Complete 72 allophone binary for testing

Previous topic - Next topic

0 Members and 1 Guest are viewing this topic.

Andrew Davie

**Call for recordings: VoxProbe 2 (AtariVox owners wanted!)**

The first VoxProbe reshaped the SpeakJet emulator: we now know how the chip voices sound, that effects follow Speed and Pitch, how Repeat works, and that the mysterious "extra sound" is the chip resetting itself.

VoxProbe 2 goes after what's left, and the biggest item is **Pitch**. The chip's frequency for a given Pitch setting doesn't follow any formula we've tried, and that's now the largest remaining error in high-pitched speech. This ROM plays every Pitch value from 0 to 254, so we can read the chip's table directly.

It also tests:
- whether the chip's ~8 ms voicing fade is the same for every sound;
- what the "fill" under the robot sounds (R4, R6–R9) is;
- the exact frequencies of all 12 DTMF tones;
- what triggers that reset.

**You'll need:** a 2600 (or 7800) with a Harmony or PlusCart, an AtariVox or AtariVox+ in the **right** port, a joystick in the **left** port (or use RESET), and something to record the AtariVox's audio output.

**To record:**
1. Connect the AtariVox audio out straight to your recorder's line in. Use WAV or AIFF, 44.1 or 48 kHz, 16-bit or better.
2. Turn **off** noise reduction, auto-gain, compression and normalising. Please leave the faint background whine in: it's the 2600's line rate, and it lets us correct for your recorder's clock.
3. Set the level so the loudest vowels just reach full scale.
4. Start recording, then run the ROM. The screen goes **blue**.
5. Wait about **5 seconds**, then press **FIRE** (or RESET). The screen goes **red** while it sends.
6. When it turns **green**, keep recording **15 more seconds**, then stop.

The full run takes about 4 minutes 20 seconds. Please send the file untouched, with no trimming or editing. It also helps to say which AtariVox (original or +) and which cart you used. Recordings from **different AtariVox units** are especially welcome, since they show which quirks belong to the SpeakJet itself and which to one particular unit.

Thanks! Every recording so far has turned guesses into measurements.

SpiceWare

probe2 results.  Still AtariVox+, I'll see if I can't locate my original AtariVox when I get off work, it's midday in Houston and I'm heading out to run some errands on my lunch break.


SpiceWare

Remembered where my original AtariVox was while I was out running errands.

Results of vox_record attached, currently capturing vox_probe

SpiceWare

Results of vox_probe attached, currently capturing vox_probe2

SpiceWare

results of vox_probe2 attached.

Note: I left the volume control set to tick #4 for these captures from my original AtariVox.

Andrew Davie

Quote from: SpiceWare on 27 Oct 2026, 03:19 AMprobe2 results.  Still AtariVox+, I'll see if I can't locate my original AtariVox when I get off work, it's midday in Houston and I'm heading out to run some errands on my lunch break.

**SpeakJet emulation update: VoxProbe 2 results**

Huge thanks for the VoxProbe 2 recording. It was built to pin down the chip's arithmetic rather than tune individual sounds, and it delivered. Both libraries are updated (C: andrew-davie/SpeakJet; Go: the PR to JetSetIlly/SpeakJet).

**How the chip sets its pitch**

The ROM played every Pitch value from 0 to 254. The chip counts each voice period in whole ticks of a clock of about 8 kHz: **8000 ÷ Pitch, rounded down, plus one**. Above Pitch 200 you can see it plainly. The pitch holds dead flat across runs of settings (201–205 all give the same 199.7 Hz), then jumps exactly where the tick count changes.

That fits all but 12 settings. Those 12, including the default Pitch 88, take exactly one extra tick, consistently, but we couldn't find a formula that produces them. They may come from a lookup table in the chip, so the emulator lists them. The emulator's pitch is now within about 0.1% of the chip everywhere from Pitch 40 to 254. It used to be up to 2.6% off.

**The chip's frequency grid**

All seven DTMF frequencies turned out to be exact whole multiples of 4.3845 Hz (697 Hz is 159 steps, 1477 Hz is 335 steps, which comes out at 1468.75 Hz). Several earlier findings fell into place on that grid:
- one step of Bend is 10 grid steps;
- the odd short step between Bend 4 and 5 is 5;
- Stress is about 19.

That points to a phase accumulator inside the chip. The emulator now runs its oscillators on the same grid and plays the chip's own DTMF tones. The "+5 Hz offset" I mentioned last time was the recorder's clock, not the chip.

