A saved F-major harmony case links Roman-numeral labels to the score. The first 12 events have applicant-confirmed review: ten marked correct, two retaining contextual or voice-leading questions.
“C7/B♭ is correct as a sonority. Review the seventh-resolution voice leading.”
Applicant-confirmed review record
IMPLEMENTATION EVIDENCE / RULE EXAMPLE
Make voice relationships available for review.
The inspected counterpoint code checks aligned voices for intervals and motion. In the parallel-fifth route, two consecutive perfect fifths with motion in the same direction produce a rule issue.
Input
Aligned note events in two or more voices
Rule
Consecutive fifths + same-direction motion
Output
Issue records with voice and time evidence
Human review
Interpret the writing in its species and phrase context
Illustration of the supplied code example: [48, 55] → [50, 57]. Both voices rise by a whole tone.
Source and rule illustration, not a fabricated product result. Comprehensive real-score validation is not reported here.
RECORDED REPORT / CANDIDATE ANALYSIS
Read proposed sections. Question their boundaries.
The mini-program report proposes A–B–A′ across 24 measures: 1–8, 9–16 and 17–24. Cadences are explicitly marked for further review. This is a recorded system suggestion, not expert-validated form recognition.
Input
Uploaded score
Analysis
Proposed section ranges and tonal context
Output
Structured section report
Human review
Check recurrence, contrast and cadence evidence
MGSA is a separate boundary-detection research line; it is not presented as the deployed form-analysis engine.
The actual mini-program workflow shows score upload, analysis and report inspection. The original sequence is retained, including the time required to complete the task.
Upload the source score.
Run the analysis.
Inspect the report and score annotations.
This recording demonstrates the workflow; it does not establish overall recognition or analysis accuracy. The saved 12-event review below is a separate evidence item.
Web tools: twelve-tone matrices, metronome, tuner and counterpoint practice. Score analysis runs in the mini-program.
Form boundary test
Beethoven · Piano Sonata No. 2, Op. 2 No. 2 · II. Largo appassionato
I ran Xunpu Studio’s existing symbolic form-analysis engine offline on the complete 80-measure MusicXML movement. The run disabled catalog lookup and online models. I compared the resulting section ends with the locally aligned top-level Taking Form annotations, excluding the piece ending.
Two timelines on the same 80-measure grid. Green marks connect exact matches. Select the figure to enlarge it.
Two local comparisons: reference after 12 vs. candidate after 14, and reference after 43 vs. candidate after 45.
Matching tolerance
Matched boundaries
Boundary F1
Exact
2
0.2857
±1 measure
2
0.2857
±2 measures
4
0.5714
The candidates after measures 31 and 67 match exactly. The candidates after 14 and 45 are two measures later than the references after 12 and 43. The system proposes fewer sections than this reference. The benchmark checks boundary placement for one interpretation of one movement. It leaves PDF recognition, key identification and form labels outside the evaluation.
Boundary positions on the score
These excerpts come from the same MusicXML used in the test, converted from Craig Stuart Sapp’s encoding of the Durand 1915 edition. The labels mark the first measure of each proposed section.
Measures 12–15. The reference starts at measure 13. Xunpu starts at measure 15.Measures 43–46. The reference starts at measure 44. Xunpu starts at measure 46.
Chord-content and inversion checks in an F-major context. The applicant confirmed this scoped review; comprehensive voice-leading and recognition accuracy remain outside it.
APPLICANT-CONFIRMED HUMAN REVIEW
12 harmonic events, examined individually
Chord content and inversion are checked in an F-major context: ten entries are marked correct; two retain voice-leading or contextual questions. No whole-system accuracy percentage is inferred.
First 12 events in the saved F-major case. E03 and E09 are flagged for contextual review.
Scope & source
Human-review status is published following the applicant’s confirmation. A review date and independent reviewer identity were not supplied.
The evidence covers the first 12 harmonic events, not the whole ten-measure, 24-event output. The original input PDF was not recovered; an unannotated cached score raster was available. The remaining events and comprehensive voice-leading checks are outside this review.
C7/Bb is correct as a sonority. The bass moves Bb–C before A; review the seventh-resolution voice leading.
4
2
2.0
V
C3 E3 C4 G4
V
Correct
C-major triad, root in bass.
