Vox Probe 16

Previous topic - Next topic

0 Members and 1 Guest are viewing this topic.

Andrew Davie

**VoxProbe 16: how the SpeakJet joins its sounds**

The 0.5.x work got the single sounds close: almost every allophone and effect now scores about as close to a real AtariVox as one recording of the chip is to another. Connected speech hasn't caught up. Scored against two real units, our joins cost about 224, while one real AtariVox scored against the other costs 144. Most of that gap is in how one sound hands over to the next.

Over the last few days I tried to close it by rules: "a vowel after MM runs 10 ms shorter", "a slowed nasal runs 10 ms longer", and so on. Each one measured better, and together they bought about 15 %. But it is plainly not how the chip works. No speech chip checks "after MM, cut 10 ms". Two real mechanisms did come out of it and replaced four of those rules:

- **The formants never jump.** Each sound starts where the last voiced sound left them, held through a hiss or a short pause. SE and TT set them to the alveolar position.
- **A vowel moves at a rate.** The move takes longer the further the formants have to travel, stretched by Speed and Slow.

The nasal joins are where it stuck. Tracking F1 and F2 on both units:
- In MM AW and NE AW, the chip hands over from the nasal 10-15 ms before we do.
- In Man Goes Down's "MM EYIY", it hands over exactly when we do.

Three recordings can't say what makes the difference: the vowel, the nasal, or something else. Any rule fitted to them would be one more patch.

So VoxProbe 16 records the joins themselves, and it takes Man Goes Down apart. It runs about 3½ minutes:

1. A reference tone and silence.
2. Every sound used below, alone, so each join can be measured against its own parts from the same session.
3. Nasals into vowels: MM, NE and NGE into IY, AW, UW and EYIY, plain and with Slow on either side.
4. Vowels into nasals: IY, AW and UW into MM and NE, plus MM AW MM and NE IY NE.
5. Vowel to vowel (every pair of IY, EH, AW and UW, EYIY in and out), LE and RR beside vowels, and SE and TT before and after vowels.
6. Speed 90 and 127, and Slow, on two of those joins.
7. Man Goes Down's three phrases as sent, then each one without its Slow, without its Pitch changes and without its Bend. That shows which control causes which part of the timing.
8. Each of Man Goes Down's joins alone, with the Speed, Pitch, Bend and Slow that are in force at it in the game.

**Recording it:** run the ROM on real hardware with an AtariVox, line out straight into the recorder (48 kHz 16-bit, no processing). Press FIRE on the blue screen, and keep recording for 10 seconds after the screen turns green. The title screen reads "VOXPROBE / 16 B1".

Every take helps. Each unit we have so far agrees with the other on these joins to within a few milliseconds, so even one more AtariVox is a strong check.

SpiceWare

6 recordings. 2 additional Stella recordings with the AtariVox USB Interface powered by the hub instead of the monitor.

vox_probe16 117 MB

Andrew Davie

@SpiceWare TY for the most useful recordings.

**VoxProbe 16: what the recordings showed, and the joins rebuilt from them**

Six takes of VoxProbe 16 came in (both AtariVoxes, on a 2600 and through Stella), and they did the job the probe was built for: every kind of join recorded on its own, the solo sounds in the same session, and Man Goes Down's phrases as sent and with each control stripped out. The two units agree with each other to within 5-10 ms on every join, so what they show is the chip, not a recording.

**What they showed**

- A nasal hands over to the vowel after it ~22 ms (MM) or ~13 ms (NE) before its own end alone, whatever vowel comes next. A vowel into a nasal hands over ~15 ms early. The library had "no overlap, unless either is slowed", a rule fitted on Man Goes Down's phrase 1, and it was wrong in both directions.
- Speed 127 is exact. Speed 90 stretches a vowel pair x1.65-1.68; the library used the effects' Speed law for speech below 114 and gave x1.84.
- Slow on an isolated pair fits the library within 10 ms. In the phrases at Speed 108 and 118 the chip's slowed vowels come out ~30 ms shorter than the library's x1.29 x Speed. That one is still open.
- Man Goes Down with every Slow stripped: phrase 1 1305 ms on the chip, 1280 on ours; phrase 2 1520 / 1510; phrase 4 790 / 820. Without its Pitch or Bend commands the chip's lengths change by at most 15 ms, so neither carries timing.
- NGE's and MM's second formants are their 2249 and 1430 Hz lines (the ones that glide into the next vowel's F2); the library had those as a third formant beside a near-silent second one.

**What I did with it**

The library had accumulated thirteen join rules fitted to Man Goes Down through a phrase score ("IY 5 ms shorter after a sound", "a vowel 10 ms shorter after MM", and so on). VoxProbe 16 showed they were compensating for each other: putting the correct nasal overlap in scored *worse* against them. So all thirteen are gone, the phrase score is retired, and the joins are now judged on VoxProbe 16's own items, by the timing of each formant move against both units' tracks and by the levels in sixteen bands. In their place, six measured mechanisms: the nasal overlaps above; LE and LO joining like nasals (WW and RR stay on the vowel rule); the speech length law below Speed 114, with a vowel's own; a stop after a pause closing 15 ms longer, not 30; and the NGE / MM formant order.

                                              0.5.4      now
joins, VoxProbe 16's 76 recorded join items   4.57 dB    4.07 dB   (one real AtariVox against the other: 2.77)
joins, VoxProbe 13's 87 pairs, length error   18.1 ms    13.5 ms   (not fitted on: an independent check)
Man Goes Down phrase 1 (units 1490 ms)        1480       1475
Man Goes Down phrase 2 (units 1690 / 1685)    1725       1725
Man Goes Down phrase 4 (units 930 / 935)      945        930
solo allophones and effects                   unchanged, except B6 1.06 -> 1.02 and B5 0.88 -> 0.86

So: joins 11 % closer to a real unit than 0.5.4, 25 % closer on the independent set, phrases 1 and 4 within 15 ms of both chips, phrase 2 still 37 ms long, and no per-phrase patches left in the engine.

**One effect, from a picture**

Andrew pointed out that 226 B6's spectrum had a mountain on the chip between 2.2 and 3.1 kHz where the library had a valley. It turned out to be the chip's tone switches: when an oscillator leaps to a new frequency it plays the midpoint of the old and new for about 1.5 ms first, and B6 switches between 1.8 and 3.5 kHz five times. A constant-rate glide drew a slope; the midpoint draws the chip's hump. B6's "4.5 kHz tone" was the 3.5 kHz tone's image at the chip's 8026 Hz sample rate (the two always add up to the rate, on every take), so it and two more like it in C2 and A4 came out of the tables. B6 1.06 -> 1.02 against a real take's 1.0.

**Still off**

Phrase 2's 37 ms (the Slow-at-Speed question above, and BE after a pause, which on the chip is a 20 ms bar 25 ms after the hiss then 95 ms of quiet); PO after a nasal; AY IY and AW IH AW about 30 ms long; IY SE ending 15 ms early. All measured, none fitted. No further probe is planned: the recordings in hand settle what they can, and these stay open.

This is 0.5.5 on the main branches, with the technical manual at 1.24. Thanks to everyone who recorded VoxProbe 16.