We explore models for translating abstract musical ideas (scores, rhythms) into expressive performances using seq2seq and recurrent variational information bottleneck (VIB) models. Though seq2seq models usually require painstakingly aligned corpora, we show that it is possible to adapt an approach from the Generative Adversarial Network (GAN) literature (e.g. Pix2Pix, Vid2Vid) to sequences, creating large volumes of paired data by performing simple transformations and training generative models to plausibly invert these transformations. Music, and drumming in particular, provides a strong test case for this approach because many common transformations (quantization, removing voices) have clear semantics, and learning to invert them has real-world applications. Focusing on the case of drum set players, we create and release a new dataset for this purpose, containing over 13 hours of recordings by professional drummers aligned with fine-grained timing and dynamics information. We also explore some of the creative potential of these models, demonstrating improvements on state-of-the-art methods for Humanization (instantiating a performance from a musical score).

We investigate a speculative future in which we celebrate happiness by capturing laughter and representing it in tangible forms. We explored technologies for capturing naturally occurring laughter as well as various physical representations of it. For several weeks, our participants collected audio samples of everyday conversations with their loved ones. We processed those samples through a machine learning algorithm and shared the resulting tangible representations (e.g., physical containers and edible displays) with our participants. In collecting, listening to, interacting with, and sharing their laughter with loved ones, participants described both joy in preserving and interacting with laughter and tension in collecting it. This study revealed that the tangibility of laughter representations matters, especially its symbolism and material quality. We discuss design implications of giving permanent forms to laughter and consider the sound of laughter as a part of our personal past that we might seek to preserve and reflect upon.

This work examines the rhetorical techniques that speakers employ during political campaigns.We introduce a new corpus of speeches from campaign events in the months leading up to the 2016 U.S. presidential election and develop new models for predicting moments of audience applause.In contrast to existing datasets, we tackle the challenge of working with transcripts that derive from uncorrected closed captioning, using associated audio recordings to automatically extract and align labels for instances of audience applause.In prediction experiments, we find that lexical features carry the most information, but that a variety of features are predictive, including prosody, long-term contextual dependencies, and theoretically motivated features designed to capture rhetorical techniques.