“Play PRBLMS”: Identifying and Correcting Less Accessible Content in Voice Interfaces
summary
A prototype Spotify voice interface on mobile exposes tracks that are popular by streams but rarely found by voice queries; anomaly detection divides streams by voice-finds over the top 5,000 tracks, log-normalizes counts, and flags outliers (about 80% precision). CrowdFlower collects pronunciations, transcriptions create aliases that map common ASR outputs to correct tracks; the process defines 12 disadvantage classes, shows hip hop and country are overrepresented, and yields per-track gains (one example ~2500% improvement after aliasing).
AI-written summary · about
chapters (4)
- 0:00Start
- 7:26Expressive and Alternative Spellings
- 9:34Overview
- 15:47How Did You Deal with Anglicized Versions of Titles or Names and Other Languages
related talks
27mEnsuring Black voices matter: Why your voice assistant is racist, and what we can do about it
46mAccessibility vs latest Web APIs. Can’t we just get along?
23mDesigning for Accessibility
29mAccessibility vs latest Web APIs. Can’t we just get along?
21mMaking Accessibility More Accessible for Your Organization
30mAI and Accessibility: the Good, the Bad, and the Bollocks
12mMachine learning for Accessibility
40mBuilding a considerate app