Ensuring Black voices matter: Why your voice assistant is racist, and what we can do about it
summary
Claim: automatic speech recognition shows word error rate almost double for Black speakers (Stanford analysis), and cloud text‑to‑speech APIs offer no recognisably African‑American or Indigenous voices. Concrete wins: Google supports Kiswahili while Amazon and others omit many African languages; Mozilla Common Voice plus open tools Elpis and Persephone target the transcription bottleneck. Proposals include a metadata schema for corpora and targeted data collection to fix the measurement and training gaps.
AI-written summary · about
chapters (17)
- 0:00Intro
- 0:37Introduction
- 4:36Voice Assistant Stack
- 6:00Languages
- 6:36Google vs Amazon
- 6:52Language support
- 8:31Human development index
- 9:09Financial affluence vs language
- 10:28Exclusion at a language level
- 11:14Machine learning
- 13:03Digital divide
- 14:02Low resource languages
- 15:41Voice synthesis
- 17:12Leverage points
- 18:28Metadata
- 19:40Arty
- 22:34Questions
related talks
42mLet's Talk about AI, ML and Bias
45mMore voice, less choice: The rise of voice interfaces and the decline of open source voice
37mThe AI 'Black Box'
28mBreaking the Chains of Oppressive Software
46mArtificial intelligence to the rescue of accessibility
41mGhost in the Machine: The Unexpected Results of Not Understanding Your Users
17m“Play PRBLMS”: Identifying and Correcting Less Accessible Content in Voice Interfaces
35mHaving a Voice on Android: How hard can it be?