CapisciAMe CNN outperforms familiar listeners on impaired speech

AI system

System: CapisciAMe speaker-dependent ASR (compact convolutional neural network, cnn-trad-fpool3) in a voice-input voice-output communication aid
Model: cnn-trad-fpool3.
Release: Exact release not established.
Application developer: University of Messina and IRCCS Centro Neurolesi Bonino-Pulejo research team (CapisciAMe app developed in Italy as a research tool)
Role: The personalized network recognized spoken words for a communication aid.
Configuration: The speech-recognition model was trained on 1,120 patient utterances.
Inputs and outputs: Not assessed
Autonomy: Not assessed
Workflow: Speaker-dependent training on hundreds of the patient's own utterances. Closed, small target vocabulary. Multiple prompting modalities during recording.
Human oversight: Not assessed
Relation to AI: The personalized network is the entire intervention. Its 72.65% word accuracy exceeded the 56.75% mean of familiar listeners. The comparison used a 13-word closed set. The authors state the single-case design and limited vocabulary constrain generalization.
AI contribution: The personalized network is the entire intervention. Its 72.65% word accuracy exceeded the 56.75% mean of familiar listeners. The comparison used a 13-word closed set. The authors state the single-case design and limited vocabulary constrain generalization.

Outcome identifiers: Not assessed
Benefit: recognized 72.65% of her words versus 56.75% for 12 familiar professionals
Category: Improved Task Accuracy
Status: Not assessed
Beneficiary groups: People. Society.
Beneficiaries: Not assessed
Reach: Not assessed
Distribution and access: Not assessed
Duration: Not assessed
Expected?: Yes
Costs and harms: Not assessed

Studies: Not assessed
Study design: Not assessed
Measurement: Not assessed
Setting and sample: Not assessed
Measures: recognized 72.65% of her words versus 56.75% for 12 familiar professionals
Comparison: 12 familiar rehabilitation professionals listening to the same utterances
Results: Reference: mean = 56.75 %, SD = 12.91. Result: 72.65 % accuracy.
Additional value from AI: Not assessed
Limitations: Not assessed
Denominators: Not assessed
Materials for replication: Not assessed

Participants: Developer: University of Messina and IRCCS Centro Neurolesi Bonino-Pulejo research team (CapisciAMe app developed in Italy as a research tool). Operator: University of Messina (therapy-room sessions).
Jurisdictions: Benefit: Messina, Italy. Developer: Italy. Operator: Italy.
Ownership and control: Not assessed
Dependencies: Not assessed
Cross-border connections: Not assessed
Geopolitical implications: Not assessed

Conditions for replication: Patient can produce consistent vocalizations for each target word. Quiet recording environment with a headset microphone. Clinician time to collect and label training data.
Replication attempts: Not assessed
Later findings: No later development found.
Availability: Not assessed
Institutional decision: Not assessed
Grounds: Not assessed
Action and result: Not assessed
Success criterion: Not assessed

Assessed benefits: People. Society.
Depth: Not assessed
Reasons: People: The result is a recognition-accuracy capability measured in test sessions. The paper documents no use of the system for the patient's real communication. Society: Single-case proof of concept with a 13-word vocabulary.
Observed or intended: Not assessed
Rubric: Not assessed
Assessor and date: Not assessed
Alternative assessments: Not assessed
Fill: The makers' claim

Kind

Not assessed

Related entries

Both decode speech for people with impaired speech; Whisper fabricated words and violent phrases when transcribing impaired speech.

Field

Speech-language rehabilitation / augmentative communication. Adults with post-stroke aphasia and dysarthria.

Purpose

Recognize the impaired speech of a person with aphasia and dysarthria so a communication aid can voice her intended words

Dates

Date basis unassessed: 2022 (precision unassessed). Updated: 22 October 2026.

Stage

Not assessed

Version

SHA-256: c7c5373cc759…