Vox Probe 13

Previous topic - Next topic

0 Members and 1 Guest are viewing this topic.

Andrew Davie

**AtariVox / SpeakJet emulation: VoxProbe 13, please record**

Here is the thirteenth probe ROM. (Earlier builds went up briefly and were withdrawn before anyone downloaded them; this is build 5.)

VoxProbe 12's recordings explained a good part of Man Goes Down's timing. Below the default Speed, speech stretches the way the effects do, and the library had a gentler rule for speech. They also showed that Slow stays on its own sound and doesn't carry to the next. Working through "man goes down" against both recordings then turned up three things the library didn't do at all. When a voiced stop follows a nasal, the chip goes almost silent for 20-30 ms (the G in "goes"). GO's release takes different formants when a vowel follows. And a stop running into a vowel lost too much of its release. With those in, "goes" sounds like "goes" rather than "gone".

Going through every two-sound pair the earlier probes recorded on both units then showed how the chip joins sounds. The further apart two vowels' formants are, the longer the chip takes to get from one to the other: AW into IY runs about 45 ms longer than the library had it, AW into UW slightly shorter. A voiced stop into a vowel with a similar F2 (BE EH, DO AYWW) runs shorter. VoxProbe 1's pause items gave the true lengths of P0, P3, P5 and P6 (P3, P5 and P6 are 30-50 ms under the manual's figures). All three Man Goes Down phrases now match both units' lengths to within about 20 ms. But several of these rules rest on a handful of items, and VoxProbe 13 is built to check them across many sounds.

It also redoes the Bend tests: VoxProbe 12 sent Bend 0-8 as effect codes by mistake, so only Bend 10, 12 and 15 were recorded right. It runs about 3 min 45 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 12: 48 kHz, 16-bit, no input processing (no noise reduction, no AGC). The probe starts with 1.5 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, and every item starts from a reset. The title screen reads VOXPROBE 13 B5; please say which build you recorded.

**What it tests**

1  reference           1.5 s of silence, the 1000 Hz tone
 2  Bend                B4's steady 1825 Hz tone at every Bend 0-15; AW at Bend 5, 6 and 7 (Man Goes Down plays
                        at 6 and 7); A5 at Bend 1 at two Pitches (its second tone at Bend 0-2)
 3  Speed near 114      AY IYUW EYIY AW OWWW UX alone at Speed 100, 108, 114, 118, 122 and 127: the speech Speed
                        law either side of the default
 4  what Slow scales    six vowels plain and Slow: does Slow stretch the whole sound or only part of it
 5  joins               vowel to vowel (AW IY AW and others), vowel to nasal and liquid both ways (IY MM, AW LE,
                        AW RR, IY NE, AW WW), and vowel into a gliding vowel (AW AYWW, IY OWWW)
 6  into a voiced stop  MM NE NO NGE, LO RR, and OW EH AW before each of GO DO BO, then OW; NE before KO TT PO;
                        NE GO OW at Speed 90, 108 and 127, and with Slow and Fast on GO
 7  a stop's release    GO GE DO DE BO BE alone and before IY AY OW UW EH; DO AYWW and BE EH with and without
                        Slow on the vowel; Fast on BE and DO
 8  the death sound     C1 as Man Goes Down sends it (Speed 62, Pitch 212), then twice at Speed 0
    slowed
 9  pauses              P2 and P6 between two AWs at Speed 100, 116 and 127, at Pitch 55 and 110, at Bend 1 and 4;
                        IY P2 IY, UX P1 EK, EH P6 DE
10  unvoiced stops      SE, NE and AW before TT; AW before KO and PO; Fast and Slow on TT and on TH; a slowed
    and hisses          IYUW and AYWW before a hiss, with and without a Slow on it; SE SE
11  Man Goes Down       phrase 1 ("man goes down") exactly as the game sends it
12  reference again     the 1000 Hz tone, 1.5 s of silence

The slowed effects planned for this probe have moved to VoxProbe 14, to keep this one under four minutes.

Thanks again for the recordings; everything the emulation knows about the chip has come from them.

SpiceWare

All 4 recordings:

vox_probe13 153 MB

I'm almost out of room on DropBox so will start deleting prior attachments when I do #14. I'll start with the video recordings I linked to in the Wen Hop forum.

Andrew Davie

Quote from: SpiceWare on 08 Nov 2026, 02:45 AMAll 4 recordings:

vox_probe13 153 MB

I'm almost out of room on DropBox so will start deleting prior attachments when I do #14. I'll start with the video recordings I linked to in the Wen Hop forum.

