Complete 72 allophone binary for testing

Previous topic - Next topic

0 Members and 1 Guest are viewing this topic.

Andrew Davie

I'm getting too keen to get PRs sent to the Go library.
There are some oddities in the latest one and I'll fix those then slow down a bit, I think.

SpiceWare

#61
Quote from: Andrew Davie on 28 Oct 2026, 12:52 PMI think we're doing OK with the current process, actually.

Sounds good, I'll stick with the current capture device.

Quoteset it lower than last time, peaks around −6 dBFS, because loud effects clipped before.

I've lowered the volume control to 3 ticks after MIN.

Probe 5 results from AtariVox+ Vectrex Enhanced


SpiceWare

Probe 5 results from original AtariVox

Andrew Davie

Quote from: SpiceWare on 29 Oct 2026, 02:01 AMProbe 5 results from original AtariVox

TYVM for those.
And with those, I think we are done.

Andrew Davie

#64
With this installment I think we're done.
Barring any jarring bugs, of course.




Fifth VoxProbe results, and a new library push.

Thanks to the two VoxProbe 5 recordings (both AtariVoxes), three things are settled:

1. What's waiting in the SpeakJet's buffer makes no difference to a sound's length. Each of ten sounds sent streamed, sent alone with nothing after it, and sent alone followed by pauses comes out the same length on both units, within 5 ms. So the 15% difference between the VoxRecord and VoxProbe sessions is between sessions, not something the chip does to isolated sounds.

2. Repeats have two lengths. The second play of a sound runs longer than the copies after it, and the later copies are mostly the manual's 70 ms. AW: 130 ms, then 72, then 70. "AW AW" as two codes is the same waveform as repeat 2. Stops are different again: EB repeated is a burst every 45 ms, with the gap in the closure, so stretching a copy evenly was wrong.

3. The chip lengthens a sound before any command byte, not only before a pause. AW before a Pitch or a Delay is as long as before a pause, and the same length whatever the pause code.

The library has all three now, plus the earlier work from today: C8/C9 as a swept noise band, DH/ZH/JH voiced a little faster than the vowels, SE and SH as broad hiss instead of two thin lines, and a start pulse at very low pitch. Scored against both chips, over the 752 recordings the fits use, the library went from 1.22 to 1.18 (on a stricter score than before, where 0 means "as close as a second chip"); over 1383 recordings it was never fitted on, from 1.36 to 1.18.


Andrew Davie

Graph of emulation accuracy, commit by commit...