Audio · 8.28.26 session
Band Practice Players 8.28
Every complete take from the 8.28.26 practice tape, one player each, with a real waveform, scrub-to-seek, speed control and an A/B loop. False starts, section drilling, tuning and chatter were removed before these were cut.
Read this before you mix off these. The 8.28 session was recorded to a roughly 4 GB WAV, then sent by text. Messages transcoded it to 24 kbps AAC on the way, an 87× reduction, and that transcode is the only copy that reached this Mac. These players are a reference for finding parts and marking timecodes. The lossless master is still on the recording machine.
Ableton master mixes
Full studio mixes of eleven songs from this session, mixed in Ableton on 2026-09-22. Press play on any row, or open the song for the waveform player.
New in this pass
These 0 takes were absent from the first read of the tape. They are
0 minutes of playing that the old detector never cut a take from.
Audio for them alone sits in ~/Music/Band-Practice/songs-8-28-NEW/.
Longer versions of takes already shipped
These 27 carry the 0 songs from the
first 12-song read, each restored to its full length. Audio in
~/Music/Band-Practice/songs-8-28-PREVIOUSLY-SHIPPED/.
What the session was
Of 120 takes, 27 were complete songs. That is 173 minutes out of 312, so about 56% of the tape was run-throughs. The rest was parts work, which is what a writing session looks like rather than a rehearsal of finished material.
| Verdict | Count | Meaning |
|---|---|---|
| SONG | 27 | Complete take, shipped as a player |
| FALSE_START | 7 | A real attempt, abandoned, replayed later |
| WORKSHOP | 13 | The band drilling one section |
| FRAGMENT | 59 | A stab, a riff check, a count-in |
| NOODLE | 14 | No steady pulse, tuning or playing over talk |
Every timecode, so the real file is a one-liner
All 120 takes were exported with exact start and end times, not just the 27 keepers. The transcode has the identical 18745.520 second duration as the master it came from, so these timecodes land on the lossless WAV unchanged. When the 4 GB file arrives, nothing needs detecting again.
./cut_8-28-26.sh /path/to/8-28-26-master.wav
That script cuts all 27 songs with stream copy, so the PCM is untouched. Files on disk:
| File | What it holds |
|---|---|
| 8-28-26-timestamps.xlsx | All 120 takes with verdict, score, BPM and reasons, plus a CutList sheet of the 27 keepers with ready ffmpeg arguments |
| 8-28-26-timestamps.json | The same data for scripts |
| cut_8-28-26.sh | Runnable cut list, takes the master as its argument |
How the takes were judged
The detector treats a false start as a take the band tried again, and compares each take head-to-head against the ones that follow using chroma-sequence similarity with lag alignment. 22 of the 27 keepers pass a from-scratch re-check measured against the room floor rather than against their own quiet passages. The other 5 carry a brief internal stop near the head; they were left whole rather than trimmed, so nothing played is lost. Measured against the source, the kept takes average −6.7 dB and the dropped fragments −17.1 dB.
Why this replaces the 12-song read
The first pass on this tape returned 12 songs and 34 minutes. That was a detector fault, not a quiet session. Calibrating what counts as playing, the finder weighted pulse salience heaviest of its four features. On this board mix that feature carries no signal at all: measured inside the 12 confirmed songs it read lower than outside them. So the heaviest voice in the clustering was noise, and it both dropped 90 minutes of playing outright and clipped the songs it did keep down to their loudest cores. Each of the original 12 turns out to sit inside a longer take here.
Two independent checks settle it. An energy and low-end profile built from scratch,
calibrated on the 12 confirmed songs, recalls 94.5% of them and marks 37 further stretches, 90 minutes,
that the detector never cut a take from. Transcribing samples of those stretches returns zero words,
matching a known song and unlike a known conversation window from the same tape, which transcribes as
speech. The stretches are the band playing. The finder now takes --calib loudness, which
drops salience from the split; the default is unchanged for every other tape.
The older 8.14.26 session is lossless and has its own set at band practice 8.14, with the method written up in the songs report.