If you’ve ever watched a child struggle to find the right word for how they’re feeling — or watched an AAC (augmentative and alternative communication) device deliver a flat, robotic “I’m sad” that doesn’t quite match the moment — you already understand the problem this new study is trying to solve.

A team of researchers led by Kevin Pitt at the University of Nebraska–Lincoln recently published a preliminary study exploring whether artificial intelligence can help AAC systems recognize a child’s emotions from their face — and eventually use that information to make communication feel more natural and personal. The study, published in the American Journal of Speech-Language Pathology, is one of the first of its kind to include not only neurotypical children but also children with autism, ages 8–13.

Why This Matters

Many AAC systems are excellent at helping a child construct a sentence, but far less equipped to convey how that sentence should sound. A child might type “I’m okay,” but their AAC device has no way of knowing whether that “okay” should sound calm, frustrated, or forced. Over time, this mismatch between message and tone can leave AAC users feeling like their device doesn’t quite “sound like they do in their heads.” Research has shown that authentic emotional expression is something AAC users themselves identify as central to feeling truly heard.

Emotion communication is also foundational for building relationships and for getting support — whether that’s a child telling a friend they’re upset, or a teen describing their emotional state to a counselor or doctor. Without an easy way to express feelings, AAC users can find themselves cut off from an important part of everyday connection.

What the Researchers Did

The team built a prototype application using two tools: DeepFace (a widely used, open-source facial emotion recognition system) and OpenCV (a video-processing library). The app could detect five basic emotions — happiness, sadness, anger, surprise, and calm/neutral — chosen because they tend to emerge earliest in childhood.

Twelve neurotypical children and four children with autism took part. Each child went through a short “calibration” process — similar to how some eye-gaze AAC systems are calibrated — where the system learned that particular child’s unique facial expressions. Children then completed two tasks:

1. Imitated (copy) task: Children copied facial expressions shown in photos.
2. Semispontaneous task: Children were given a short prompt (e.g., “Imagine you lost your favorite hoodie — how would your face show you’re a little sad?”) and asked to react naturally.

Afterward, kids rated how much they enjoyed the experience and how satisfied they were with the system’s speed and accuracy.

What They Found

The results were cautiously encouraging:

• Overall accuracy was around 80–87%, on average, across both tasks and both groups of children — above the general 70% benchmark researchers often use to judge whether new assistive technology shows real promise.
• Happiness, sadness, and calm/neutral were recognized most reliably. These emotions tend to come with clear, stable facial cues (like a relaxed or smiling face), making them easier for the system to pick up.
• Anger and surprise were harder for the system to detect consistently, likely because they rely on more subtle or variable facial movements — something that may be especially true for children.
• Children with autism performed comparably to neurotypical children overall, though the very small sample (just four children) means this finding should be treated as preliminary rather than conclusive.
• Kids generally enjoyed using the system. Average enjoyment ratings were high (around 4.2–4.3 out of 5) for both groups, suggesting the novelty and “game-like” quality of the task was motivating rather than frustrating — even when the system got things wrong sometimes.
• Satisfaction with the system’s accuracy was rated a bit more modestly, especially among children with autism, hinting that expectations and experiences varied quite a bit from child to child.

What This Could Mean for Families and Clinicians

This is early-stage, proof-of-concept research — not a finished product — but it points toward a few meaningful takeaways:

• Personalization matters. The study’s calibration step, which tailored the AI model to each child’s individual facial expressions, echoes a principle SLPs already know well from other assistive technologies (like eye-gaze systems): one-size-fits-all rarely works for AAC users.
• User control is essential. The prototype let users turn emotion detection on or off and manually correct a wrong guess. The researchers were intentional about this — emotion data is deeply personal, and no child (or adult) should have their feelings “broadcast” without consent.
• This is a starting point, not a solution. The system was tested in a quiet lab setting with children who were not everyday AAC users. Real-world performance — with lighting changes, movement, glasses, or more subtle expressions — is still an open question. The authors themselves are careful to frame DeepFace as a feasibility test, not a clinical tool ready for classrooms or therapy rooms.
• Future systems may combine multiple cues. The researchers suggest that pairing facial analysis with voice tone or other signals (a “multimodal” approach) could improve accuracy and better reflect the nuanced, sometimes-mixed emotions real people experience.

The Bottom Line

For families and clinicians who’ve long wished AAC devices could better capture the emotional layer of communication, this study offers a hopeful, if early, glimpse of what’s possible. It’s not about replacing a child’s own voice or judgment — the researchers were explicit that any future system should let users stay in control of what gets shared and when. Instead, it’s about giving AAC systems one more tool to help a child’s device sound a little more like they do in their head.

As with most emerging AI applications in speech-language pathology, the message here is measured optimism: promising early results, a clear need for more research with larger and more diverse groups of AAC users, and a strong emphasis on keeping the user’s autonomy and privacy at the center of the design.

Source

Pitt, K. M., Ousley, C., Gibbons, C., Martinez, E., Kirkpatrick, C. E., & Bubak, A. (2026). Artificial intelligence for enhancing emotion expression in augmentative and alternative communication systems: A preliminary study including children both with and without autism. American Journal of Speech-Language Pathology, 35, 1713–1729. https://doi.org/10.1044/2026_AJSLP-25-00469