SANE 2026 poster
Discrete Parameter Estimation for DDSP Source Separation
Poster presented with Andrea Cogliati at SANE 2026 (Speech and Audio in the Northeast) at MIT. It studies how to fit the integer loop delay and pluck position of a Karplus–Strong string model, where gradient descent breaks down, comparing exhaustive search, pitch estimation, continuous relaxations, and learned categorical estimators (REINFORCE, Gumbel-Softmax) on mixtures of up to seven strings.
Audio information research
Self-Supervised Representation Learning for Audio Information Research
This presentation covers the development of self-supervised representation learning for audio information research, including WaveNet, wav2vec, VQ-wav2vec, wav2vec 2.0, HuBERT, and WavLM.