5
2
3.0
I6
A2 F3 C4 F4
I6
Correct
F-major triad, first inversion.
6
3
1.0
IV
Bb2 F3 D4 Bb4
IV
Correct
Bb-major triad, root in bass.
7
3
2.0
I6
A2 F3 C4 A4
I6
Correct
F-major triad, third in bass.
8
3
3.0
V
C3 E3 C4 G4
V
Correct
C-major triad, root in bass.
9
3
4.0
V7
C3 E3 C4 Bb4
V7 (no fifth)
Requires contextual review
Root-position dominant seventh; G is omitted. Make the incomplete voicing explicit; assess its resolution in context.
10
4
1.0
I6
A2 F3 C4 A4
I6
Correct
F-major triad, first inversion follows the dominant seventh.
11
4
3.0
V
C3 E3 C4 G4
V
Correct
C-major triad, root in bass.
12
5
1.0
V42
Bb2 E3 G3 C5
V42
Correct
Complete C7/Bb; the next bass A supports the third-inversion resolution.
This saved harmony case and the full product recording are presented as separate evidence items; no frame-by-frame case correspondence is claimed.
03 / CLOSED-PILOT EVIDENCE
A system used beyond the workspace.
50Participants
1,247Core-function calls
8Weeks · Dec 2025–Feb 2026
Project records report an eight-week closed pilot. These figures describe participation and function use; calls are not unique users or daily active users.
Participant breakdown, task distribution and systematic feedback outcomes are not reported here. The recording and saved analysis demonstrate product behaviour separately from pilot usage.
04 / SYSTEM & RESEARCH BOUNDARIES
Symbolic representation is the bridge.
Implemented processing routes depend on file type, configuration and service availability.
INPUTScore / PDF / image
MusicXML · MXL · MIDI
IMPORTParsing / OMR
Configured perception route
REPRESENTATIONSymbolic music
Pitch · time · measures · voices
REASONINGMusical modules
Harmony · counterpoint · form
REVIEWEvidence + explanation
Annotations → human judgement
INDEPENDENT RESEARCH / NOT DEPLOYED PRODUCT INFERENCE
AI 工作台 / AI WorkbenchTask orchestration with visible tool status
Feature overview
INPUT → PROCESS → REVIEW
01 / INPUT
Score + task
Import the score, choose the module, and describe the musical question.
02 / PROCESS
Module tools
The workbench calls the relevant analysis tools and displays task status.
03 / OUTPUT
Answer + evidence
Read the final answer and visible tool status; rule conclusions come from the analysis module.
FUNCTIONS IN THE IMPLEMENTATION
AI task orchestration, with module-specific analysis.
The desktop workbench connects a user task to existing modules. The module advisor provides quick or detailed explanations, with SATB audition controls in the desktop interface.
Task orchestrationTool statusModule advisor
Saved harmony case · candidate annotations
Historical OMR output: 10 measures, 24 harmonic events, F-major key context. The original system labels are preserved. The first 12 events have applicant-confirmed human review; remaining events are outside that review.
AI assistance requires configured services. This website presents the module workflow and does not execute a task.
Feature overview · based on desktop modules and sourceInteractive portfolio preview
Switch modules to inspect their actual purpose, processing steps, and outputs.
Founder · Product Research Lead2024—present
Xunpu Studio
Integrated music learning, analysis, creation & practice workbench
An integrated music-learning, analysis, creation and practice workbench for music students, composers and emerging creators. Reviewable analyses of harmony, counterpoint and form connect with AI-assisted composition, twelve-tone tools, orchestration support, ear training, rhythm feedback and an AI music advisor.
MusicXML/MXL/MIDI and PDF/image input via OMR are entry routes into the workbench. Musical rules, score annotations and explanations allow users to inspect, question and revise the system's reasoning. I lead creator research, product design, prototyping and development coordination.
Music technologyHuman–AI interactionProduct design
Closed pilot Project records report 50 participants and 1,247 core-function calls over eight weeks (Dec 2025–Feb 2026).
I lead creator research, product design, prototyping and development coordination. My composition training informs the musical questions the system makes available for review.
The interface connects module-specific analysis to source scores, visible evidence and explanations. The review case illustrates the principle: a plausible chord label can still require a contextual musical decision.
