missed-songs-audit · six tapes · 2026-09-01

Jam Tape Audit

Every long band tape on this Mac, ranked by how much frequency survived the recorder, then run through the missed-songs audit. One tape carried the same silent failure that cost the 8.28 board tape ninety minutes of playing.

1. Which tape has no frequency loss

Container and bitrate describe a file. They do not measure what reached the disk. The test here samples nine windows spread across each tape, averages the spectrum, and reads the level at 16 kHz against the 1 kHz to 4 kHz band that every one of these recordings has. A codec cutoff shows up as a wall, with energy tens of dB below anything an instrument produces.

The two greyed rows are copies rather than separate sessions. 7:14:26.m4a matches the 8-14 WAV at an envelope correlation of 0.998 once a constant 70 ms AAC priming delay is removed, so it is a lossy sibling of the same performance under a different date in its name. The 5.8.26 WAV matches its m4a at 1.0000 with no lag at all, so it carries no information the m4a lacks.

2. The audit, tape by tape

Coverage is the share of the tape the detector called a song. On a practice tape, anything under about 40 percent is the alarm. Uncovered regions are stretches of at least 45 seconds that measure as playing and carry no take, each one confirmed by transcription against a talk control from the same tape.

3. 7.31.26 carried the 8.28 failure

Re-running the detector with --calib loudness, the same flag that repaired 8.28, took the tape from 10 songs to 35 songs and coverage from 10.6 percent to 63.6 percent. The audit then reported zero uncovered regions. Thirty of the 35 songs are ones the old run never held in any form.

Two songs from the old run do not appear whole in the new one, and both are accounted for. The take at 1:36:57 is now split across two songs that together cover the same audio. The take at 0:30:08 is a 53 second stretch the new run calls two fragments, which is the one genuine downgrade.

4. The 8-14 cut that shipped is short

The 8-14 detector run is healthy on its own terms. Coverage reads 71.6 percent, 80.3 percent counting every take, and the audit reports zero uncovered regions. The problem sits downstream of the detector.

Those eleven were exported by the detector on 2026-08-18 and sit as WAVs in ~/Projects/_outputs/real-song-finder/band-practice-8-14-26/songs/, 210 MB in total. The gap is in the curation step rather than the export.

They were dropped for a reason worth stating. The curated twenty run 74 to 2129 seconds with a median of 274, and score 71.3 to 92.1 with a median of 81.3. The extra eleven run 45 to 146 seconds with a median of 75, and score 59.8 to 72.6 with a median of 63.0. They are short borderline takes sitting just under the strict pass, and the two score bands almost touch at 71.3 against 72.6. Judging them belongs to the ear rather than the detector.

5. The two residual gaps, now recovered

Three tapes had never been through the song detector at all. Two of them came back with playing the new run did not hold, and those stretches are now cut to audio.

• 5.8.26 reads 36.0 percent coverage across 20 songs. Six regions totalling 10 minutes measured as playing and carried no take. All six transcribed to zero words against a talk control that returned six words.
• 7.3.26 reads 38.2 percent coverage across 17 songs. Seven regions totalling 12 minutes carried no take. All seven transcribed to zero words against a talk control that returned 20 words at 0.50 words per second.
• 5-22-26 reads 81.7 percent coverage across 24 songs, the highest here. One region of two minutes carried no take, and it is recovered too.

On both of the tapes that fire the talk control at 0.0 percent, the audit test rejects conversation cleanly, so the zero-word readings in those regions carry their full weight.

6. The calibration flag decides every lossy tape

The 8.28 repair added --calib loudness, which drops pulse salience from the split that decides playing against not playing. That was treated as a fix for one tape. Running both calibrations across this whole set shows it is the correct setting for every phone recording here.

The split follows fidelity rather than session. Five lossy tapes all improve under loudness, by 19 to 75 percentage points. The one full band master is the only tape where the default calibration already reads correctly, at 71.6 percent. Pulse salience carries information when the recording has real high frequency content, and carries noise when a 23.6 kbps encoder has stripped the transient detail the autocorrelation depends on.

7. Method, controls, and limits

Each tape went through the same four steps. The profile is built from loudness and a mid-band ratio, deliberately not the detector's own feature set, so the audit does not inherit the blind spot it exists to find. Coverage calibrates on the detector's own confirmed songs rather than a hand-picked threshold. Every disputed region is then settled by transcription.

python3 scripts/profile_tape.py --audio TAPE --out profile.npz
python3 scripts/audit_coverage.py --profile profile.npz --manifest MANIFEST.json --talk START END
~/.venvs/rsf311/bin/python scripts/speech_probe.py --audio TAPE --windows windows.json
python3 scripts/compare_runs.py --old OLD.json --new NEW.json

The talk control is the load-bearing part. The first probe on 7.31.26 returned zero words in all eleven windows, including the ones meant to be conversation. The script refused to let that stand, because a batch with no talk control cannot prove that a zero-word reading means music. A 24 window sweep across the whole tape then found speech at 1:48, and the audit was re-run against it.

Three limits belong with these numbers. Recall of the audit test against each detector's own songs runs 80 to 83 percent, so it under-reports rather than over-reports. The talk control fires the test between 0 and 33 percent depending on tape, which is why every region here was transcribed rather than believed on spectrum alone. And each probe samples 75 seconds from the middle of a region, so a long region is confirmed at its centre rather than end to end.