Qingyang (Tom) Xi

Machine Learning Researcher | Time-Series, Signal Processing

Profile

Machine learning researcher with PhD training at NYU MARL, specializing in time-series modeling and segmentation, signal processing, representation learning, and evaluation for noisy temporal data.

Built methods, datasets, and validation tools for temporal segmentation in audio ML, with experience spanning deep learning, uncertainty calibration, and ML research tooling in Python.

Motivated to apply robust time-series ML and signal-processing methods to wearable sensing and health-focused products.

Publications

ISMIR 2025 Lose the Frames: Event-based Metrics for Efficient Music Structure Analysis Evaluations

• • Qingyang Xi, Brian McFee
• Introduced event-based evaluation methods that improved efficiency, accuracy, and reproducibility for music structure analysis.

ISMIR (LBD) 2024 Zero-Shot Structure Labeling with Audio and Language Model Embeddings

• • Morgan Buisson, Christopher Ick, Qingyang Xi, Brian McFee
• Applied audio-language representation learning to zero-shot labeling for temporal segmentation.

ISMIR (LBD) 2021 Beyond Hard Decisions: Accounting for Uncertainty in Deep MIR Models​

• • Qingyang Xi, Brian McFee
• Studied probability calibration for uncertainty quantification in deep music information retrieval models.

ISMIR 2018 Guitarset: A Dataset for Guitar Transcription

• • Qingyang Xi, Rachel M. Bittner, Johan Pauwels, Xuzhou Ye, and Juan Pablo Bello
• Designed, recorded, and annotated GuitarSet, a widely used dataset for guitar transcription.
• Developed machine-assisted annotation and human-in-the-loop correction and validation workflows.

Experiences

2017-Present Researcher, Music and Audio Research Lab, NYU, New York, NY

• Developed ML methods for structured prediction on temporal signals, including segmentation, representation learning, uncertainty modeling, and evaluation.
• Built Python-based experimentation, benchmarking, and validation pipelines for large-scale comparison of models and annotations.
• Developed efficient evaluation methods (frameless_eval) for temporal segmentation, improving reproducibility and reducing computational cost.
• Led dataset design, annotation workflows, and human-in-the-loop validation for GuitarSet, a widely used dataset for guitar transcription.
• Contributed to open-source Python tools for audio and MIR research, including librosa, mirdata, mir_eval, and JAMS.

2016-2017 Adjunct Lecturer, New York University, New York, NY

• Developed and taught course material in digital and analog electronics for audio engineers.

2014-2015 Software Developer, SynthWorks, Brookline, MA

• Developed software tools and plugins for real-time audio workflows, translating user and production needs into maintainable software.
• Built performance-sensitive solutions for live systems where timing, reliability, and usability were critical.

Education

2018-2026 Ph.D., Music Technology, New York University

2015-2018 M.M., Music Technology, New York University

2010-2014 B.M., Piano Performance and Theory/Composition, Boston University

Skills

• Python, PyTorch, NumPy, pandas, SciPy, scikit-learn, Weights & Biases
• Time-series modeling, signal processing, representation learning
• Benchmarking, uncertainty calibration, dataset curation and annotation
• Git, slurm/HPC, singularity/docker

Service

• Reviewer for TASLP, ICASSP, ISMIR, and TISMIR.