**Formants measured directly**

At Pitch 40 each voice pulse stands alone: a few milliseconds of three pure formant tones, then silence. That let us read the formant frequencies straight off the waveform for 18 sounds, with no estimation from spectra. Using them, the emulator's sustained sounds moved closer to the first probe recording, which wasn't used for the measurement: average spectrum error fell from 4.5 to 3.7 dB. LE, long the worst-matching sound, improved by 7 dB.

**Tried and not kept**
- **A voice pulse at the start of every sound.** The chip does this at very low Pitch, so Pitch 0 isn't silent on real hardware. But at normal pitches it doesn't restart the voicing between repeats, and modelling the low-pitch behaviour across the board made things worse. Pitch 0 stays silent in the emulator for now.
- **A voice-driven noise component in the robot sounds.** R7–R9 have a part that follows Pitch but ignores Bend, the opposite of their main tones. Modelling it as a noise oscillator helped on this recording but hurt on the earlier one, so the robots are unchanged.

**The reset mystery**

That odd "DTMF 0-1-0 and the chip changes state" event showed up in 3 of 5 test variants, and nowhere else in the 4½-minute recording. It's neither a fixed position in the byte stream nor a fixed time. It needs a long run of slow speech. My best guess is an overflow in the input buffer handshake under those conditions, but that's unproven. The emulator doesn't reproduce it.

**Results so far** (against the clean VoxRecord take):
- Vowels, nasals, liquids and diphthongs now match at 84% on the spectrogram (82% before).
- Average spectrum error across all speech sounds is 4.2 dB. The original emulator was 6.4.

**Recordings from other AtariVox units are now the most valuable thing.** Every measurement so far comes from one unit. A second unit would show which of these quirks (the 12 odd Pitch ticks, the reset, the grid) belong to the SpeakJet itself and which to one particular chip. If you have an AtariVox or AtariVox+ and can run VoxProbe 2, please do. The same ROM and instructions apply.


Andrew Davie

#21
**SpeakJet emulation: where we're at**

A big thank-you for the recordings from a second AtariVox (an original unit). Having two chips changed what we can do: we can now tell the SpeakJet itself apart from one particular unit, and fit to both at once.

**What two chips told us**
- **Same firmware.** Both units count the voice pitch identically, including the same 12 odd settings, and use the same internal frequency steps and formants. The second unit's clock just runs about 0.3% slow, which is normal for the chip's internal oscillator.
- **The mystery reset is one unit's fault.** The "DTMF 0-1-0, then the chip changes state" event never happened on the second unit, in any of its three recordings. It's a quirk of the first unit, not SpeakJet behaviour.
- **Two real chips sound almost identical.** On vowels, nasals and diphthongs they match each other about 97–98%, and on the sound effects 95%. That gives us a true target. The emulator is at about 75% overall, so there's still real headroom. Earlier I'd suggested hiss sounds had a low ceiling; that came from a clipped recording and was wrong.

**What's improved since the last update**
- **Pitch** now follows the chip's own tick-count rule. It's within about 0.1% everywhere it can be measured, and both chips confirm the 12 exceptions.
- **Oscillators** run on the chip's 4.38 Hz frequency grid, so the DTMF tones are the chip's exact values.
- **Formants** for 16 vowels and liquids, and their relative strengths, are measured directly from single voice pulses.
- **SE, SO and SH** were fixed: I'd had their onset upside down. On the chip it's a quiet, low start, then full-level hiss.
- **Speed, Repeat, pauses and effects** follow the chip's rules, and every vowel's length is "manual time + 28 ms".

**Numbers, against both chips** (spectrogram match, average spectrum error):

| | original emulator | now | two real chips |
|---|---|---|---|
| allophones | 57% / 6.4 dB | 75% / 4.0 dB | 90% / 1.6 dB |
| hiss sounds | | 57% / 2.9 dB | about 65% |

**Where the remaining gap is**

In vowels it's mostly in the "valleys" between formants, which are 8–17 dB too deep in the emulator. We've ruled out the fade length, the fade's shape and the oscillator restart phase. Fitting single voice pulses explains only 80–93% of the chip's waveform, and the missing part is probably what fills those valleys. That's what I'm looking at next. After that comes one joint fit of the shared engine settings across both chips.

