What the tests found
Accuracy & limitations
The analyzer always returns its best estimate after a successful analysis. That is useful behavior, but it is not the same as proven accuracy.
- Best-effort estimates
- Measured limitations
- Blind holdout not run
Current benchmark snapshot
On September 9, 2026, the analyzer completed all 71 items in its calibration set. It included 56 short GuitarSet acoustic-guitar recordings and 15 generated stress tests for atonal audio, drums, modulation, noise and speech. This was not a collection of 71 full songs across many genres, and the project's accuracy targets were not met.
Finishing every analysis means the tool returned an answer. It does not mean every answer was correct.
These items were used while evaluating the analyzer. They were not an untouched blind test. The reserved 29-item holdout has not been run, so no blind-validated accuracy claim is made.
What do generated test examples show?
These two 12-second fixtures were generated locally by the committed evaluation script and checked in the release browser on September 9, 2026. They are controlled diagnostic probes, not evidence of real-song accuracy.
Generated example 01
120 BPM drum pattern
- Designed signal
- 120 BPM, no tonal key
- Observed output
- 120 BPM, F# minor
The tempo matched the generated pulse. The key label is not musically meaningful because a drums-only signal has no intended tonal center.
Generated example 02
Deterministic noise
- Designed signal
- No stable BPM or key
- Observed output
- 107 BPM, Db major
The plausible-looking output is a false interpretation. It demonstrates why the always-output mode must be described as best-effort.
Why the tool still returns a result
Tempo & Key Finder follows an always-output product mode: when decoding and analysis succeed, it displays BPM, key and Camelot as a best-effort estimate. The 100% output coverage above describes availability of an answer, not correctness. Silence, malformed detector output and technical failures still produce errors rather than invented values.
Common tempo ambiguities
- Half-time and double-time can describe the same groove.
- Syncopation may make the detector follow a secondary pulse.
- Live performances and tempo drift do not have one fixed BPM.
- Quiet intros, speech and noise can resemble periodic rhythm.
Common musical-key ambiguities
- Relative major and minor keys share the same key signature.
- A song may modulate instead of staying in one key.
- Percussion and speech may contain no reliable tonal center.
- Short, sparse or noisy sections can distort the key profile.
When the two harmonic profiles disagree, the second result is shown only as a possible alternative. Neither result should replace listening or musical judgment.
How can the evidence be checked?
Real-audio calibration uses the GuitarSet dataset under CC BY 4.0. The generated fixtures, selection rules, raw observations and release checks are included in the corresponding source archive.
The calibration cohort is not an untouched test set. The reserved 29-item blind holdout remains unrun, so these materials support reproducibility and limitation reporting rather than an accuracy guarantee.