What we’ve
checked.
A few results from development and testing. Each result keeps its sample and scope alongside it.
Earlier detector study
Matched note timing.
Average difference from the checked note-start times.
Lower means closer timing.
Earlier Transkun, Tsumugi and ARIA tests on five development recordings, using the same frozen selector. Values cover matched notes only; they are not a current Piano CRNN accuracy score.
Sample and method
These exposed development passages were already used while building the system. Different engines matched different subsets of notes. This chart shows timing on those matches, not missing-note coverage or an overall model ranking.
Download the full comparison CSV ↗A small research resultOne more note
One more note
from a repeated phrase.
A source-audio recurrence experiment found one additional checked attack in The Spring opening, with no unmatched selections added in that passage.
Research test · not the production routeBefore7 / 12
After8 / 12
12 checked melody attacks · one Spring opening
0 unmatched selections in both versions
Five-pipeline check · two original game recordings
Were the reference notes available?
Pitch-class and note-start matches in a 20-second window.
Higher means more reference attacks had a candidate.
This checks candidate availability, not a finished melody. Octave differences are allowed and extra detections are not penalised. It is not exact-pitch or whole-song accuracy.
Sample and method
Five pipelines; two original game excerpts; one-to-one matching within 70 ms against a source-aligned piano guide. These are provisional guide checks, not a certified reference for every musical part.
Selector training experiment · 16 held-out songs
Research · not enabled in the appTraining closer to detector output.
Two models trained on 64 songs, validated on 16 and tested on 16.
F1 balances missing and extra notes. Higher is better.
Clean-matched and detector-matched training arms use the same protocol. The historical model used different training data and is not pooled into this comparison. Neither new model passed all production acceptance checks.
Sample and method
Same-instrument rendered piano; symbolic attack/context features; song-grouped split. Each new model has 8,897 parameters, 12 epochs and 192 updates. These scores concern the selector experiment, not current Piano CRNN transcription accuracy.
Game checks and acceptance evidence ↗There is more than one chart.
21 recorded passage/model rows, including opening and expanded checks. Full counts remain in the comparison CSV.
10 rows across two original game excerpts. The candidate-coverage chart above shows both recordings.
Two matched training arms, public test scores, game checks and acceptance decisions.
The Spring opening: 7 to 8 of 12 checked attacks, with no unmatched selections added.
Older recorded figures · summary evidence only
These figures survive in an archived handover. Detailed metric receipts were not recovered for this website. The first screen covered seven excerpts in total, but the piano/mixed subgroup counts, matching tolerances and averaging method are missing. These are not verified current accuracy percentages.
The second screen covered two mixed-audio excerpts. It retains onset and release scores, but the full evaluation protocol is not restated in the summary. Do not combine these values with the 70 ms checked-passage results.
Complete recorded data and source records
The bundle also contains the 72-entry research ledger. It is a work log, not 72 successful benchmarks.
Complete research data ↗All cached comparison rows ↗Source records ↗What these numbers mean.
This page highlights selected development results. The automated checks test software behaviour; they do not measure musical accuracy. The research cases cover small, named passages rather than a representative test set of songs.
View source data and testing notes
- Current software checks: 676 available frontend tests and 25 Python tests. Historical tests whose required archives are absent are not counted as passes.
- Earlier timing comparison: the five original checked passages in the cached comparison, with exact-register pitch/onset matching inside the stated tolerance.
- Recurrence case: 7 to 8 of 12 checked targets; unmatched selections remain zero. This proposal was not promoted into production.