Computational Techniques for Making Karaoke Harder

Robot Karaoke is a live comedy show in which the lyrics of well-known songs are swapped for fragments of text drawn from an unusual catalog of datasets. Jamie Brew explains where that material comes from: copied pages of reviews, competition data published by companies, forum-style questions, government tax forms, wiki articles, a collection of early internet banner ads, and submissions gathered from a community maker space. Each candidate fragment must match the syllable count of the line it replaces, which is worked out with the CMU pronouncing dictionary and extended with manual overrides for missing words.

The middle of the talk traces how the software evolved, from a terminal prototype that needed a keypress for every line, to a browser version with syllable-level timing, a persistent header, and a queue of requested songs. Brew described how the show runs at a regular bar venue and how audience requests come in through a QR code. Several demos followed, including a Blink-182 song rewritten with lines from a federal tax form and a Total Eclipse of the Heart parody built from a hacker film script, with audience members singing some parts.

In the question period, Brew said phrase selection mostly follows existing sentence breaks, with noun-phrase extraction used when sentences are too long to fit a line. Licensing questions led to a plan to open-source the data standard and datasets while keeping song lyrics off public hosting. The takeaway is that noticing and collecting everyday text can be enough to build a usable dataset.