal older individuals. Neuropsychologia 2012; 50: 2880–2886.

74.

Konig

Satt

Sorin

, et al. Use of speech analyses within a mobile application for the assessment of cognitive impairment in elderly people. Curr Alzheimer Res 2018; 15: 120–129.

75.

Kavé

Goral

. Word retrieval in picture descriptions produced by individuals with Alzheimer’s disease. J Clin Exp Neuropsychol 2016; 38: 958–966.

76.

Hernández-Domínguez

Ratté

Sierra-Martínez

, et al. Computer-based evaluation of Alzheimer’s disease and mild cognitive impairment patients during a picture description task. Alzheimers Dement (Amst) 2018; 10: 260–268.

77.

Rojas

Kefalianos

Vogel

. How does our voice change as we age? A systematic review and meta-analysis of acoustic and perceptual voice data from healthy adults over 50 years of age. J Speech Lang Hear Res 2020; 63: 533–551.

78.

Schultz

Rojas

St John

, et al. A cross-sectional study of perceptual and acoustic voice characteristics in healthy aging. J Voice 2023; 37: 969.e23–969.e41.

79.

Livingston

Huntley

Sommerlad

, et al. Dementia prevention, intervention, and care: 2020 report of the Lancet Commission. Lancet 2020; 396: 413–446.

80.

Ivanova

Meilán

JJG

Martínez-Sánchez

, et al. Discriminating speech traits of Alzheimer’s disease assessed through a corpus of reading task for Spanish language. Comput Speech Lang 2022; 73: 101341.

81.

Mueller

Van Hulle

Koscik

, et al. Amyloid beta associations with connected speech in cognitively unimpaired adults. Alzheimers Dement (Amst) 2021; 13: e12203.

82.

Verfaillie

SCJ

Witteman

Slot

RER

, et al. High amyloid burden is associated with fewer specific words during spontaneous speech in individuals with subjective cognitive decline. Neuropsychologia 2019; 131: 184–192.

83.

Vogel

Sobanska

Gupta

, et al. Quantitative speech assessment in ataxia—consensus recommendations by the ataxia global initiative working group on digital-motor markers. Cerebellum 2024; 23: 1128–1134.

84.

Gonzalez-Ortiz

Kac

Brum

, et al. Plasma phospho-tau in Alzheimer’s disease: towards diagnostic and therapeutic trial applications. Mol Neurodegener 2023; 18: 18.

85.

Mueller

Koscik

Hermann

, et al. Declines in connected language are associated with very early mild cognitive impairment: results from the Wisconsin Registry for Alzheimer’s Prevention. Front Aging Neurosci 2018; 9: 437.

86.

Xiang

Zhang

, et al. Language dysfunction correlates with cognitive impairments in older adults without dementia mediated by amyloid pathology. Front Neurol 2023; 14: 1051382.

87.

Wang

J-T

Zhi

, et al. Speech digital biomarker combined with fluid biomarkers predict cognitive impairment through machine learning. Alzheimers Res Ther 2025; 17: 229.

88.

Pang

Chen

Dodge

, et al. Bridging language markers and pathology: correlations between digital speech measures and surrogate CSF biomarkers in Alzheimer’s disease. bioRxiv 2025; https://doi.org/10.1101/2025.11.05.686472 [Preprint]. Posted November 06, 2025.

89.

Young

Smith

Karjadi

, et al. Speech patterns during memory recall relates to early tau burden across adulthood. Alzheimers Dement 2024; 20: 2552–2563.

90.

Chen

Mueller

, et al. Articulatory precision from connected speech as a marker of cognitive decline in Alzheimer’s disease risk-enriched cohorts. J Alzheimers Dis 2025; 103: 476–486.

91.

Siddiqui

Snyder

Alber

, et al. Speech, language, p-tau217 and neurofilament light chain in an Alzheimer’s disease at risk population. bioRxiv 2025; https://doi.org/10.64898/2025.12.04.25341021 [Preprint]. Posted December 04, 2025.

92.

Nyongesa

Hogarth

. Artificial intelligence-driven natural language processing for identifying linguistic patterns in Alzheimer’s disease and mild cognitive impairment: a study of lexical, syntactic, and cohesive features of speech through picture description tasks. J Alzheimers Dis 2025; 106: 120–138.

93.

Iacono

Feltis

. Idea density and grammatical complexity as neurocognitive markers. Brain Sci 2025; 15: 1022.

94.

Pompili

Abad

de Matos

, et al. Pragmatic aspects of discourse production for the automatic identification of Alzheimer’s disease. IEEE J Sel Top Signal Process 2020; 14: 261–271.

95.

Burke

Gunstad

Hamrick

. Comparing global and local semantic coherence of spontaneous speech in persons with Alzheimer’s disease and healthy controls. Appl Corpus Linguist 2023; 3: 100064.

96.

Martínez-Nicolás

Martínez-Sánchez

Ivanova

, et al. Reading and lexical–semantic retrieval tasks outperforms single task speech analysis in the screening of mild cognitive impairment and Alzheimer’s disease. Sci Rep 2023; 13: 9728.

