How Not to Develop an "umm" Detector Using Create ML
summary
Attempts to build an “umm” detector use Apple speech recognition (which filters fillers), a Core ML classifier trained with Create ML, audio slicing with ffmpeg, waveform augmentation via SoX, and active learning on large corpora. Dataset starts ~736 labeled clips with ~1:11 um:clean imbalance, ums are synthetically increased ~4×, feature extractor switches from Apple’s 5 KB audio feature print to a VGG (4.6 MB), validation peaks ~95.7%, and a live demo fails in a noisy room.
AI-written summary · about
chapters (11)
- 0:00Intro
- 6:11Speech Recognition
- 7:20Apple Speech Recognition
- 8:34Machine Learning
- 14:54Create ML
- 18:40Problems
- 20:28Swift Script
- 23:33Augmented Data
- 27:44More Data
- 29:34Active Learning
- 32:39Why Not
related talks
31mPower-Up Your App With Apple’s Machine Learning Frameworks
25mLooking at Sound
27mMachine Learning Methods in Detecting Alzheimer’s Disease from Speech and Language
45mAuditory Intelligence : beyond speech recognition
40mNavigating the Machine Learning Landscape for Apple Devices
22mGetting Started with Apple ML
18mUsing Machine Learning to Enrich and Analyze Real-time Communications
30mStarting deep learning projects without enough labelled data