Vox Probe 11

Previous topic - Next topic

0 Members and 1 Guest are viewing this topic.

Andrew Davie

**AtariVox / SpeakJet emulation: VoxProbe 11, please record**

Here is the eleventh probe ROM. Probe 10 was meant to be the last, but the work since has left a short list of questions only the chip can answer. These come from hand-drawing the effects' structure over the recordings: where the noise "bed" under the effects comes from, what the striped noise columns in C2 and B7 follow, and how repeatable the chip is against itself. It runs about 3 min 10 s. As before, I'd like it from both units if possible: the original AtariVox and the AtariVox+.

**Recording.** The same setup as VoxProbe 7 to 10: 48 kHz, 16-bit, no input processing (no noise reduction, no AGC). The probe starts with 2 s of silence and then a 1000 Hz tone; set the level so that tone sits at about -10 dBFS and nothing clips. Start recording, press FIRE (or RESET) to start the sequence, and leave it running until the closing tone and silence. Each section is announced by touch-tone digits, so the recording can be matched back to the ROM. The title screen reads VOXPROBE 11 B1; please say which build you recorded.

The first two sections and the last drive the synthesizer directly through the serial protocol. The emulation now supports that protocol, so in Gopher2600 (on the next upate hopefully) they play as tones; emulators running an older SpeakJet library stay silent there.

**What it tests**

1  reference           2 s of silence, the 1000 Hz tone
 2  the bed             one plain tone at 300 to 3950 Hz, at four volumes, then two and three tones together
                        (does the noise under the effects follow the tone's frequency and level, and add up?)
 3  effects with beds   R1, C5 and B8, twice each, to set against section 2
 4  banded columns      C2 and B7 at Pitch 60, 88 and 150, and at Speed 60 (do the stripes follow Pitch,
                        do the columns stretch with Speed?)
 5  low Pitch           C3 and C6 at Pitch 40, 50, 60, 70 and 88
 6  M2                  at Pitch 60, 88 and 150 (its sidebands against Pitch)
 7  repeatability       AW, SE, C5, A8 and C3, six takes each, each sent alone: the chip against itself, which
                        is the closest any emulation can get
 8  connected speech    the four Man Goes Down phrases, three takes each from a reset, then all four in a row
                        twice
 9  effect onsets       B7 and A8 alone after silence, then each straight after A0 (does the opening column
                        belong to the effect, or to what came before?)
10  reference again     the 1000 Hz tone, 2 s of silence

ROM: vox_probe11.bin (4 KB, NTSC). Thanks again for the recordings; everything the emulation knows about the chip has come from them.

SpiceWare

Probe 11 results from AtariVox+ Vectrex Enhanced on real Atari

Would it be helpful to recapture any of the prior runs now that we have the volume dialed in at tick 3.5 and can make clearer recordings using Stella?

SpiceWare

Probe 11 results from original AtariVox on real Atari

SpiceWare

Probe 11 results from AtariVox+ Vectrex Enhanced on Stella

SpiceWare

Probe 11 results from original AtariVox on Stella

SpiceWare

This time for the Stella recordings the AtariVox Interface and the Audio Capture device are both plugged into the monitor.  My thinking is if they're powered by the same source it might reduce noise in the recording.

Andrew Davie

Quote from: SpiceWare on 04 Nov 2026, 07:00 AMProbe 11 results from AtariVox+ Vectrex Enhanced on real Atari

Would it be helpful to recapture any of the prior runs now that we have the volume dialed in at tick 3.5 and can make clearer recordings using Stella?


Immensely.  A like--for-like recording with (say) Stella would lower the noise level significantly, and allow averaging of many of the samples that are duplicated and lead, I think, to a fair gain in accuracy overall.

Andrew Davie

**AtariVox / SpeakJet emulation: VoxProbe 11 results**

Thanks for the VoxProbe 11 recordings: four takes this time, both units directly and both units through Stella. All 87 items were found and placed on every take. Here is what they showed, and why the emulation has not got much better from them.

**What it answered**

question                         answer
the "bed" under the effects      not from the oscillators. A plain tone set through the registers has no
                                 broadband noise under it at any frequency (300-3950 Hz) or volume; near the
                                 4 kHz half rate it carries a few discrete spurious lines instead. The effects'
                                 bed (20-40 dB over the background) comes from how the chip runs an effect
register volume                  linear: 6 dB per doubling (4, 8, 16, 31)
B7's striped noise column        stripes ~500 Hz apart at Pitch 60, 88 and 150: they don't follow Pitch;
                                 the column comes later at Speed 60
C2's stripes                     still too faint to measure automatically
the chip against itself          six takes of AW SE C5 A8 C3: spectrum differences of 0.4-1.1 dB (SE 2.3).
                                 That is the closest any emulation can get, and it is now measured
C6 at low Pitch                  close; its pulse runs 1-3 % slow at Pitch 40-70 in the emulation
C3 at Pitch 40-50                badly wrong: the chip keeps broad energy where the emulation has only its tone
M2                               two forms. On VoxRecord it has a low formant (F1); on VoxProbe 4 and on all six
                                 VoxProbe 11 takes that formant is silent. The emulation was fitted to VoxRecord,
                                 so it is ~10 dB off on 14 of the 16 recordings of M2
controls persist                 Speed, Pitch and Bend carry across chunks until a Reset: Man Goes Down's last
                                 settings carried into the closing sections, shifted their touch-tones and
                                 silenced the final reference tone
Man Goes Down timing             measured sound by sound for the first time, from clean takes: the chip runs
                                 180-240 ms longer than the emulation on phrases 2-4, mostly on sounds after a
                                 Slow code and on the death sound at Speed 62

**Why it hasn't helped much**

- Most of its questions asked *where* something comes from, and the answers were mostly "not from here". The bed isn't the oscillator; the B7 stripes don't follow Pitch. Those answers rule out the fixes I had planned (a bed tied to the tone, a stripe spacing tied to Pitch) but don't hand over a replacement. The change to the stripe rule is in, and it's correct, but no shipped sound uses it at Pitch 88, so nothing scores differently.
- The biggest single finding, M2's two forms, can't be used yet. Moving the emulation to the common form would improve 14 recordings and spoil the 2 it matches now, and the rule that picks the form needs a recording to find. The quality gate rightly refuses the change until then.
- The tests were too small and too scattered: one sound at three settings, one question per section. A probe that answers rules has to set one thing at a time against a fixed reference. VoxProbe 11 mostly gave single examples.
- The losses that matter most weren't in it. Bend 0 and 15 and Man Goes Down's timing cost more score than everything VoxProbe 11 tested together. The Bend tests and the default-pitch test were planned for an earlier probe and never recorded.

**What it set up**

The Man Goes Down timing is the useful part: for the first time we know, sound by sound, which lengths are wrong and by how much. That shaped VoxProbe 12. It takes the phrases apart one control at a time, measures Slow and Fast properly, measures the effects' Speed law and Bend across their ranges, and puts M2 in nine settings to find what decides its form. VoxProbe 12 is posted separately.

SpiceWare

#8
Quote from: Andrew Davie on 04 Nov 2026, 12:31 PM
Quote from: SpiceWare on 04 Nov 2026, 07:00 AMWould it be helpful to recapture any of the prior runs...

Immensely.

On my lunch break I hacked all the older ROMS to show the # and have recorded 1-6 using Stella and the original AtariVox.

Time to get back to work, I'll finish the re-recordings (and the new #12) this evening. I'll zip up the results and put them on my Dropbox.