Daniel Dopp is a digital audio producer at ESPN, where he has worked for four years producing several different podcasts including Fantasy Focus Football, The Bill Barnwell Show, and Adam Schefter’s Know Them From Adam. He does everything from generating story ideas, to coordinating and booking guests, to mixing and publishing final cuts. In order to promote the content from the finished shows, he takes short, juicy clips and sends them — along with a transcript generated by Pop Up Archive — to the ESPN editorial team for use in other content such as TV shows or online stories.
Earlier this year, Pop Up worked with Tanya Clement and Steve McLaughlin of the UT-Austin School of Information on a massive effort to use machine learning to identify notable speakers’ voices (for example, Martin Luther King, Jr.) from within the American Archive of Public Broadcasting’s 70,000 digitized audio and video recordings. Now, Tanya and Steve are sharing DIY techniques for using free machine learning algorithms to help label speakers in “unheard” audio.
This is a huge and hugely important effort: a model that can identify a single speaker’s voice has vast potential implications for the ability to see inside audio, making the content to accessible to researchers, organizations, and the general public. That the toolkit they’re sharing is DIY means its appropriate for use by programming novices who may be working on their first audio project. Continue reading