Music AI Lab
Piano-domain evaluation and interpretable musical intelligence.
Piano Evaluator MVP
Piano Evaluator is the first practical direction of the Audience-Centered Brain-like Music Cognition and Performance System: a piano-focused evaluator that turns audio, optional score input, and expert listening goals into traceable reports, rubrics, and improvement suggestions.
- Audio-first analysis with optional score input
- Evidence-based reports and rubrics
- Goal tags before open-ended LLM interpretation
Evidence architecture
The MVP is organized around Music Understanding Files: raw audio events, audio note candidates, Audio-MUF, Score-MUF, Aligned-MUF, and future Performance-Plan-MUF. The point is to keep audio evidence, score expectation, alignment confidence, and performer deviation inspectable rather than hiding them inside one opaque model output.
Audio-MUF stabilization
The current implementation notes prioritize robust candidate handling: Basic Pitch events should be sorted before any cap is applied, short-note recall should be tested around an 80 ms default, and fast-note groups, onset-only candidates, coverage warnings, and beat-level summaries should stay visible for later alignment.
Prototype directions
The next public-facing materials can be demo clips, annotated listening cases, parameter-sweep summaries, interface sketches, and a GitHub or project link once the prototype is ready to be shown without overclaiming its musical judgment.
- Demo video or walkthrough
- Annotated case studies
- GitHub, paper draft, or prototype notes