Vox Probe 12

Previous topic - Next topic

0 Members and 1 Guest are viewing this topic.

Andrew Davie

VoxProbe 12

Build 1: vox_probe12.bin (generator make_probe12.py, manifest probe12.json), 9 sections, 37 chunks, 1559 bytes, about 248 s. Title "VOXPROBE / 12 B1". Not yet handed over; frozen once it is.

Every section was chosen from the measured losses (E per recording group, 2026-11-03, work/dehum/xramp0.json; Man Goes Down's timing, probes/VoxProbe11/analysis/p11mgdtime.py), and asks a question the existing recordings cannot answer. Lessons from VoxProbe 11 built in: a Reset before every marker and every item (VoxProbe 11's sections 9-10 inherited Man Goes Down's Pitch 48 / Bend 7 / Speed 62, moved their markers and lost the last reference); ablation instead of single contexts; every section checked placeable on a library dry run with a chain background.

Why: where the score is lost

group (E, sum over items of the excess)    excess   what is open
p4 2  (weak effects at Pitch/Bend)          +89     Bend 0-2 low-tone line, Bend 15
p7 bend 0                                   +68     Bend 0-2 low-tone line
p4 7                                        +33
p7 bend 15                                  +27     Bend 15 (no wrap there: unexplained)
mgd                                         +20     timing (below), C1
p7 pitch 200                                +19     effects' period at 200
Man Goes Down, chip against library, sound by sound (VoxProbe 11 section 8, 3 takes from a Reset, A and B agree to ~20 ms, so the differences are the chip's):

phrase  sound              library  chip   diff   context
P1      AY                    190    275    +85   Slow, Speed 108
P1      NE                    114    160    +46   Slow
P1      OWWW                  248    320    +72   after Slow GO (the library's Slow does nothing on GO, nor OWWW)
P1      NE (end)              118    150    +32
P2      EYIY                  206    285    +79   after MM, Speed 118, Pitch 64, Bend 3, no Slow
P2      WW IH                         +25 +16
P3      C1                    747    980   +233   Speed 62, Pitch 212
P4      IYUW                  272    410   +138   Slow, Pitch 104
P4      UX (3rd)              135    195    +60   after Pitch 48, phrase end
total   P2 / P3 / P4                  +184 / +236 / +190 ms
Library Slow is x1.22 (VoxProbe 8, single sounds at Speed 114); in Man Goes Down the chip's slowed sounds run x1.36-1.83 of the library's plain length. Either the factor is wrong, or Slow does more than lengthen one sound.

