'99% accurate' is a sentence without a test. Here's how to evaluate transcription quality the way engineers do, with nothing but your ears and a piano roll.
Four numbers that matter
Pitch correctness: what share of detected notes are the right pitch. Onset precision: do notes start when the performance starts, or smeared around it. Recall: how many notes actually played were found. And phantom rate: how many detected notes were never played at all. A converter can win any of these by sacrificing another — high thresholds find fewer, cleaner notes; low thresholds find everything and hallucinate.
The octave problem
The most common transcription error is not a wrong note but a wrong octave — the model hears energy at a frequency and its harmonic double. When reviewing a result, scan for notes an octave above or below the musical line; deleting a handful of phantom octaves fixes most 'wrong' transcriptions.
Run your own benchmark
Take three sources you actually care about — a clean solo, a room recording, a dense mix. Convert each at three sensitivity levels and time how long manual cleanup takes. The tool with the smallest total time (conversion plus cleanup) on your material is your best converter, regardless of any leaderboard.