Andrew Davie

**VoxProbe 3: what we're testing, and why**

**How the SpeakJet makes a vowel**

From the earlier probes, we now know the voiced sound works like this. At each "pitch pulse", the chip restarts three sine-wave oscillators, one for each formant (the resonances that make an "ee" different from an "ah"). Each sine starts at zero and fades away in a straight line over about 8 ms. Then everything waits for the next pulse.

At normal speaking pitch the pulses come about every 11 ms, so they nearly run together. At Pitch 40 they're 25 ms apart, so each pulse stands on its own and we can see exactly what the chip produces from a single one. That's why everything in VoxProbe 3 plays at Pitch 40.

**What's still wrong**

The emulator now gets each formant's frequency, strength and fade right. But a real chip's vowel spectrum is "fuller" between formants. On IY ("ee"), the gap between the low first formant and the high second one has 10–24 dB more energy on hardware than in the emulator. Other vowels show the same thing to a lesser degree.

I've ruled out the obvious explanations on two separate AtariVoxes:
- the fade length and the fade's shape;
- the phase at which the sines restart;
- a stepped fade, or coarse internal arithmetic;
- the spare oscillators being used as extra tones;
- the output filter.

Something else is putting energy into those gaps, and this ROM is designed to corner it.

**The tests**

1. **Every vowel, nasal and liquid, held for ten repeats.** That gives about 30 separate pulses each, which can be averaged for a very clean picture of one pulse's spectrum. The previous probe had about 9 per sound.
2. **The same sounds at different Bend settings.** Bend shifts all three formants up or down by the same number of hertz.
  - If the extra energy moves with them, it comes from the oscillators themselves, something about how the chip generates or fades each sine.
  - If it stays put, it's added later in the signal path, around the mixing, filtering or output stage.
3. **The same sounds at different Volumes.** If the fill comes from distortion or limited precision in the chip's arithmetic, its level relative to the formants will change with Volume. If it's part of the sound itself, it won't.
4. **Two vowels at Pitches 25 to 60.** This checks whether a single pulse's shape depends on how often the pulses come.
5. **A side mystery: the "first pulse".** At very low Pitch (0 to about 15), the chip fires a voice pulse at the start of every sound, and of every repeat, so even Pitch 0 isn't silent. At Pitch 33 and above, the pulses run on evenly through repeats. This section steps through Pitch 0 to 40 to find exactly where that behaviour changes, so the emulator can copy it rather than guess.

**Why more units help**

Two AtariVoxes have now shown identical firmware: the same pitch arithmetic, frequency steps and formants. Their voiced sounds match each other about 97–98%. That makes every extra recording useful in two ways:
- if a new unit agrees, it confirms the target;
- if it differs, it shows which details belong to one particular chip (like the reset glitch one unit had) and which belong to the SpeakJet itself.

So if you have an AtariVox or AtariVox+, the full run is only about a minute and a half. Recording instructions are in the earlier post and in the README.

Thanks, as always. Every recording so far has replaced a guess in the emulator with a measurement.


SpiceWare

Probe 3 results from AtariVox+ Vectrex Enhanced

SpiceWare

Probe 3 results from original AtariVox

Andrew Davie

#25
Quote from: SpiceWare on 27 Oct 2026, 07:47 AMProbe 3 results from original AtariVox

Many thanks. Results are in...


**SpeakJet emulation: VoxProbe 3 results, and the chip's real voice "fade"**

Thanks for running VoxProbe 3 on both AtariVoxes. It answered the question it was built for, and produced a real improvement.

**Quick recap of the problem**

Each voiced sound on the SpeakJet is a train of pulses. At every pulse, three formant oscillators restart and then fade away. The emulator was getting each formant's frequency and strength right, but real vowels had extra energy in the gaps between formants that we couldn't account for.

**What the ROM showed**

At Pitch 40 each pulse stands alone, and holding each sound for ten repeats let us average about 24 pulses per sound.

- **The two units agree to within about 1 dB** at this level of detail. Whatever differs from the emulator is a property of the SpeakJet itself, not of one particular unit.
- **The missing energy is narrow:** one band per sound, 10–15 dB, often right where the emulator had a deep notch.
- **Under Bend, it moves exactly with the formants.** AW's band sits at 660, 858, 1077 and 1296 Hz at Bend 0, 5, 10 and 15. So it belongs to the formant oscillators themselves, not to a later stage such as mixing or output.
- **Volume doesn't change it from 32 up,** so it isn't distortion.
- **Pitch doesn't change it** from 25 to 60.

