To help make questions answered during the Governance & Algorithmic Design in Speech AI Mozfest session more searchable and allow folks watching asynchronously to participate, we’re experimenting with running Q&A…right here!
Couple of questions that I had listening to the other excellent speakers:
For Caroline - what are the characteristics of bias that you think are most often overlooked, that have the biggest impact on the data value of a dataset?
For Britone - I love the idea of the writing competitions! Did you find this gave you a better lexical or phonetic coverage of text prompts?
For Wei - the ecosystem diagram really resonated! So useful. What sort of metrics or measures would you use for assessing the health of the ecosystem or for targeting actions?
Thanks @kathyreid for your question! A big one that comes to mind for voice AI, which you touched on in your presentation, is sociolinguistic variation in speech, especially commonly spoken sociolects, and interaction effect - when two or more sociolinguistic factors collide and result in variations of speech that is more than just the combination of the individual factors. One well-known example is the tendency for an interaction between class and gender identity whereby women, especially working-class women, style-shift to avoid stigmatized “working-class” linguistic forms more than men do.