Vox Probe 17 - Definitive Reference Recording

Previous topic - Next topic

0 Members and 2 Guests are viewing this topic.

Andrew Davie

**VoxProbe 17: one AtariVox+, one sitting, every sound three times**

Until now the emulation has been fitted against four recordings of every SpeakJet sound: an AtariVox+ and an original AtariVox, each recorded in two separate sessions. The idea was that four takes average out the chip's own randomness and tell us how close "close" can get.

Looking hard at the differences between those four takes has shown a problem. They don't only differ by the chip's randomness (its noise, its oscillators starting at random phases). They differ in ways that are consistent and real:

- **BE:** the voice bar before the release is about 7 ms longer in one recording session than the other.
- **OWIY:** the attack is three times slower on the original AtariVox than on the AtariVox+.
- **IY:** the AtariVox+ plays it about 10 ms longer than the original.

Averaging four takes like that gives a target that no real chip actually produces: half of one unit's IY length and half of the other's. And measuring "how close is close enough" from the spread between them sets the bar too loose, because the spread includes the differences between units and sessions, not just the chip's own variation.

**The fix is a cleaner reference**: every sound recorded three times, on one AtariVox, in one sitting, through the same recording path. Then:

- **The target is one real chip,** not an average of two units and two sessions.
- **The spread between the three takes is that chip's own randomness and nothing else,** so "as close as another recording of the same chip" becomes a precise, strict standard.

**The preferred unit is the AtariVox+.** It's already the reference for the emulation's levels and spectrum, so recording it again keeps everything measured against the same chip. The original AtariVox stays in use as a check. If we match the AtariVox+ but fall outside what the original does, that sound gets flagged, not fitted.

**What's on the ROM**

- **Running time:** about 4¾ minutes.
- **The title screen** reads VOXPROBE 17 B1.
- **A start reference:** a second and a half of silence, then a 1 kHz tone.
- **Then three rounds:** in each, all 72 speech sounds, then all 55 sound effects, each played on its own.
- **Before every sound,** a short marker: a moment of silence, a two-note tone, and another silence. The tone pair is one the SpeakJet never produces in its own sounds, so every item can be found and identified automatically in the recording.

**If you're recording it**, please use the AtariVox+, keep the same setup from start to finish, and don't stop partway. A single unbroken recording is what makes it work.

Andrew Davie

To clarify; one recording only is needed, from a genuine Atari 2600 unit, using the AtariVox+.

SpiceWare

Sorry, won't be able to get to this until next week.

My sister and her family are coming for the weekend. Usually she stays with me while her son & his family stay with my folks, but they decided to switch it around this time. He has 2 little ones so I've been baby proofing the house. The Atari and capture equipment has already been put away, it's setup on the floor and I didn't want any of it to get stepped on.

I would have held off on stashing the Atari except the last reply on probe 16 had:

Quote from: Andrew Davie on 18 Nov 2026, 07:38 AMNo further probe is planned: the recordings in hand settle what they can, and these stay open.

so I thought we were done.

Andrew Davie

Quote from: SpiceWare on 20 Nov 2026, 03:08 AMSorry, won't be able to get to this until next week.


All good. I've appreciated the fantastic assistance.
I thought it was the last too but Claude keeps throwing spanners in the works.

It's probably good enough now. I just get frustrated when it starts "transcribing" to make things work instead of trying to find the underlying "mechanisms".  If only someone would share the firmware, and all this would be moot.

JetSetIlly

Quote from: Andrew Davie on 20 Nov 2026, 03:13 AMIt's probably good enough now. I just get frustrated when it starts "transcribing" to make things work instead of trying to find the underlying "mechanisms".
The way it's going, I don't think it'll halt the process on its own. I don't think the underlying mechanisms can be found with this method.

I think the emulation is good enough.
https://github.com/JetSetIlly/Gopher2600
@JetSetIlly@mastodon.gamedev.place
@jetsetilly.bsky.social

Andrew Davie

Quote from: JetSetIlly on 20 Nov 2026, 09:01 AM
Quote from: Andrew Davie on 20 Nov 2026, 03:13 AMIt's probably good enough now. I just get frustrated when it starts "transcribing" to make things work instead of trying to find the underlying "mechanisms".
The way it's going, I don't think it'll halt the process on its own. I don't think the underlying mechanisms can be found with this method.

I think the emulation is good enough.


I'll soon put up the latest round of fixes, which are nonetheless significant, and then maybe call a halt to it all and get back to WenHop.