TY.  Results will be posted tomorrow, as I have a busy day ahead.
I have copies of all the probe recordings safely archived, so they can be zapped.
Cheers
A

Andrew Davie

#3
**AtariVox / SpeakJet emulation 0.4.3: what VoxProbe 13 taught us**

Thanks to everyone who recorded VoxProbe 13 on both units. Those recordings drive this release: version 0.4.3 of the C library (andrew-davie/SpeakJet) and the matching Go library.

**What changed, in one line.** No individual sound has changed: every sound played alone is sample for sample what 0.4.2 played. What changed is how the chip joins sounds, pauses and stretches them, which is where speech lives.

**What the probe showed**

- **Speed is not one straight line.** Below the default (114) everything stretches the way the sound effects do. Just above 114 the chip steps down: at Speed 118 the vowels already run x0.80, where a straight line from 114 gives x0.88. By 127 both agree at about x0.61. Pauses follow the same law. The consonants don't seem to shrink at all at 118. One item at 127 disagrees, so that part is marked as an assumption.
- **Pauses.** P2 is 188 ms, not 200. With that, the second AW of "AW P2 AW" lands within about 10 ms of both units at Speed 100, 114, 116 and 127.
- **Unvoiced stops are preceded by silence.** Before TT, KO and PO, both chips stop the vowel and go completely quiet for 20-40 ms before the burst. The library faded the vowel all the way into the burst. This was the biggest single structural difference VoxProbe 13 showed.
- **The chip reads ahead.** A Slow placed on a TT (or on a GO after NE) makes the sound before it about 40 ms longer, not the stop itself.
- **Joins between sounds.** NE and NO run longer into GO, KO and PO than into TT. A nasal joins a plain vowel with no overlap. The liquids (LE LO WW RR) join vowels the way vowels join each other. A stop runs into a vowel longer when F2 has further to travel.
- **After a hiss, the next sound starts clean.** No formant sweep down from SE's 4-6 kHz into the next sound, and no second low thump when one SE runs into another. That removes the dark blob in Man Goes Down's double SE and the dark streak where it meets UX.
- **A Slow on the second of two identical sounds breaks the Repeat.** Slow SE Slow SE plays two whole SEs, not one SE and a copy.
- **GE glides into a following vowel.** Its F2 falls from about 2000 Hz into the vowel's over 40-60 ms.

**Results**

```
                            unit A  unit B    0.4.2    0.4.3
Man Goes Down phrase 1      1495    1502    1502    1485 ms
Man Goes Down phrase 2      1695    1700    1700    1700 ms
Man Goes Down phrase 4        942      945      925      940 ms

timing inside the phrases (ms off the units, on average)
phrase 1                                      14.7    12.4
phrase 2                                      16.7      9.9
phrase 4                                        9.9      6.8

VoxProbe 13, length error per section, summed  120.9    55.0 ms
```

On the VoxProbe 13 items that break into separate runs of sound (a vowel, a gap, a burst, a vowel...), every run now matches the chips', and each run's start and end lands on average 8 ms from both units.

**Pictures.** graphs/compare/phrase_mgd_p1/p2/p4_before_after.png show unit A, unit B, 0.4.2 and 0.4.3 for Man Goes Down. phrase_vp13_*.png show single VoxProbe 13 items the same way, with the red line at the end of the sound.

**Two lessons about measuring**

- **A quiet hiss after a vowel can't be timed by loudness alone.** SE after AW sits under the recording's hum, so a plain "where does the sound end" measure stops at the vowel. AW SE SE read 128 ms when it really runs about 290. Lengths are now taken band by band, so a hiss counts wherever it rises above that band's own background.
- **Check every item's placement against the other unit.** In one take a stray extra sound shifted five items by one place. The item count still matched, so nothing flagged it, and for a while it looked as if the two units disagreed at Speed 116. They don't.

**Marked as assumptions.** Several of these rules rest on a handful of items: the consonants above Speed 114, the NE-into-GO run-on, KO after a nasal, the TH lengthening, and the GE glide. They're marked ASSUMPTION in the source, and the technical manual (version 1.13, section 6.9) lists every rule and what it was fitted on.

**Still open.** AW IH AW and AY IY run 30-40 ms long. IY after a voiced stop and RR after a sound run 15-30 ms short. In phrase 4 the slowed IYUW ends about 25 ms early. There is also the consonant question at Speeds 118-127. VoxProbe 14 is being planned around exactly these, with every hiss placed so it can be timed.

Thanks again: everything the emulation knows about the chip has come from these recordings.