That leaves one explanation: the shape of each formant's fade. A fade's shape determines the faint energy that spreads out either side of each formant, and our fade was simply the wrong shape.

**The fix: the chip's measured fade**

Instead of guessing shapes, I let the data draw it. I fitted a free curve, one value every half millisecond and only allowed to fall, to all 18 sounds' single pulses on both units at once. The chip's fade:
- drops quickly to about 0.9 and holds there until about 2 ms;
- falls through about a third and then a sixth;
- holds near an eighth;
- switches off at about 9.5 ms.

Previously we'd used a straight line down to zero over 8 ms.

The curve describes the chip, not one recording. Fitted to either unit alone, it predicts the other just as well.

**Results**
- Single pulses: average spectrum error from 3.7 to 3.0 dB on both chips.
- Voiced sounds across both chips' VoxRecord recordings: 85.7% → 87.0% spectrogram match.
- All speech sounds: 74.7% → 75.5%.
- Sustained sounds from the first probe, which played no part in the fit: 3.55 → 3.33 dB.

**Two corrections and side findings**
- **The "first pulse".** I'd said earlier that the chip restarts its voicing at every repeat of a sound. It doesn't. It fires a pulse at the start of each new sound (so Pitch 0 gives one pulse per sound, not silence), but repeats carry on smoothly. At low Pitch the first gap after that pulse also runs 10–17 ms long. The emulator doesn't copy this yet, because doing it without that longer gap made things slightly worse.
- **At very low Volume (8–16)** the chip adds a faint noise floor between formants. That's limited precision in its arithmetic, and only matters for very quiet speech.

**Where we stand**

On voiced sounds, two real chips agree with each other at about 97–98%. The emulator has moved from about 81% to 87% over the last few rounds, and each step has replaced a guess with a measured chip behaviour: the pitch tick count, the frequency grid, the formants, and now the fade.

Andrew Davie

Update.  I asked Claude to discard all "uncertain" findings and fittings and re-examine all of the measured data again...  It looks like we have a significant improvement so far. Here's an in-progress update.

The first phase of the rebuild is done and in both libraries; nothing is committed or pushed.

| Both units | At the waypoint | Now |
|---|---|---|
| Allophones | 75.4%, 3.89 dB | **79.5%, 3.16 dB** |
| Effects | 70.5%, 4.21 dB | **80.3%, 3.48 dB** |
| Connected-speech lengths | 14 ms | 16 ms |

**Voicing:** fitting the VoxProbe3 pulses as waveforms rather than spectra found three things:
- **Fade:** it holds for 2 ms and is gone by 7.75 ms. The old fade's long faint tail was a compromise of the spectral fit.
- **Output filter:** it agrees with the one measured from the tone sweeps to within 2–3 dB, so the library now uses it.
- **Formants:** each sound's formant frequencies and levels are the average of both units.

Fitted on either unit, the result fits the other just as well. Run on the library's own pulses, the method returns the library's known values, so it isn't inventing the shape. The library's pulses now match the chip's waveforms about five times more closely (−8.6 → −16.4 dB). F1 and F2 also turn out to sit on steps of about 88 Hz.

**Voiced sounds pulse by pulse:**
- **Method:** at the default Pitch every voicing pulse stands alone, so I measured the chip's formants pulse by pulse for every voiced sound.
- **What the chip does:** voiced stops are a low voice bar and then a short vowel-like release, with no noise burst to speak of. Voiced fricatives are mostly voicing, and diphthongs move in steps.
- **Which sounds use it:** 23 sounds now play from these measured frames: the diphthongs, ZZ, ZH, DH, BO, GE, GO, EG, OG, IH, NO and LO. I kept each one only where it scored better on both units. Vowels kept the steady model, which scored better for them.

**Effects:**
- **Background removed:** silencing the recording's background at the start and end of each effect was worth +3 points.
- **Tone effects retraced** on both units, each tone on its own oscillator, starting and stopping in place. For example, B5 went from 54% to 94%, B8 from 62% to 99% and R7 from 34% to 87%.

**State of the code:**
- C tests pass, with the samples regenerated. The duration test now checks only that the framed sounds have a length, as it already did for the traced effects.
- Go matches C to 3e-8; Go's own test file still doesn't build, as before.
- text2speakjet's length tables are updated and its tests pass.
- design.md and the CHANGELOG describe the rebuild.
- Waypoint for reverting: `refs/checkpoints/20261027121810-before-rebuild` in both repos.

