LAPE
A large language model-based paralinguistic enhancement framework for automatic Alzheimer's disease detection from speech. It incorporates three techniques—prosodic event textualization, lexical-prosodic unit formation and chunking, and text-anchored paralinguistic fusion—to jointly model acoustic features such as pauses and word elongation with linguistic content.
Timeline (last 90 days)
A paper published on arXiv proposed a framework called LAPE (LLM-Anchored Paralinguistic Enrichment) for Alzheimer's disease speech detection.
The core concept of LAPE is to use linguistic representations extracted by a large language model as an anchor, and then integrate paralinguistic cues into them.
The LAPE framework contains three synergistic innovations: prosodic event textualization, lexical-prosodic unitization and chunking, and text-anchored paralinguistic fusion.
Prosodic event textualization encodes prosodic events such as pauses and word lengthening as explicit tokens, and uses a duration-aware bounded repetition mechanism to represent event intensity.
The lexical-prosodic unitization and chunking method performs pooling only on consecutive word units to preserve event identity and magnitude in both the text and speech modalities.
Text-anchored paralinguistic fusion integrates local-level and discourse-level speech features and introduces the NormGate dynamic modulation mechanism.
The research team evaluated LAPE on two standard datasets, ADReSS and ADReSSo.
The evaluation used participant-level cross-validation and a leave-one-subject-out method to prevent data leakage.
LAPE achieved state-of-the-art (SOTA) performance across all four main evaluation settings.
The research team stated that LAPE's code will be publicly released after the paper is accepted.
All Facts (10)
The research team stated that LAPE's code will be publicly released after the paper is accepted.
50%UnverifiedThe core concept of LAPE is to use linguistic representations extracted by a large language model as an anchor, and then integrate paralinguistic cues into them.
50%UnverifiedThe LAPE framework contains three synergistic innovations: prosodic event textualization, lexical-prosodic unitization and chunking, and text-anchored paralinguistic fusion.
50%UnverifiedProsodic event textualization encodes prosodic events such as pauses and word lengthening as explicit tokens, and uses a duration-aware bounded repetition mechanism to represent event intensity.
50%UnverifiedThe lexical-prosodic unitization and chunking method performs pooling only on consecutive word units to preserve event identity and magnitude in both the text and speech modalities.
50%UnverifiedText-anchored paralinguistic fusion integrates local-level and discourse-level speech features and introduces the NormGate dynamic modulation mechanism.
50%UnverifiedThe research team evaluated LAPE on two standard datasets, ADReSS and ADReSSo.
50%UnverifiedThe evaluation used participant-level cross-validation and a leave-one-subject-out method to prevent data leakage.
50%UnverifiedLAPE achieved state-of-the-art (SOTA) performance across all four main evaluation settings.
50%UnverifiedA paper published on arXiv proposed a framework called LAPE (LLM-Anchored Paralinguistic Enrichment) for Alzheimer's disease speech detection.
50%