Guitar Transcription Workstation

1. This Mac is Intel, not Apple Silicon

uname -m returns x86_64 on macOS 13.7.8. The Apple Silicon rationale for pinning Basic Pitch to Python 3.10 does not apply here.

2. A new 3.10 Basic Pitch environment would rebuild a working one and reintroduce a crash

~/venvs/audio_midi_311 holds basic-pitch 0.4.0 on Python 3.11.15 with onnxruntime 1.23.2. The constraint on this machine is the model format. The Python version is not what breaks. Basic Pitch defaults to:

/Users/drwu/venvs/audio_midi_311/lib/python3.11/site-packages/
    basic_pitch/saved_models/icassp_2022/nmp.mlpackage

That CoreML package SIGFPEs on this Mac. nmp.onnx ships in the same directory, so no download is needed. env.sh pins BASIC_PITCH_MODEL_TYPE=onnx, and a pytest case fails if that pin is ever dropped.

3. The validator stack was installed before this session, minus one package

Python 3.13.6 at /usr/local/bin/python3 carried pyguitarpro, lxml, numpy, scipy, librosa, soundfile, audioread, mido, pretty_midi, pandas, matplotlib, pydub, and music21. Only pytest was missing, now installed at 9.1.1. Building a fresh 3.12 venv would have duplicated thirteen working packages. Demucs 4.0.1 is installed in ~/venvs/demucs on Python 3.12.13 with torch 2.2.2.

4. validate_song.py mostly exists

The gates for pitch preservation, impossible spans, string collisions, ownership, octave doubling, tuning, and candidate versus source statistics are written and tuned. Writing a second set from scratch would have thrown away the recorded regression cases. The new script is an orchestrator over them, roughly 130 lines.

The finding the run surfaced

The existing two staff split of this song put 585 notes on Guitar and 1 note on Guitar 2. The split ran and produced a file, and the file does almost nothing. The 42 percent hand skip and the nine ownership flags remain on the first staff. Exit zero and a written file prove only that the tool ran. Neither measures whether the output is good.

Another session is editing that same tree right now. The delivery README changed on disk mid read. Nothing here writes into ~/Projects/_outputs/APPLESEED-DELIVERY-2026-08-28/, and coordinating before any repair pass is the safe move.

No octave verdict yet, because the beat map fails its own control

On Sentence, with the score tempo of 200 bpm at offset 0:

map anchor 1  notated onsets to nearest real attack:  median 88 ms
              SAME map pointed 11 s off the mark:    median 83 ms
              -> MAP DOES NOT BEAT ITS CONTROL
map anchor 2  notated end 153.5 s vs audio 177.5 s (13.5% drift)
MAP           SUSPECT

Pointing the map eleven seconds away from the music aligns slightly better than pointing it at the music. A map that cannot beat its own wrong time control does not locate notes, so the per note ratios below carry no information. The wrong time ratio control agrees: median 1.31 against 1.122 for the flagged notes.

The probe prints MAP SUSPECT and withholds the verdict. That is the gate doing its job. The existing 04_TABLES/octave_test.json shows offenders at median 2.17 against a decoy control at 1.77, which is very likely the same problem: with a broken time map, offenders and decoys measure the same thing.

Reading it honestly

Four of the seven point to a note an octave below, two are ambiguous, one matches the written pitch. Since MIDI 88 cannot be played anywhere on a 22 fret guitar, the one reading at 2.835 cannot be a guitar fundamental at that pitch either; the likelier explanations are another instrument bleeding into the stem, or a partial of a different note. The majority evidence supports the harmonic hypothesis on this song, and one song is not five.