HASS: Hierarchical Simulation of Logopenic Aphasic Speech for Scalable PPA Detection

This paper introduces HASS, a novel, clinically grounded hierarchical simulation framework that generates scalable, multi-level logopenic aphasic speech data to overcome data scarcity and improve the accuracy and generalizability of primary progressive aphasia detection models.

The Gist

The Big Problem: The "Empty Library"

Imagine you are trying to teach a robot to recognize a specific type of cough that indicates a rare disease. To do this well, the robot needs to listen to thousands of recordings of people with that cough.

But here's the catch: The people with this disease (Primary Progressive Aphasia, or PPA) are a vulnerable group. It's hard, expensive, and ethically tricky to record enough of them to build a massive library of data. Without enough data, the robot stays "dumb" and can't diagnose the disease accurately.

The Old Solution: The "Glitchy Tape"

Previously, scientists tried to solve this by taking normal, fluent speech and artificially adding "glitches" to it. They would randomly insert pauses, repeat words, or stutter.

The Analogy: Imagine trying to teach someone what a broken car sounds like by taking a perfectly running engine and randomly pressing the "squeak" button on a toy.

• The Flaw: Real brain diseases don't just add random glitches. They break the engine in a specific, logical way. A real PPA patient doesn't just stutter randomly; their brain struggles to find the right word, which causes them to pause, then say the wrong sound, then pause again. The old "glitchy tape" method missed this chain reaction, so the robot learned the wrong patterns.

The New Solution: HASS (The "Digital Twin" Factory)

The authors of this paper built a new system called HASS (Hierarchical Simulation of Logopenic Aphasic Speech). Think of HASS not as a glitch machine, but as a highly realistic "Digital Twin" factory.

Instead of just breaking the speech, HASS simulates the brain of a person with a specific type of PPA (called the logopenic variant). It works in two layers, like a two-step assembly line:

1. Layer 1: The Word Hunt (The Content Level)
   Imagine a person trying to describe a "hearth fire." Their brain knows the concept, but the "word file" is missing.

   • What HASS does: It simulates the struggle to find the word. It makes the speaker say, "The... uh... the place where you burn wood..." instead of just "hearth." It adds "false starts" and "circumlocutions" (talking around the word).
   • The Rule: It only does this on hard words (like "amber" or "hearth"), just like a real human would.

2. Layer 2: The Sound Fumble (The Phoneme Level)
   Once the speaker finally grabs the word "amber," their brain is still tired. They might trip over the sounds.

   • What HASS does: It simulates the mouth tripping. Maybe they say "am-ber" but pause in the middle, or swap the 'm' for an 'n'.
   • The Connection: Crucially, this layer only happens because of the struggle in Layer 1. The errors are linked, just like in real life.

The Result: A "Super-Teacher"

The researchers used this factory to create 4,773 hours of synthetic speech. They created "sick" voices (with the simulated disease) and "healthy" voices (using the same factory but without the disease).

They then trained a new AI doctor (a machine learning model) using this synthetic data.

The Magic Trick:
When they tested this AI doctor on real patients from different hospitals (which it had never seen before), it performed better than doctors trained only on the tiny amount of real data available.

• Why? Because the HASS factory taught the AI the logic of the disease (how the brain breaks down), not just the sound of the disease. It learned the "rules of the game" rather than just memorizing specific recordings.

The Takeaway

Think of HASS as a flight simulator for speech disorders.

• Before, we tried to teach pilots (AI models) to handle engine failure by throwing wrenches at real planes (adding random glitches).
• Now, we have a simulator that perfectly mimics the physics of a failing engine (the HASS framework).
• The pilots trained in the simulator are so good that when they finally get into a real plane, they handle the crisis better than pilots who only practiced on a few real, but limited, flights.

This approach solves the data shortage problem, protects patient privacy (since the data is synthetic), and creates a more reliable tool for diagnosing neurodegenerative diseases.

Technical Summary

1. Problem Statement

Primary Progressive Aphasia (PPA) is a neurodegenerative disorder characterized by progressive language impairment. Automated screening for PPA, particularly the logopenic variant (lvPPA), is hindered by a critical data scarcity bottleneck.

• Limitations of Real Data: Collecting large-scale clinical speech data is difficult due to patient vulnerability, high costs of expert labeling, and strict ethical/privacy constraints. Existing public datasets (e.g., DementiaBank, AphasiaBank) are small and lack cross-institutional diversity.
• Limitations of Existing Synthetic Data: Previous attempts to generate synthetic dysfluent speech rely on injecting isolated errors (e.g., simple pauses or repetitions) into fluent speech. These approaches fail to capture the holistic, multi-level nature of PPA, where semantic, phonological, and temporal deficits interact. Furthermore, many synthetic methods lack grounding in specific clinical phenotypes, causing models trained on them to learn generic dysfluency patterns rather than disorder-specific signatures.

