MusicReasonResource
Why Audio Recognition Is Ambiguous
An audio estimate is a bounded inference from a signal under stated conditions, never a fact extracted directly from a recording.
MusicReasonResourceWhy Audio Recognition Is Ambiguous
Measuring a signal and inferring a musical fact from it are not the same operation. A microphone's captured waveform is a measurement. Deciding that waveform represents "the note A4, played cleanly" is an inference, built on assumptions about signal quality, harmonic structure, and how much of the true source made it into the recording. Every audio-analysis engine on MusicReason keeps that distinction visible instead of collapsing it into a single confident-looking number.
Three certainty labels, never mixed up
| Label | What it expresses |
|---|---|
| Calibrated confidence | A probability-like estimate validated against a representative test corpus |
| Match score | How closely a candidate matched a reference pattern, without a probability claim |
| Confidence band | A coarse category (for example, strong, moderate, weak) describing signal-evidence quality |
No engine invents a fourth label or silently renames one of these three. Each recognition engine is defined to use exactly one of these modes, and MusicReason never presents one mode's number as if it meant another.
What a confidence band does not mean
A confidence band or match score is not:
- A guarantee that the reported answer is correct.
- A percentage chance computed the same way for every engine on the site.
- Evidence about anything other than the specific signal you provided, at that moment.
- A substitute for stating alternatives when the evidence genuinely supports more than one candidate.
Why alternatives appear instead of a single winner
A struck note still decaying, background noise, or a signal rich in harmonics can all support more than one plausible reading. Microphone Tuner and the file-based analysis engines report a next-best candidate and its supporting evidence when one exists, rather than silently discarding a genuinely close alternative to present a falsely clean single answer.
Signal quality states are part of the answer
Silence, clipping, background noise, and overlapping notes (polyphony) each degrade what a detector can honestly claim. Peak and RMS Meter and Spectral Peak Finder report these quality signals directly, rather than only reporting a final result and hiding the conditions it was extracted under.
How a claim like this gets tested
Every recognition-style claim on this site is checked against synthetic signals with a known correct answer before it ships, plus edge cases designed to produce low confidence on purpose (heavy noise, near-silence, ambiguous harmonic content). See How MusicReason Tests Its Answers for the full testing approach.
Frequently asked questions
No. A confidence band on MusicReason describes how strong and clean the supporting signal evidence was for a given estimate, not a calibrated probability that the answer is true. The three labels this site uses, calibrated confidence, match score, and confidence band, are never interchanged or renamed per engine.
A real signal can carry more than one plausible fundamental, especially with harmonics, background noise, or a struck note still decaying. Showing the next-best candidate and its evidence is more honest than silently discarding a genuinely close alternative.
No. Full-song chord-progression extraction is explicitly excluded from this product (Specs/unresolved-decision-register.md, MR-DEC-004) until a licensed evaluation corpus proves the method meets a stated accuracy threshold with honest uncertainty, and none currently ships.