Take 2, then Take 3 -- new emulation - request for recordings

Previous topic - Next topic

0 Members and 2 Guests are viewing this topic.

SpiceWare

Wonder if that changes are due to the RF Noise Suppressor - it's one of these though shaped as a 3 cm cube.

I noticed a like and clicked to see it but got a "you're banned" popup.

SpiceWare

More weirdness - while I was posting my reply the forum said something along the lines of "there's a new reply you should review it".  It was from Andrew.  After I clicked Post the forum showed errors instead of posts.  After restarting Safari the forum shows the posts again, but the new reply from Andrew is missing.

Andrew's reply was something about how values were fetched and applied.

Andrew Davie

Sorry really bad timing.  I made a post about a finding that may or may not be true, but I wanted to have it properly verified and realised I posted too early, in my excitement. So I deleted it -- and the forum is now unhappy because the post isn't there.  SHould be OK now but if you have issues, log out and in again and that's fixed issues for me when I've had similar.

SpiceWare

Logged out, restarted Safari, logged in. 

Still get the "sorry you're banned" popup if I click the "1 person likes this" on my reply.

Andrew Davie

Quote from: SpiceWare on 01 Nov 2026, 03:05 AMLogged out, restarted Safari, logged in. 

Still get the "sorry you're banned" popup if I click the "1 person likes this" on my reply.

Ah, that's another issue. I get that too.
I have a catch-all on the like link, sorry --- there's a Russian bot-net or whatever you call it polling the like links thousands of times a day.  If you *really* want to like something, then copy the URL and witch around two of the terms for example these two... "ltype=msg;sa=like" to "sa=like;ltype=msg".  That will bypass the bot-blocker URL thing. I really should update the forum code but for now it is what it is.  Bloody annoying assholes driving the bots, though :(

Andrew Davie

VoxProbe 8 went well: both units recorded cleanly, and it's already making a measurable difference where it counts, in real game speech. Here's Man Goes Down's phrase-length error before and after the changes VoxProbe 8 drove (both units):

phrase   length off (ms)            P score
         fitted  before  now       before  now
P1         55     160    175        2.01   2.05   unchanged: Slow-heavy, not solved yet
P2         50     110     40        1.82   1.26   now closer than fitted
P3         10      50     25        1.98   1.72   halved (effects at Speed 62)
P4         15      35     25        1.53   1.40

Overall, Man Goes Down went from P 1.84 to 1.62 and E 7.93 to 7.33, closing about half its gap to the fitted build. P2 now beats fitted outright. On all 1,382 recordings the library leads the fitted build by P 1.046 to 1.115.

Where the gain came from:

Pauses shortened to the chip's measured gaps: most of P2's improvement.
The effects' Speed law refitted: P3.
Read-ahead discovered: the chip acts on a Speed or Volume that follows a sound while that sound is still playing. That's a real mechanism no single-sound recording could have shown. It's neutral on the existing scores, but correct, and it explains odd results across 552 older recordings.
Two wrong conclusions corrected: the "curved Speed law" and "speech compresses with Volume" were artefacts of that same read-ahead.
What it didn't settle:

P1 is still 175 ms off. It's dominated by Slow between sounds, and VoxProbe 8's layout (controls right next to the items under test) couldn't isolate that. The next probe needs pauses between each control and its test item.
Nasal pulse shape: a real clue, the chip's mid band decaying more slowly. Testing a fix now.
Still to come: round 3's anneals now train on VoxProbe 8's 164 items too, so a little more should follow. The overnight results will show how much.

SpiceWare

Looks like I can like ok, just did on your two replies. 

The "banned" popup occurs when I tried to see who liked my replies; though, I figured out I can see that in the Alerts history:

Andrew Davie

**AtariVox / SpeakJet emulation: what VoxProbe 8 showed**

Overall, this was a huge win for the emulation!

VoxProbe 8 has been recorded on both units, an original AtariVox and an AtariVox+, and the recordings have been worked through. Thanks to @SpiceWare who recorded it, and for recording it a second time when the first takes clipped. Here's what it showed about the chip, and what changed in the emulation (the C library and the Go version used by Gopher2600) as a result.

