Limits of Knowledge...

Previous topic - Next topic

0 Members and 1 Guest are viewing this topic.

Andrew Davie

**Why P won't go to 0**

P compares how far the library is from a chip with how far one chip is from the other. Our two AtariVoxes run the same firmware and differ only in clock (×0.997) and analog noise. So "0" means "as close as two identical chips", which is a very tight target. Current median error ratios:

              spectrum  tf    tone  env
effects        2.2x    1.75x  1.8x  1.6x
allophones    1.8x    ~1.0-1.2x (others)

The spectrum is most of the gap.

**What the shortfall is**

- **No global fix is left.** The library-minus-chip mean spectrum is flat within ±1.5 dB for effects, with the sign split evenly across bands, so no shared filter is missing. I tried four engine-wide fixes and all made P worse:
  - Bend 0-2 gain
  - a voice-pulse offset
  - the voicing-envelope drop-in
  - an output high-pass
- **The gap is per sound, and some of it is structure, not tuning:**
  - the Bend 0 / high-Pitch mechanisms (no wrap, period rounding at Pitch 200)
  - the pulsed openings on 235 C5 and R0-R3
  - nasals and voiced stops (BE DO NE) about 3 dB short at 300-420 Hz
  - the excluded set: R5 R8 R9 C8 C9 B0 and D0-D11

  Annealing can't close a structure gap. It just games E and P around it.
- **The voice source is wrong, and we can't see the right one.** The chip's pitch pulse rises over 1.3-2.6 ms and still keeps its low harmonics. In our engine the step rise is what makes those harmonics. So the chip gets its low end from something we haven't identified. Every voiced table is tuned around the step, so the source can't be changed by itself: it means changing it and refitting all the voiced sounds together, and that refit hasn't been run.

**What the recordings can't tell us**

1. **Output only.** We hear the sum of source, formant filters, noise and envelope after the AtariVox analog path. Different internal splits give the same output, so the fit can't tell a wrong mechanism from a wrong number.
2. **Unknown analog path.** The coupling, the reconstruction filter and the record chain colour everything. The low-band pulse shape shows that, but it isn't a simple high-pass.
3. **Contamination.** There's 60 Hz hum and its harmonics in every recording, and noise sounds are random. That puts a floor under any per-sound comparison, and raw averages give false shortfalls.
4. **Only two units.** That gives a tight, maybe unrepresentative "chip-to-chip" yardstick, and nothing to show which differences come from the firmware and which from the part.
5. **Context effects.** The chip reads ahead, so a sound's tail depends on what comes next. Solo takes miss some of the behaviour and connected takes mix it up.
6. **Few takes per condition.** The searches end up fitting noise in the recordings.

**What could fix it**

- **Tap the DAC/PWM pin directly**, ahead of the AtariVox filtering, on both units. That takes out the analog path, and the hum if we keep the ground clean. This is the biggest gain for understanding the voice source.
- **Probes that isolate one part at a time.** Use SCP register writes to drive single oscillators and filters with the rest off. That measures source, filter and envelope separately instead of inferring them from the mix. It's VoxProbe 10 or later material, alongside the super-slow and low-Speed C1 ideas already listed.
- **Repeat takes per condition (×8 or more) and average them**, which removes the random part of the noise sounds and gives real confidence bands.
- **A third unit**, if we can borrow one, for a more honest yardstick.
- **Order of work:** fix the structure on the listed sounds first, then refit the voice source together with the voiced tables. Annealing alone won't move P much further.

SpiceWare

Quote from: Andrew Davie on 02 Nov 2026, 03:58 PM- **Tap the DAC/PWM pin directly**, ahead of the AtariVox filtering, on both units. That takes out the analog path, and the hum if we keep the ground clean. This is the biggest gain for understanding the voice source.

What's needed to do this?

Also might having the computer drive the AtariVox eliminate some of the noise, such as the 15.7 kHz?  I have some adapters somewhere, though I do recall having a flow control issue.

I'm getting ready to head into town for my monthly work meeting, so will be offline for the next 6+ hours.

Andrew Davie

Quote from: SpiceWare on 03 Nov 2026, 03:07 AM
Quote from: Andrew Davie on 02 Nov 2026, 03:58 PM- **Tap the DAC/PWM pin directly**, ahead of the AtariVox filtering, on both units. That takes out the analog path, and the hum if we keep the ground clean. This is the biggest gain for understanding the voice source.

What's needed to do this?

Also might having the computer drive the AtariVox eliminate some of the noise, such as the 15.7 kHz?  I have some adapters somewhere, though I do recall having a flow control issue.

I'm getting ready to head into town for my monthly work meeting, so will be offline for the next 6+ hours.



The very most desirable thing to do would be for someone to pop the SPeakJet chip into a PIC chip reader and see if the firmware is protected or not. If not, then apparently we could dump it and do a *perfect* emulation.  The problem is, only you have put up your hand to assist with recordings, and as my forum is rather obscure there are not many people seeing this work.



SpiceWare

I pinged Nathan about this forum yesterday suggesting he post about it in his AtariVox club. Maybe @JetSetIlly could post something about Gopher2600's recent update to support it in that club.

Heading out now, will check back in when I return from my meeting.

alex_79

Quote from: Andrew Davie on 03 Nov 2026, 03:17 AMThe very most desirable thing to do would be for someone to pop the SPeakJet chip into a PIC chip reader and see if the firmware is protected or not.
The firmware is definitely protected. I remember reading a few years ago that Kevin Horton managed to dump it (and some cores he wrote for the analogue fpga systems did include AtariVox support, if the dumps were provided by the user), but that requires some advanced, and usually destructive for the chip, techniques.