§ 01 about — a short bio, set in serif

M. Stella
Tavella.

AI · Music · ML · Audio · Music Information Retrieval

Senior AI Engineering Manager at Musixmatch — building systems that understand music through lyrics, audio, and everything in between.

Research in agent orchestration, music information retrieval, and electronic music production.

Ten-plus years in music tech. Currently shipping live AI performance tools for macOS.

§ 02 research — peer-reviewed papers & patents

Peer-reviewed papers

№ 01 2023 3 cit.

Exploiting Music Source Separation for Singing Voice Detection

F. Bonzi, M. Mancusi, S. Del Deo, P. Melucci, M.S. Tavella, L. Parisi, E. Rodolà

IEEE 33rd MLSP Workshop

→

№ 02 2020 8 cit.

Butter Lyrics Over Hominy Grit — Comparing Audio and Psychology-Based Text Features in MIR Tasks

J. Kim, A.M. Demetriou, S. Manolios, M.S. Tavella, C.C.S. Liem

ISMIR 2020 — pp. 861–868

→

№ 03 2019 21 cit.

Exploiting Synchronized Lyrics and Vocal Features for Music Emotion Detection

L. Parisi, S. Francia, S. Olivastri, M.S. Tavella

arXiv:1901.04831

→

Patents

№ 04 2023 Granted

Method and System for Analyzing, Classifying, and Node-Ranking Content in Audio Tracks

L. Parisi, M. Paglia, A. Albano, P. Magnani, P. Melucci, M.S. Tavella

US Patent App. 17/869,053

→

№ 05 2023 Granted

Systems and Methods for Artificial Intelligence Assistant Publishing

M.S. Tavella, L. Parisi, M. Bracci, G. Costantino, F. Delfino, P. Melucci

US Patent App. 18/337,905

→

§ 03 projects — open-source work

№ 01 Featured

aevum — live AI music performance, for macOS

A macOS app for live AI music performance, built on Magenta RealTime 2. Load songs, auto-extract musical loops, morph between them in real time, and control every generation parameter live with a MIDI controller.

stack: SwiftUI · Obj-C++ · Magenta RealTime 2
model: mrt2_small (230M) · mrt2_base (2.4B)
release: v0.1.0 · Apple Silicon · unsigned DMG
license: MIT (model weights CC-BY-4.0)

github→ aevum.app→ model↦

№ 02 Open Source

cocotagger — semantic music tagging, zero-shot

Tag audio with genre, mood, instruments, vocals, and texture — zero-shot via 768-dimensional MusicCoCa cosine similarity. 178 tags across 10 categories; runs entirely on a MacBook CPU, no GPU required.

stack: Python · TFLite · XNNPACK
model: MuLan → CoCa (MusicCoCa) · 44M YouTube tracks
tags: 178 across genre / mood / instruments / vocals / texture / harmonic / tempo / era / production / energy
license: MIT

github→ ★ 8

Also open-sourced № 03 — 06

№ 03 broski Visual browser automation for humans — drag-and-drop web scraping, testing, and monitoring workflows that drive Chromium live. → № 04 Modulr A modular music generation and synthesis framework. → № 05 musicnnCondVAE Conditional Variational Autoencoder for music generation with deep neural networks. → № 06 midi2img2vae MIDI → image conversion with a VAE for music generation and reconstruction. →

§ 04 contact — write, if you would like

Write.