Publications
See my Google Scholar profile for more information.
TAPE: Learning General Audio Effect Representations
Phillip Long, Yongyi Zang, Qihui Yang, Christodoulos Benetatos, Junfei Liu, Randal Leistikow, Julian McAuley, Perry R. Cook
arXiv Preprint, https://arxiv.org/abs/2610.40759
Presented TAPE (Transformations of Audio Parameter Embeddings), a system that listens to a processed recording, figures out what effects, like an equalizer, compressor, or reverb, were applied to it, and reapplies that same processing to a different recording, all without ever being told the names of the plugins involved. Trained the system on 68 real studio effects plugins so it could recognize them by ear, then had it translate what it heard into the actual knob settings of its own interpretable, built-in effects chain. Showed it matched reference recordings more closely than prior methods, in a single 0.5-second pass rather than thousands of trial-and-error renders. Furthermore, every listener in blind tests preferred its output, rating it 72.5 out of 100 versus 41-48 for competing approaches. Also built a lightweight version that processes audio in 128-sample chunks 1.76 times faster than real time on a single CPU core.
JazzSAMBA: A Synchronous and Asynchronous Multi-take Band Audio Dataset of Jazz Standards for Live Music Models
Phillip Long, Jacob Nguyen, Jace Hosto, Gage Hosto, Jett Takazawa, Fares Nofal, Sebastian Stade, Nithya Shikarpur, Julian McAuley, Cheng-Zhi Anna Huang, Stephen Brade, Aleksandra Teng Ma
arXiv Preprint, https://arxiv.org/abs/2609.34931
Introduced JazzSAMBA, a studio multitrack dataset of 76 jazz standards played by an eight-musician combo. Each song was recorded twice: once as asynchronous overdubs, and once with the full band live. The release includes preferred and alternate takes, stem audio, mixtures, and MIDI, aligned to timed bar, chord, section, and soloist annotations. Employed those assets in a jazz-combo source-separation baseline and a chart-conditioned accompaniment benchmark.
Benchmarking Language Modeling for Lossless Compression of Full-Fidelity Audio
Phillip Long, Zachary Novack, Chris Donahue
Interspeech 2026, https://arxiv.org/abs/2603.08683
Introduced Trilobyte, a hierarchical byte-level tokenization scheme enabling the first tractable language model compression of 24-bit audio. Benchmarked language model-based lossless compression across diverse domains and bit depths. Illustrated that learned approaches can outperform FLAC on CD-quality audio, though those gains diminish substantially relative to lower bit depths. Conducted extensive ablations characterizing traditional and neural codec baselines, including FLAC, pretrained LLMs, and neural audio codecs as drop-in replacements for FLAC's linear predictive coding.
MuseTok: Symbolic Music Tokenization for Generation and Semantic Understanding
Jingyue Huang, Zachary Novack, Phillip Long, Yupeng Hou, Ke Chen, Taylor Berg-Kirkpatrick, Julian McAuley
IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP) 2026, https://arxiv.org/abs/2510.16273
Proposed a new tokenization method for symbolic music, which employs a residual vector quantized-variational autoencoder (RQ-VAE) on bar-wise music segments within a Transformer-based encoder-decoder framework, producing music codes that achieve high-fidelity music reconstruction and accurate understanding of music theory. Evaluated the semantic understanding of the framework at the note, bar and song levels by training classifiers on MuseTok's encoded-latents for three music information retrieval tasks: melody extraction, chord detection, and emotion recognition.
MusPyExpress: Extending MusPy with Enhanced Expression Text Support
Phillip Long, Hao-Wen Dong, Julian McAuley, Zachary Novack
NeurIPS 2025 Workshop on AI for Music, https://arxiv.org/abs/2608.21678
Created an extension to the popular symbolic music processing library MusPy, which we term MuspyExpress, to better extract expression text, symbolic annotations common in western sheet music. While often omitted by the dominant MIDI standard, expression text is abundant in MusicXML (and MusicXML-adjacent) formats yet often goes unharnessed because current symbolic music processing libraries struggle, or even fail, to parse it. By integrating expression text into the MusPy ecosystem, MusPyExpress enables the extraction of expression text along with symbolic music for downstream modeling.
PDMX: A Large Scale Public Domain MusicXML Dataset for Symbolic Music Processing
Phillip Long, Zachary Novack, Taylor Berg-Kirkpatrick, Julian McAuley
IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP) 2025, https://arxiv.org/abs/2409.10831
Assembled the largest known copyright-free MusicXML dataset, PDMX, consisting of over 250,00 sheet music pieces for modeling symbolic music. Demonstrated the utility of PDMX's deduplication and rating data as a means for data distillation by training decoder-only transformers to generate symbolic music from four different-filtered subsets of PDMX. PDMX also features an abundance of fine-grained performance directives, which could be harnessed in future work as expression-text controls or natural language captions.
The utility of a closed breeding colony of Peromyscus leucopus for dissecting complex traits
Phillip Long, Vanessa J Cook, Arundhati Majumder, Alan G Barbour, Anthony D Long
Genetics, Volume 221, Issue 1, May 2022, iyac026, https://doi.org/10.1093/genetics/iyac026
Used R, Python, and Bash scripts on UC Irvine's HPC3 Cluster to analyze terabytes of genetic data collected in Peromyscus leucopus, the primary reservoir for Lyme disease. Tested a new computational framework for imputing genetic data in closed-breeding colonies. Created figures and tables summarizing datasets for the paper.