Artificial intelligence has transformed numerous creative fields, but music remains one of the most complex territories for automation. For years, symbolic music analysis systems have focused on large-scale prediction — processing an entire score and generating a harmonic analysis without human intervention. However, the real workflows of musicologists, composers, and educators are rarely linear: they need to correct partial sections, complete missing fragments, or refine interpretations through successive iterations. This gap between high-performance models and the flexibility demanded by day-to-day work has motivated a new approach: moving from pure prediction to interactive human-machine collaboration.
This paradigm shift resonates with what we at Q2BSTUDIO have been applying in enterprise software development. Just as a music analysis system must be able to work with incomplete parts and allow local corrections, a custom software application for business management must adapt to evolving processes, not rigid pre-conceived flows. The key is to design systems that reuse expensive precomputed representations during iterative refinement phases, minimizing latency and maximizing responsiveness. In the musical domain, this means that a powerful model — trained on thousands of scores with Roman numeral annotations — can serve not only for full analyses, but also to complete missing labels in a specific measure or to revise a harmonic progression without recalculating everything from scratch.
The unified framework proposed here enables three usage modes covering the entire spectrum of interactive analysis: full-score prediction, inference of missing annotations from partial context (something like harmonic autocomplete), and localized revision of existing labels. To achieve this versatility, the system separates heavy computation — feature extraction via pre-trained deep networks — from lightweight inference performed in real time. This way, the analyst can select a fragment, request a suggestion, adjust it manually, and see how the model updates the rest of the harmony without perceptible delays.
This approach is not exclusive to musicology. In the business world, tools like AI agents are beginning to offer similar experiences: assistants that complete reports, correct forecasts, or suggest optimization paths while the user retains control. The same philosophy of 'expensive compute once, lightweight inference many times' is what we use in cloud AWS/Azure projects to process large data volumes in batch and then serve interactive queries with millisecond latency. Cybersecurity also benefits: an intrusion detection system can precompute behavior models and then update rules on the fly without restarting monitoring.
Experiments on Dilemmadata — the largest and most heterogeneous benchmark for Roman numeral analysis — show that the proposed framework not only matches state-of-the-art models in accuracy, but also offers the ability to complete partial masks with quality comparable to full prediction. This is crucial because in practice no musical analysis is definitive on the first pass: experts always want to try variants, correct ambiguities, and explore alternative interpretations. An interactive system that enables that dialogue becomes a collaboration tool, not a black box.
From a technical perspective, the architecture relies on pre-trained representations that capture both rhythmic structure and long-range harmonic relationships. These representations are stored in a vector database queried during interaction. The lightweight inference model is a small transformer operating on the fixed representation, generating Roman numeral labels (degree, inversion, function) for each chord or for a set of masked chords. Iteration is achieved through a feedback loop: the user modifies a label, the system updates the local representation, and re-infers adjacent labels, all in under 100 milliseconds per measure on conventional hardware.
This rapid iteration capability opens the door to educational applications where harmony students can experiment with different options and receive contextual suggestions. It also allows composers and arrangers to test harmonic progressions over a given melody, instantly seeing how it affects tonal coherence. And for musicologists, it offers a semi-automatic annotation tool that drastically reduces labeling time without losing control over the result.
At Q2BSTUDIO we understand that technology should not impose a workflow, but adapt to each user's needs. That is why, when developing custom software solutions, we apply similar principles: we separate intensive batch processes from real-time interactions, offer modular interfaces, and allow users to correct, complete, or refine data without relying on costly reprocessing. We carry this same approach into areas like Business Intelligence, where BI/Power BI dashboards can benefit from precomputed prediction models that update with new data without rebuilding the entire panel.
The future of symbolic music analysis lies in collaboration between human intuition and computational power. Systems that achieve this balance will not only improve analyst productivity but also enrich musical practice by offering new harmonic perspectives. The same logic can apply to any domain requiring structured yet flexible analysis — from genomics to logistics. In all cases, the goal is not to replace the expert, but to provide a digital companion that learns from their corrections and suggests paths they might not have considered.
The research described here demonstrates that it is feasible to build a unified framework combining high-quality prediction with real-time interaction. The results on Dilemmadata consolidate the proposal as a new baseline for Roman numeral analysis, but its real value lies in the user experience: being able to complete, correct, and refine without friction. That is the transition from prediction to collaboration, a shift we are also driving in enterprise software development at Q2BSTUDIO, where every project is conceived as an ongoing dialogue between machine and person, optimized for efficiency and creativity.
