The Apple Watch haptics turned out to be perfect for this... subtle enough that only you notice, but enough to catch yourself before you've been rambling at 200 WPM for five minutes.
The journey was messier than I expected:
Started by trying to run speech recognition entirely on-device. Trained various models, got to about 85% accuracy on clean English audio. Then I tested it on actual meeting recordings... turns out clean training data and real-world messy audio are very different problems. Accents, background noise, crosstalk... the accuracy dropped hard.
Tried throwing bigger models at it. Left my MacBook Air running for 3 days straight training a larger model. Result? 1% improvement. That's when I realised I needed to rethink the approach entirely.
The final version uses a different method that works reliably across different speaking styles and environments. Still does the core job... monitors your pace and taps you when you're speeding up or slowing down too much.
What I learned:
- On-device ML sounds great until you need it to work on messy real-world data - Bigger models aren't always the answer - Sometimes the unsexy solution is the right one - Apple Watch haptics are genuinely underrated as a feedback mechanism
Would love feedback from anyone who's tackled similar on-device ML challenges or built accessibility/feedback tools.
App Store: https://apps.apple.com/gb/app/pacecoach/id6758337168 Website: https://pacecoach.co.uk