01 / IMPLEMENTATION & CONTRIBUTION
What the work involves
Donald built video ingestion, job orchestration, library deduplication, and MIDI postprocessing around upstream Transkun transcription.
Technical depth
Downloaded audio, model failures, GPU availability, overlapping notes, and variable timing need a coherent conversion flow.
Data and assumptions
User-supplied video/audio, predicted MIDI, estimated tempo, timing and velocity values. No private library or copyrighted recording is republished.
Audio transcription is provided by Transkun. The project does not claim a new transcription model.
The project family
pianovision_midi_converter02 / THE EXPERIMENTAL RECORD
What the experiment taught
CUDA transcription can fail; the implementation retries on CPU. Musical transcription accuracy has not been independently evaluated.
What changed
Postprocessing adds partial sixteenth-note quantization, velocity smoothing, and same-pitch overlap trimming.
03 / RESULTS
What came out of it
A conversion workflow with serial queueing, CPU fallback after CUDA failure, optional timing and velocity cleanup, overlap trimming, and a saved MIDI library.
The lesson in the work
The useful product includes the handling around the model: input, recovery, cleanup, and retrieval.
04 / SUPPORTING EVIDENCE
Follow the source
Implementation notes, project records, and supporting artifacts.
Source context & project scope
Transkun is upstream, not a model Donald trained.
No hand-splitting implementation found.
The project does not establish that Donald plays piano or owns PianoVision.
User library/audio files were not inspected.
Transcription accuracy remains unmeasured.
Integration and product engineering; Transkun is an upstream dependency. The project is not evidence of Donald’s musical biography.
MIDI processing estimates tempo, moves timing 80% toward a sixteenth-note grid, smooths velocity, and trims same-pitch overlaps.
SOURCE · 2026-09-17