Accuracy Limitations and Best Results

Accuracy, Limitations and Best Results

A2M estimates a performance from sound. It can make a useful editable starting point, but it cannot recover musical information that the recording does not make distinguishable.

Give the engine a clearer problem

Results are generally easier to interpret when the input has:

• one piano rather than a full mix;
• little clipping, distortion, or aggressive compression;
• audible attacks and natural dynamics;
• limited speech, percussion, crowd noise, or strong effects;
• enough level to reveal quiet notes without hard limiting;
• no unnecessary chain of lossy re-encoding.

Automatic denoising is not always an improvement. Processing that smears attacks can remove timing evidence the transcription needs.

What commonly needs editing

Area	Why it is difficult	What to inspect
Quiet notes	Masked by louder harmonics or noise	Missing inner voices
Repeated notes	A new attack can blend into the previous decay	Note splits and short gaps
Note endings	Resonance, pedal, and room response blur release	Durations and overlaps
Dense chords	Related harmonics compete	Extra or missing chord tones
Pedals	Acoustic resonance is indirect evidence of controller state	CC64 and CC67 timing
Mixed audio	Other instruments create piano-like spectral energy	False notes and unstable timing

The current engine models all 88 piano pitches, MIDI 21 through 108, plus sustain and soft pedal symbols. It does not model sostenuto and is not a general instrument recognizer.

CPU versus GPU output

Changing provider does not select a different model. Small numerical differences are possible because providers and hardware can evaluate floating-point operations differently, but GPU mode should not be described as a more accurate engine. Compare output on the actual system if exact provider parity matters.

Review a transcription

1. Listen while following the MIDI in a piano roll.
2. Check melody and bass structure first.
3. Inspect repeated notes, quiet voices, and dense chords.
4. Compare releases with pedal lanes temporarily hidden.
5. Judge velocity as a musical contour before polishing exact values.
6. Correct structure before expression.

The Worked Example and Validation page demonstrates this process with an MP3, source MIDI, A2M output, waveform, log-frequency spectrogram, aligned piano rolls, and explicitly defined onset metrics.