# Matteo Spanio

> AI research engineer working on audio, music and multimodal generative models — from cultural-heritage preservation to symbolic music and LLMs.

AI Research Engineer and PhD candidate in Brain, Mind and Computer Science at the University of Padua, based at the Centro di Sonologia Computazionale (Department of Information Engineering). ORCID 0000-0002-2436-7208.

Research areas:
- AI for audio cultural heritage preservation
- Multimodal and cross-modal generative AI
- Symbolic music and large language models
- Audio digital signal processing
- Music information retrieval

## Selected publications

- [A multimodal symphony: integrating taste and sound through generative AI](https://matteospanio.github.io/publications/#spanio_frontiers_2025): Frontiers in Computer Science 2025.
- [A novel derivative-based approach for the automatic detection of time-reversed audio in the MPAI/IEEE-CAE ARP international standard](https://matteospanio.github.io/publications/#bosi2024a): Proceedings of the 157th Audio Engineering Society Convention (AES) 2024.
- [Enhancing Preservation and Restoration of Open Reel Audio Tapes Through Computer Vision](https://matteospanio.github.io/publications/#10.1007/978-3-031-51026-7_26): Image Analysis and Processing - ICIAP 2023 Workshops 2024.

## Recent publications

- [A Quantized Native Runtime for On-Device Semantic Audio Generation](https://matteospanio.github.io/publications/#spanio2026quantizednativeruntimeondevice): Proceedings of the IEEE International Symposium on the Internet of Sounds (IS2) 2026.
- [BMdataset: A Musicologically Curated LilyPond Dataset](https://matteospanio.github.io/publications/#spanio2026bmdataset): Proceedings of the 23rd Sound and Music Computing Conference (SMC) 2026.
- [Can LLMs understand LilyPond? A benchmark for symbolic music generation and understanding](https://matteospanio.github.io/publications/#spanio2026lilypond): Proceedings of the Italian Conference on Artificial Intelligence (Ital-IA 2026) 2026.
- [Cross-cultural evaluation of taste-sound correspondences in AI-generated music](https://matteospanio.github.io/publications/#spanio2026): PLOS ONE 2026.
- [Does Notation Matter? MusicXML Tokenisation for Symbolic Music Generation and Understanding](https://matteospanio.github.io/publications/#poltronieri2026notation): Proceedings of the 27th International Society for Music Information Retrieval Conference (ISMIR) 2026.
- [Fluxion: A Differentiable DSP Runtime from Training GPU to Real-Time Edge](https://matteospanio.github.io/publications/#spanio2026fluxion): Proceedings of the IEEE International Symposium on the Internet of Sounds (IS2) 2026.
- [Food and Flavor Representation Learning: A Systematic Mapping Study and Research Agenda](https://matteospanio.github.io/publications/#spanio2026review): Artificial Intelligence Review 2026.
- [Multimodal Dataset Normalization and Perceptual Validation for Music-Taste Correspondences](https://matteospanio.github.io/publications/#spanio2026multimodal): arXiv preprint arXiv:2604.10632 2026.

## Projects

- [TorchFX](https://matteospanio.github.io/projects/torchfx/): gpu-accelerated audio dsp, filters as torch modules. Python, PyTorch, CUDA, C++.
- [lilyBERT](https://matteospanio.github.io/projects/lilybert/): codebert adapted to lilypond notation. Python, PyTorch, Transformers, Hydra.
- [tasty-musicgen](https://matteospanio.github.io/projects/tasty-musicgen/): musicgen fine-tuned to generate music from taste descriptions. Python, PyTorch, Transformers, AudioCraft.
- [LilyBench](https://matteospanio.github.io/projects/lilybench/): a generation and understanding benchmark for LLMs on LilyPond. Python, PyTorch, Transformers, music21.
- [spam-analyzer](https://matteospanio.github.io/projects/spam-analyzer/): a command line spam filter with a random forest under the hood. Python, scikit-learn, Click, NLTK.
- [ML Project template](https://matteospanio.github.io/projects/ml-project-template/): a copier template to train ML/DL models on a SLURM cluster. Python, Copier, Hydra, submitit.

## Writing

- [Making scientific Python blazingly fast with PyTorch](https://matteospanio.github.io/blog/2025/slow-python/): Swapping NumPy and SciPy for PyTorch buys GPU acceleration, an object-oriented API and AI interoperability at once.
- [Python's virtual environments](https://matteospanio.github.io/blog/2024/python-environments/): A tour of pip, venv, pipx, conda and poetry, and of what each one keeps out of your system Python.
- [Python's Static Typing Safari: In Search of Code Clarity](https://matteospanio.github.io/blog/2023/python-static-typing/): Growing a one-line addition function into annotations, overloads and TypeVar generics, so type mismatches surface in the editor instead of at runtime.
- [Principles of statistics](https://matteospanio.github.io/blog/2022/Principles-of-statistic/): Populations, samples, estimators and the errors that separate an estimate from the parameter it chases.
- [DSP: il suono in digitale](https://matteospanio.github.io/blog/2021/DSP-il-suono-in-digitale/): Come il suono diventa una sequenza di numeri, dal campionamento alla quantizzazione fino alla trasformata di Fourier e alla codifica PCM

## News

- 2026-05-28 — Ital-IA 2026
- 2026-04-12 — Lilybert is available on Hugging Face
- 2025-08-28 — DAFx 2025
- 2025-03-05 — New DL model released
- 2024-11-03 — AIxIA 2024
- 2024-08-18 — AES 2024

## Optional

- [Full text of everything](https://matteospanio.github.io/llms-full.txt): every page as plain text.
- [Machine-readable profile](https://matteospanio.github.io/profile.json): JSON Resume with research and software metadata.
- [Publications as JSON](https://matteospanio.github.io/api/publications.json)
- [CV](https://matteospanio.github.io/cv/)
- [Contact](https://matteospanio.github.io/contact/)