**What the chip does**

- **It reads ahead.** The chip takes the command that follows a sound early, while that sound is still finishing: at an effect's closing fade, otherwise in the sound's last 30 ms or so. A Speed or Volume command read this way changes the rest of the sound already playing. Pitch is not read ahead. This cleared up two things I had wrong from earlier probes. One was a "curved" Speed law for the effects; the other was speech that seemed to get shorter with Volume. Both were the read-ahead acting on test sounds that had a control command straight after them. Lesson for future probes: put a pause between a test sound and the next command.
- **Pauses.** P1 leaves a 90 ms gap and P4 15 ms.
- **The effects' Speed.** Lengths scale as (147 - Speed) / 33. Once the read-ahead is accounted for, that holds within 5-7 % from Speed 20 to 127.
- **Volume** is linear.
- **The output stage.** Tones set directly through the serial control protocol, from 150 to 3950 Hz, fall off exactly as two simple low-pass stages (near 1.1 and 2.4 kHz) predict, within 0.4 dB on both units.
- **The hiss under the effects.** A tone set directly is almost clean, but the built-in effects carry 35-45 dB more background noise under their tones. Rewriting the volume register while a tone plays doesn't add any, so the noise comes from the way the chip plays its effects. The noise grows with the tone's pitch up to about 1.5 kHz, then levels off.
- **A puzzle.** A directly set tone has faint 2nd and 3rd harmonics whose level stays the same when that oscillator's volume is turned down 32 dB. They do follow the master volume. So a quiet oscillator has relatively strong harmonics: at volume 1 they are only 17 dB under the tone. The two units differ by up to 20 dB here. It isn't simple distortion, and it is probably what gives the DTMF digits their faint extra tones, such as the sum of the row and column frequencies at about -44 dB. That's a question for a later probe.
- **Still open:** how Slow (and Fast) stretch sounds inside a phrase. On their own, the fricatives do stretch; between sounds it isn't settled. It is the main timing error left in Man Goes Down's long phrase.

**What changed in the emulation**

- Read-ahead, the pauses and the effects' Speed law, as measured.
- The glides ED, OHIH, EHLL, WH, EYIY, OWIY and CH were 10-20 ms too long; fixed.
- Five effects were rebuilt from their own clean recordings: 208 R8, 210 A0, 211 A1, 212 A2 and 235 C5.
- The hiss under the effects now levels off above 1.5 kHz, as on the chip.
- **The output stage, the biggest single gain.** The emulation used to approximate the AtariVox's output filter with a digital filter applied to the chip's samples. That approximation was too steep near the top. The chip holds each sample until the next, which puts faint copies (images) of every sound above 4 kHz, and those images came out 1-8 dB too quiet between 5 and 14 kHz. The filter is now worked out exactly as the analogue circuit it is, plus one small extra term near 14 kHz that the images called for. Every sound improved; not one got worse.

**How close it is now**

Every recorded item, now 1382 of them including VoxProbe 8's, is scored against the real chips on a scale where 0 means "as close as a second real chip". "previous" is the emulation before the rewrite (the "fitted" implementation), scored the same way:

                       previous    now
all                      1.12      0.93
single sounds            1.03      0.76
connected speech         1.08      0.91
VoxProbe 7 items         1.29      1.01
sound effects            1.18      1.05
Man Goes Down            1.40      1.56

That's 17 % closer overall and 26 % on single sounds. Man Goes Down's long phrases are the one place still behind the previous version, but the gap has more than halved since the last update. Most of what's left there is Slow inside phrases.

As far as object size, measured the same way as the other two versions:

                          first Go    fitted      now
Go package (compiled)    78.8 KB    140.4 KB    22.0 KB
  of which code          70.7 KB    109.8 KB    16.9 KB
C library (compiled)        -        91.7 KB    18.9 KB
sound tables              8.1 KB*      18 KB*    2.7 KB

*plus data its startup code builds



I plan to release the new implementation sometime later today.


Andrew Davie

**What the scores mean**

Each recording is compared with the emulation on several measures:
- how long it lasts;
- its loudness over time;
- how its spectrum changes over time;
- its overall spectrum;
- for speech, its pitch and its formant peaks; for sound effects, its tones and how pure they are.

Each difference is then divided by how much the two real AtariVoxes differ from each other on that same measure. Anything within that range counts as zero: the emulation is as close as a second real chip would be. Beyond it, every doubling of the difference adds 1. So a measure scores 1 when the emulation is twice as far from the chip as the two chips are from each other, 2 at four times, and 3 at eight times. An item's P is the root-mean-square of its measures, and the tables give the average over all items.

So 1.0 is not a threshold or a pass mark. It's roughly the point where the emulation is, on average, twice as far from a real chip as the two real chips are from each other. The scale is continuous and logarithmic: every 0.1 off P means the gap has shrunk by about 7 %, wherever on the scale that happens. The other end, 0, is the floor. It doesn't mean identical, only that no measure differs by more than the two real units differ from each other.