Group Project
UmSo
A native iOS app that listens in real time and highlights your filler words and gives practice drills to help you speak more clearly and confidently.
Project
Group
Platform
iOS
Tech Stack
Swift, SwiftUI, CoreData, Combine, Speech Framework, SpeechAnalyzer, AVFoundation
Role
iOS Developer
What the app does
- Live Filler Word Detection
Listens to speech in real time and detects common filler words using Apple’s SpeechAnalyzer and lightweight heuristics.
- Live Feedback Overlay
Displays real-time filler word counts, the most recent filler word, and a confidence indicator.
- Session Recording & Review
Records sessions and stores timestamps of filler occurrences for later playback and review.
- Privacy-First Export
Allows users to export annotated transcripts and delete all recordings locally, with all processing done on-device.
Key takeaways
- Real-time speech tools must balance helpfulness and non-intrusive design; subtle signals and short drills work better than harsh alerts.
- Combining SpeechAnalyzer with lightweight heuristics created a fast, reliable, fully on-device detection pipeline without heavy machine learning models.
- Small, frequent practice through micro-drills leads to better habit change than long, infrequent sessions.
Tech stack
Links
Long story
UmSo started in a very different place than most of my projects. It was built during the final challenge of the Apple Developer Academy, which made it feel like both a capstone and a goodbye project at the same time. Because of that, we wanted to build something that wasn’t just “cool,” but genuinely useful in the real world.
For the first three weeks of the challenge, my team and I didn’t build anything at all. We were stuck in long brainstorming loops, exploring ideas around communication and self-improvement. Everything felt either too big, too boring, or already overdone. It was frustrating, because time kept moving and we didn’t want our final Academy project to feel rushed or shallow.
The turning point came from a session with Bali Toastmasters. We joined one of their practice sessions and observed people training their public speaking skills. One exercise stuck with us: speakers were trained to become hyper-aware of their filler words like “um,” “uh,” “like,” and “you know.” The feedback was instant and visibly effective.
We dug into research and found that speakers who receive immediate feedback use up to 60% fewer filler words than those who don’t. That insight became the core of UmSo. Instead of building a generic speech improvement app, we focused everything around real-time awareness.
Initially, I planned to build a custom machine learning model for filler word detection. Because of time constraints, I decided to use Apple’s SpeechAnalyzer instead. What started as a compromise became a major advantage: the app stayed lightweight, fast, and efficient compared to what it would have been with a heavy CoreML model.
There was heavy pressure to ship the app in time for the end-of-challenge exhibition. That meant real App Store deadlines, real review processes, and no room for last-minute fixes.
We managed to publish UmSo to the App Store just in time for the exhibition, and visitors were able to download and try the app directly at our booth. Seeing people speak into their phones and instantly react to the live feedback was one of the most satisfying moments of my time at the Academy.
More than anything, UmSo represents persistence to me — weeks of uncertainty, a late pivot, technical trade-offs, and still shipping a real, working product under pressure.