Community Driven Data Collection and Consent in AI
summary
Common Voice builds an open source speech corpus released under CC0 donation-led approach, now covering 118 languages and reaching 125 next month. The claim is concrete: crowdsourced voice donations power models rather than scraped web text. It shows speech assistants still work reliably for roughly 20 languages while volunteers expand coverage. Alternatives discussed include paid licensing, governance via data trusts, and explicit licensing instead of stealthy scraping by large model builders.
AI-written summary · about
chapters (10)
- 0:00Consent in AI data
- 2:51AI hype cycle problems
- 11:07AI training data sources
- 16:17Common Voice data collection
- 18:26Wikipedia licensing approach
- 20:16Data trusts and consent
- 22:01Pay contributors instead of scraping
- 23:06Ethical AI data needs
- 24:31Community enforcement against theft
- 25:12Hope for AI’s future
related talks
49mPanel Discussion: Navigating the Ethical Landscape: Responsible...
15mMore empathy and health data protection in AI
21mDemocratizing AI with Open Source Multi-Agent Systems
37mThe Shadows That Follow AI Generative Models
18mV.O.I.C.E.: A Data-Driven Risk Taxonomy for Synthetic Voice Generation
15mEnterprise-Scale Privacy for AI: How Canva Scaled Customer Control of Data for AI...
39mSynthetic Trust: Exploiting Biases at Scale
24mCommanderSong: A Systematic Approach for Practical Adversarial Voice Recognition