The sections, what each settles, how it is read

Reference (SCP 1000 Hz, silence): the clocks, the background.
Slow and Fast. AY IYUW OWWW NE AW UX alone and Slow (Fast on AY, IYUW); AY and IYUW at Speed 108 plain and Slow; Slow before a stop (GO OWWW, DO AYWW) and a fricative (SE UX) against the same pairs plain; Slow AY then AW (does Slow persist past its sound?), with a P2 between; Slow Slow AY and Slow AY Slow AY (stacking); AW Slow IY AW, AW Fast IY AW (between sounds). Read: each sound's length (p11mgdtime.py's method: library starts from the engine, chip by DTW). Settles: Slow's factor per sound class and Speed, persistence, pass-through, stacking.
Controls between sounds. AW AW AW against AW Pitch 88 AW Pitch 88 AW (same value: only the command's cost), Bend 5, Speed 114; UX Pitch 56 UX Pitch 48 UX (P4's end) against UX UX UX. Settles: whether a control between sounds adds time or changes the join (P4's last UX +60).
Man Goes Down taken apart. P1 P2 P4 as sent, then without their Slow codes, Pitch codes, Bend codes, Speed codes (5 x 3). Read: sound-by-sound lengths per variant against the full phrase. Settles: which control makes each mis-timed sound (EYIY's +79 with no Slow in reach: Pitch 64? Bend 3? Speed 118?).
The effects' Speed law (open item: the chip's law is curved, the library's (147-s)/33 is a fit). B2 (402 ms, six parts with clear edges) and M2 (518 ms, two parts) at Speed 20 40 62 90 127; C1 at Speed 40 62 90 114 by Pitch 88 and 212; P3 exactly. Read: total and per-part lengths. Settles: the law's curve, whether parts scale alike, and whether C1's length follows Pitch (it is the own-pitch effect).
The default pitch (open item 4: the effects run 92 ticks at the reset default, which ceil(8000/88) does not give). M2 (voiced, 0.5 s, period readable to a fraction of a tick), C5, AW at the default and after Pitch 88, 87, 89. Settles: whether the default is 87, or special; this decides the period of every pulsed effect at default, i.e. nearly every recording.
Bend. 252 M0 (one steady 1843 Hz tone for 1.2 s) at Bend 0-6, 8, 10, 12, 15: the law over the whole range at one frequency, to 1 Hz. 211 A1 (a steady 168-176 Hz tone under two sweeps, 957 ms) and 224 B4 (a steady 1827 Hz tone beside one swept 1990 -> 300 -> 1998 Hz) at Bend 0 1 2 3 4 5 8 12 15: a low tone held below the zero point, and one swept through it, at every low Bend, against Bend 5. 215 A5 at Bend 0 by Pitch 40/88/150 (does the K - shift line move with Pitch?). 210 A0, 218 A8, 235 C5 (the worst Bend-0 losers) at Bend 1-4 (0 and 15 are on VoxProbe 4 and 7). Settles: the low-tone line's K (FACTS §8: 1885-2352 Hz per part, rule unknown) and level against the bent frequency, and what changes at Bend 15.
M2's two forms (FACTS §24: F1 present only on VoxRecord). M2 from a Reset alone; after P2 P2; after M1 P2 P2 (VoxRecord's context); after M0 P2 P2 M1 P2 P2 (VoxRecord's stream); after Pitch 88; after Bend 5; after M2; after AW; M1 M2 joined. Settles: what turns F1 on; then one table serves every recording (P 1.74 -> ~1.55 is waiting).
Reference again.
Left out, and why

Anything on the voiced stops (OG EG EB OB GE ED lose +6..+30 each): their error isn't yet pinned to a mechanism; a probe now would be guessing. Analyse first (component-wise: tf or env?), then VoxProbe 13.
233 C3 at low Pitch (one sound, +17), the bed (VoxProbe 11 answered it: not the oscillator), the buzz rate (VoxProbe 11: fixed), Volume (no scored recording uses it).

SpiceWare

Took longer than I thought it would but finally done:

vox_probe.zip 1.48 GB

Andrew Davie

MANY thanks for this extensive bit of work. Much appreciated.
I am currently totally reworking the algorithm, after an epiphany.  Results tomorrow probably.

Andrew Davie

#3
**VoxProbe 12: what it was for and what it found**

Built and released on 2026-11-04 (build 1, sha1 97ac7b29, 248 s, 9 sections), after VoxProbe 11 left the real unknowns open. Its targets were Man Goes Down's timing errors, Slow/Fast, the effects' Speed law, the default pitch, the Bend law and 254 M2's two forms. It was recorded on 2026-11-05: four takes (2600 and Stella, on unit A, the AtariVox+, and unit B, the original). That was the first set with both hosts on both units.

**The recordings**
- 133 items placed on unit B and 114 on unit A. Section 5 wouldn't split on unit A, and section 8 split on neither unit. Lengths were read in the chip's own time, using section 1's 1000 Hz reference as each take's clock.
- **Clocks:** unit B 8008–8009 Hz on both hosts. Unit A 8049 Hz (2600) and 8111 Hz (Stella), so it runs 0.4–0.8% fast on the USB interface's power, while unit B doesn't move. That is now corrected for in every comparison.
- **A generator bug:** names were looked up as sounds before being read as controls, so "B0"–"B8", meant as Bend 0–8, went out as effects 220–228. Every Bend 0–8 item in section 7, and the Bend items in sections 3 and 8, played an effect instead. Only Bend 10, 12 and 15 were recorded correctly.