2. Methodology: The HASS Framework

The authors propose HASS (Hierarchical Aphasic Speech Simulation), a clinically grounded, two-layer simulation framework designed to generate severity-controlled synthetic lvPPA speech. The pipeline consists of three main stages:

A. Clinician-Guided Dysfluent Text Generation

The system uses a Large Language Model (LLM, specifically Gemini 3) supervised by Speech-Language Pathologists (SLPs) to simulate pathological behaviors. The generation is constrained by three clinical rules:

1. Lexical Bias: Dysfluencies are non-uniformly applied, heavily targeting high-lexical-demand loci (low-frequency words, multisyllabic targets).
2. Syntactic Adherence: Disruptions occur only at plausible syntactic boundaries (e.g., clause boundaries).
3. Phenotype Exclusion: Features of other PPA variants (e.g., agrammatism typical of non-fluent PPA) are explicitly penalized to prevent "phenotype drift."

The generation is factorized into two layers:

• Layer 1: Lexical Retrieval (Content Level): The LLM introduces word-level errors such as circumlocutions, false starts, and filled pauses while preserving the intended message.
• Layer 2: Phonological Encoding (Phoneme Level): The output text is converted to a word-aligned IPA representation. The LLM then inserts inline markers for six error types based on a clinical hierarchy:
  • Primary Markers: [PAU] (pause), [SUB] (substitution), [DEL] (deletion).
  • Secondary Markers: [REP] (repetition), [PRO] (prolongation), [INS] (insertion).
  • Severity Conditioning: Marker rates and types are conditioned on severity (Mild, Moderate, Severe), with disruption probability increasing with word length and syllable complexity.

B. Speech Synthesis

The marked IPA strings are synthesized into audio using VITS (Variational Inference with adversarial learning for Text-to-Speech).

• Implementation: Phoneme-level markers ([DEL], [SUB], [INS], [REP]) are applied during IPA generation. [PAU] is realized by inserting silence segments, and [PRO] is realized by prolonging the target phoneme during inference.
• Output: The system produces both sentence-level audio and concatenated utterances (with 50ms crossfades).

C. Dataset Construction

• Corpus Size: 4,773 sentence-level clips totaling 12.81 hours of audio.
• Composition: 2,007 control utterances (fluent) and 2,766 dysfluent utterances (871 mild, 1,101 moderate, 794 severe).
• Speaker Diversity: Synthesized using 95 speakers from the VCTK corpus across 40 unique prompts. Controls are generated via the same pipeline without impairment injection to ensure differences are attributable to the simulated disorder, not speaker identity.

3. Key Contributions

1. First Holistic Simulation Framework: HASS is the first framework to model a neurodegenerative language disorder (lvPPA) as a structural, multi-level disease rather than a collection of isolated dysfluencies. It explicitly models the cascade from lexical retrieval failure to phonological errors.
2. Clinically Grounded Pipeline: The simulation is guided by SLPs and adheres to specific clinical rules, ensuring the synthetic data reflects the true "phenotype" of lvPPA.
3. Scalable Data Augmentation: The framework provides a scalable recipe for generating severity-controlled synthetic data, addressing the scarcity of real clinical recordings.
4. Open Release: All data, models, and code are released to support reproducibility.

4. Experimental Results

The authors evaluated the utility of HASS by training a Wav2Vec 2.0 model (fine-tuned with LoRA) on the synthetic data and testing its ability to detect real-world lvPPA.

• Evaluation Protocol: A strict cross-site design was used. Models were trained on one institutional dataset (e.g., Baycrest) and tested on another (e.g., Hopkins/JHU), and vice versa. This tests generalization across different recording protocols and environments.
• Performance Metrics:
  • HASS Model: AUC = 0.892, F1 = 0.800, Recall = 0.899.
  • Baseline (Real Data Only): AUC = 0.850, F1 = 0.778, Recall = 0.659.
• Key Findings:
  • Models trained on HASS-generated data outperformed models trained solely on limited real-world clinical data across all metrics.
  • HASS-trained models demonstrated superior stability (tighter variance across folds) and robust cross-site generalization.
  • The synthetic data successfully captured the core clinical phenotype of lvPPA without overfitting to the acoustic artifacts or demographic biases of a single clinical site.

5. Significance and Impact

• Clinical Utility: HASS offers a viable solution to the data scarcity problem in neurodegenerative disease research, enabling the training of more robust and generalizable diagnostic AI models.
• Privacy Preservation: By generating synthetic data that mimics real patients, the framework allows for data augmentation without compromising patient privacy.
• Future Directions: While the framework is highly effective, the authors note limitations regarding standard TTS architectures (like VITS) struggling with severe phonological errors and the inherent complexity of individual symptom progression. Future work may focus on refining the synthesis of extreme severity cases.

In conclusion, HASS represents a paradigm shift from simple dysfluency injection to hierarchical, clinically grounded simulation, proving that high-quality synthetic data can not only match but exceed the performance of models trained on scarce real-world clinical data for PPA detection.