What note detection options does Piano Trainer offer?

Updated

Piano Trainer offers five selectable note-detection models: Standard (instant), Basic Pitch (~80% F1), Transkun V1 (~94% F1), Transkun V2 SOTA (~98% F1), and Kong/Bytedance (~90% F1). The model you pick changes how accurately the app identifies the notes you play through your microphone, and higher-accuracy models generally trade speed for precision. You can switch models at any time from the detection menu, and the choice matters most when you want reliable scoring rather than instant visual feedback.

The detection models compared

Model Reported accuracy Best suited for
Standard (instant) Not stated Fast, low-latency feedback while playing
Basic Pitch ~80% F1 Lightweight detection when speed matters more than precision
Transkun V1 ~94% F1 More accurate scoring for general practice
Transkun V2 SOTA ~98% F1 The most accurate option when you want the cleanest scoring
Kong/Bytedance ~90% F1 A middle ground between speed and accuracy

The F1 figures come from the app's own labels, so treat them as relative indicators rather than guarantees. In practice, a higher F1 means fewer notes are misread as correct, missed, or wrong — which directly affects the "correct / missed / wrong" counters shown at the end of a session.

How to choose a detection model

  • If you want immediate visual feedback while playing, start with Standard (instant). It responds fastest, which helps when you're sight-reading or just watching notes fall.
  • If your session scores look noisy or unfair, move up to Transkun V1 or V2. The jump from ~80% to ~94–98% F1 is the difference between frequent misreads and mostly clean scoring.
  • If your machine struggles with the heavier models, drop back to Basic Pitch or Kong/Bytedance to keep playback smooth.
  • If you're practicing a piece with fast passages, higher-accuracy models reduce the chance that quick notes are missed entirely.

There's no single "best" model — it depends on whether you prioritize latency or scoring fidelity in that session.

Settings that affect detection

Detection accuracy isn't only about the model. Several controls in the interface shape what the microphone picks up:

  • Noise Cancel — options include None, Spectral, and DeepFilterNet3. Enabling noise cancellation helps isolate piano notes from room noise, which can improve detection in less-than-ideal environments.
  • Tune Clarity — shown at 0.63 by default. This governs how strictly the detector interprets pitch clarity; adjusting it can help if your piano's tone is muddy or heavily reverberant.
  • Speed (3fr) and Release (8fr) — frame-based timing controls that influence how note onsets and endings are tracked.
  • Sens (0.50) — the sensitivity threshold for registering a note.

If detection feels unreliable, adjusting these alongside the model is often more effective than changing the model alone.

Calibrating before you rely on detection

The app includes a tuning calibration step: play A4 on your piano and hold it for about one second. This aligns the detector with your instrument's actual pitch, which matters because an out-of-tune piano will cause even a high-accuracy model to misread notes. Calibrate first, then choose your model.

Practical takeaway

Start with Standard (instant) to get playing, then switch to Transkun V2 SOTA when you care about accurate scoring. Use Noise Cancel and Tune Clarity to clean up the input before blaming the model, and calibrate tuning so the detector hears your piano correctly. The model menu is the main lever, but the surrounding settings determine how well any model performs in your room.