97.

Fraser

Meltzer

Rudzicz

. Linguistic features identify Alzheimer’s disease in narrative speech. J Alzheimers Dis 2016; 49: 407–422.

98.

Yancheva

Fraser

Rudzicz

. Using linguistic features longitudinally to predict clinical scores for Alzheimer’s disease and related dementias. In: Proceedings of SLPAT 2015: 6th workshop on speech and language processing for assistive technologies. Dresden, Germany: Association for Computational Linguistics, 2015, pp.134–139.

99.

Orimaye

Wong

JS-M

Golden

, et al. Predicting probable Alzheimer’s disease using linguistic deficits and biomarkers. BMC Bioinformatics 2017; 18: 34.

100.

Thomas

Keselj

Cercone

, et al. Automatic detection and rating of dementia of Alzheimer type through lexical analysis of spontaneous speech. In: IEEE International Conference Mechatronics and Automation, 2005, 2005, pp.1569–1574. Niagara Falls, ON, Canada.

101.

Sharafeldeen

Keowen

Shaffie

. Machine learning approaches for speech-based Alzheimer’s detection: a comprehensive survey. Computers 2025; 14: 36.

102.

Kleiman

Galvin

. High frequency post-pause word choices and task-dependent speech behavior characterize connected speech in individuals with mild cognitive impairment. J Alzheimers Dis 2024; 102: 815–829.

103.

García-Gutiérrez

Alegret

Marquié

, et al. Unveiling the sound of the cognitive status: machine Learning-based speech analysis in the Alzheimer’s disease spectrum. Alzheimers Res Ther 2024; 16: 26.

104.

Ahn

Cho

Kim

, et al. Deep learning of speech data for early detection of Alzheimer’s disease in the elderly. Bioengineering 2023; 10: 1093–1093.

105.

Syed

MSS

Lech

, et al. Tackling the ADRESSO challenge 2021: the MUET-RMIT system for Alzheimer's dementia recognition from spontaneous speech. In: Interspeech 2021. ISCA, 2021, pp.3815–3819.

106.

Luz

Haider

Fromm

, et al. An overview of the ADReSS-M signal processing grand challenge on multilingual Alzheimer’s dementia recognition through spontaneous speech. IEEE Open J Signal Process 2024; 5: 738–749.

107.

Steinhubl

Muse

Topol

. Can mobile health technologies transform health care? JAMA 2013; 310: 2395.

108.

Petti

Baker

Korhonen

. A systematic literature review of automatic Alzheimer’s disease detection from speech and language. J Am Med Inform Assoc 2020; 27: 1784–1797.

109.

Kourtis

Regele

Wright

, et al. Digital biomarkers for Alzheimer’s disease: the mobile/wearable devices opportunity. NPJ Digit Med 2019; 2: 9.

110.

Hampel

O’Bryant

Molinuevo

, et al. Blood-based biomarkers for Alzheimer disease: mapping the road to the clinic. Nat Rev Neurol 2018; 14: 639–652.

111.

Laske

Sohrabi

Frost

, et al. Innovative diagnostic tools for early detection of Alzheimer’s disease. Alzheimers Dement 2014; 11: 561–578.

112.

Awan

Shaikh

Awan

, et al. Smartphone recordings are comparable to “gold standard” recordings for acoustic measurements of voice. J Voice 2023; 39: 1019–1032.

113.

Bucks

Singh

Cuerden

, et al. Analysis of spontaneous, conversational speech in dementia of Alzheimer type: evaluation of an objective technique for analysing lexical performance. Aphasiology 2000; 14: 71–91.

114.

Vogel

Morgan

. Factors affecting the quality of sound recording for speech and voice analysis. Int J Speech Lang Pathol 2009; 11: 431–437.

115.

Chen

Herskovits

. Machine-learning techniques for building a diagnostic model for very mild dementia. Neuroimage 2010; 52: 234–244.

116.

Kurdi

. Automatic diagnosis of Alzheimer’s disease using lexical features extracted from language samples. J Med Artif Intell 2024; 7: 13.

117.

Sluis

Angus

Wiles

, et al. An automated approach to examining pausing in the speech of people with dementia. Am J Alzheimers Dis Other Demen 2020; 35: 1533317520939773.

118.

König

Linz

Baykara

, et al. Screening over speech in unselected populations for clinical trials in AD (PROSPECT-AD): study design and protocol. J Prev Alzheimers Dis 2023; 10: 314–321.

119.

Iulita

Streel

Harrison

. Digital biomarkers: redefining clinical outcomes and the concept of meaningful change. Alzheimers Dement Transl Res Clin Interv 2025; 11: e70114.

120.

Van den Berg

Butterbrod

De Boer

, et al. Amyloid-related changes in fluency in patients with subjective cognitive decline. Alzheimers Dement (Amst) 2025; 17: e70063.

121.

Ivanova

Martínez-Nicolás

Meilán

JJG

. Speech changes in old age: methodological considerations for speech-based discrimination of healthy ageing and Alzheimer’s disease. Int J Lang Commun Disord 2024; 59: 13–37.