Voice Intent — Assistive Communication for Autism
When I first joined a volunteer program supporting people with autism, my task seemed simple: play games, read picture books, keep company. One afternoon, I played a strategy board game with a young man who worked at a autistism-friendly bubble tea shop. He smiled politely but soon froze, staring at the maze of rules. Watching him struggle, I realized it wasn’t only the game that confused him. It was the unspoken rules that often govern everyday talk.
Many individuals with autism can understand words by their literal meaning but find it hard to read tones, ironics, or implications. That gap between what is said and what is meant fascinated me. I wanted to bridge it, not through therapy, but through language itself. So I taught myself Flutter, an app development framework, and built Voice Intent, an app powered by GPT-4’s API. Users can type or record puzzling sentences, and the app explains whether they are sarcastic, figurative, or sincere.
When Voice Intent was released on the App Store and Google Play, it became more than a tool but a statement that communication should be accessible to everyone. It also taught me that leadership isn’t always about “organizing people” or speaking loudly. Sometimes it begins with noticing a silent problem and daring to fix it.
Home Screen
Designed for individuals on the autism spectrum — four ways to explore the intent behind communication.
Pragmatic Language
Speak or type — Voice Intent offers real-time insights on appropriateness, emotional expressiveness, and how to respond.
Emotions in Speech
Record speech and detect nuanced emotional signals using state-of-the-art affective AI.
Literal vs Figurative
Is it sarcastic? Literal? Figurative? Voice Intent helps users make sense of ambiguous language.
Rip Current Detection Drone
Abstract
Rip currents pose significant hazards to beachgoers, often catching untrained individuals unaware and leading to dangerous situations. The research aims to develop an effective method for detecting rip currents using artificial intelligence, thereby enhancing beach safety.
The method involves training a binary classification model on aerial images of the ocean, categorizing them into those with rip currents and without rip currents. Data augmentations such as cutting, rotating, and flipping the image are employed to enhance the model’s generalization ability.
We utilized Multi-Layer Perceptron classifiers, achieving an accuracy of 73.9% initially on the testing set, and 92.5% after applying a confidence threshold.
The integration of the model with a Tello Talent Robomaster TT drone introduced significant flexibility and efficiency, with an average processing time of 3.09 milliseconds per 1000 images.
With a True Positive Rate (TPR) of 93.06% and True Negative Rate (TNR) of 91.10%, the model demonstrated high performance. This innovative integration of AI with rip current detection is flexible, efficient, and effective, making a meaningful contribution to beach safety.