**What it settled**

1. **The default pitch is Pitch 88** (section 6). The default and Pitch 88 gave identical periods on every take (92 ticks for 136 AW, 92.2 for 254 M2). The engine's period rule already produced exactly that, so the open "92 ticks at default" question was closed with no change.

2. **Slow and Fast were wrong** (section 2).

                     chip              library then
Slow factor          x1.27-1.36 (1.32) x1.23
Fast factor          x0.63-0.70        x0.75
Slow on a stop       +12 ms (GO OWWW)  +0
Slow on a fricative  none              none
Slow Slow AY         = Slow AY         (Slow doesn't stack, doesn't carry to the next sound)

   → The engine changed to **Slow ×4/3 and Fast ×2/3** of the played length, plus Slow on voiced stops (commit 9608c9d).

3. **A sound repeated after itself was too long** (sections 2 and 3). The library made each same-sound join about 31 ms too long (AW AW AW −62 ms, UX UX UX −69, AY AY −64). A Pitch or Speed command between two sounds costs no time on the chip.
   → The engine changed to **play a sound straight after itself as a Repeat copy**, the same-sound join (commit 2688ab2).

4. **Man Goes Down taken apart** (section 4), with each control removed in turn:

                  chip    library   chip - library
P1 as sent        1487    1297      +190
P1 without Slow   1294    1200       +94   Slow carried ~96 ms of the error
P1 without Speed  1274    1132      +142
P2 as sent        1684    1720       -37
P2 without Speed  2100    1892      +207   the error depends on Speed
P1/P2 without Pitch or Bend: within 14-24 ms - neither causes the errors

  Sound by sound, using VoxProbe 11's alignment method, the errors have three causes, and both units agree:
  - the Speed law near the default (the chip hardly shortens at Speed 108 or 118, where the library took 5–12% off)
  - Slow's factor (fixed in item 2)
  - joins that no control explains (P1's GO → OWWW +47–72 ms, P4's IYUW at a phrase start +138–158 ms)

5. **The effects' Speed law is curved** (section 5, unit B only). 222 B2, 254 M2 and 231 C1 agree within about 4%. The chip is 2–6% longer than the library's straight line at Speed 40, and 3–5% shorter at 90–127. Effects scale part by part, which the library already does (241 D1 within 2%), but 230 C0 is too flat at both ends. The engine wasn't changed for this.

6. **Bend 10, 12 and 15 follow the engine's law**: about +44 Hz a step on 252 M0's tone, confirmed on 224 B4 and 211 A1. Below Bend 10 nothing could be learned, because of the bug.

7. **254 M2's two forms** (section 8). In every VoxProbe 12 context M2 played its **low form**: alone, after P2, after M1, M2 M2, AW M2, the VoxRecord stream. That held on both units and both hosts (band ratio −4.4 to −6.9 dB). Each of those started from a Reset. VoxProbe 4's M2s, sent without a Reset, are the **high form** (+12 to +13 dB) whatever the Pitch (60 = 150) or Bend (0 = 15). So Pitch and Bend don't set the form; the state since the last Reset does.

**What it couldn't answer, and where that went**

VoxProbe 13 (build 1, sha1 bc663f87, built 2026-11-05, not yet handed over) carries the leftovers, with the generator fixed so a control can only be written as `Bend=n` and anything unrecognised stops the build:
- **Bend done right:** 252 M0 at every Bend 0–15, plus the low-tone line.
- **M2's form:** VoxProbe 4's exact sequence with and without a Reset in front.
- **Speech near the default Speed:** six sounds at Speed 100–127. That's Man Goes Down's largest remaining error.
- **What Slow scales:** twelve vowels plain and Slow (AY still runs about 20 ms long under Slow).
- **Joins between different sounds:** AW IY AW is +81 ms on the chip against the library.

In short, VoxProbe 12 produced two engine changes (Slow/Fast factors, and the same-sound join), confirmed the default pitch and the upper Bend law, and showed that M2's form is state, not settings. It also split Man Goes Down's error into named causes. Its Bend section was lost to a generator bug, and VoxProbe 13 is built to redo it.