**Still to do:**
- **Noise sounds and voiceless stops** need refitting through the new filter. KE and KO are 7–11 dB off.
- **Voiced stops still on the old model:** EB, OB, DE, DO, ED and OD, plus JH. The measured frames scored worse for them.
- **Buzz and noise effects:** A1–A3, M1, M2, most of the C series, and R8 and R9.
- **Transitions between sounds in connected speech.**

Andrew Davie

#27
Update...

Third round done, and it found what was behind the long-standing dips between formants: the chip doesn't restart all three formants at the same phase. Nothing is committed or pushed.

| Both units | At the waypoint | Last report | Now |
|---|---|---|---|
| Allophones | 75.4%, 3.89 dB | 82.3%, 2.68 dB | **82.2%, 2.53 dB** |
| Voiced sounds | — | 91.7%, 2.09 dB | **92.1%, 1.90 dB** |
| Effects | 70.5%, 4.21 dB | 80.5%, 3.45 dB | 80.5%, 3.45 dB |
| Connected-speech timing | 14 ms | 16 ms | 16 ms |

**Formant phases.**
- Each sound's F2 and F3 start at their own phase relative to F1, the same on both units. For example, IH's are about 0.13 and 0.17 of a cycle ahead, and MM's about a quarter cycle.
- Those phases set how deep the dips between formants are. Forcing all three to start together, as the library did, more than doubles the model's error (1.96 → 4.52 dB), and costs up to 12 dB for IY, MM and NE.
- The library's voicing pulses now restart each formant at its measured phase, for the 16 steady sounds.

**Fade onset.** I'd smoothed the measured fade's first 2 ms into a flat start. The chip's real onset, which dips to about 0.7 within half a millisecond, shapes the whole spectrum, so the library now uses the measured curve as it is. Together with the phases, the voiced sounds go from 2.09 to 1.90 dB.

**Code:** C tests pass with the samples regenerated, Go matches C on every test stream, and the design doc, CHANGELOG and notes are updated.

**Tried and not kept:**
- per-pulse phases for the 26 framed sounds (82.2% → 81.7%);
- an automatic loop calibrating the library against the model, which diverged;
- retracing the gated C-series effects or switching them to the voicing envelope. C0, C2 and C4 do gate with the voicing fade at their own rates, but that alone changes nothing; their traces are the problem.

**What's left is smaller per item:**
- the gated and buzz effects (C-series, A1, A3, M1, M2, R8, R9);
- transitions between sounds in connected speech;
- the remaining gap between the library and the model on single pulses (2.4 against 2.0 dB), spread across the 11-tap filter, the 8192 Hz rendering and the frequency grid.


Andrew Davie

**AtariVox recordings wanted: VoxProbe 4**

Thanks to everyone who recorded VoxProbe 3. Those single voicing pulses, from two different AtariVoxes, turned out to be enough to rebuild a large part of the emulator from the recordings themselves instead of guesses.

**What the recordings showed**
- **The voicing fade.** Each voicing pulse is three formant tones that start together and fade out in a measured envelope. That envelope rises to full level, holds for about 2 ms and is gone by 7.5 ms. It's identical on both units.
- **Formant phases.** The second and third formants don't start in step with the first. Each sound has its own fixed phase offsets, the same on both chips. That turned out to be why the emulator's dips between formants were too deep.
- **The output filter.** It came out of the same data: a smooth low-pass about 5 dB down at 2 kHz and 12 dB down at 3 kHz.
- **Voiced stops.** Sounds like BO and GE are a low voice bar followed by a short vowel-like release, with no noise burst to speak of.
- **Sound effects.** I re-traced them from both units, with every tone on its own oscillator, so tones now start and stop instantly as the chip does instead of gliding in.

**The emulator is now much closer**
- Speech sounds: 75% → 82% average spectrogram match against the chips.
- Sound effects: 70% → 80%.
- For reference, two real chips agree with each other at about 90% on speech sounds and 95% on effects. So there's still room to improve, and that's what VoxProbe 4 is for.

**What VoxProbe 4 tests (about 3 minutes)**
1. The noisy and gated effects (the whole C series, A1, A3 and M1), played twice to average out their randomness.
2. The 16 effects the emulator still gets most wrong, at Pitch 60 and 150 and at Bend 0 and 15. This shows which parts of each effect follow Pitch and which follow Bend, and so how the chip builds them from its oscillators.
3. All 16 diphthongs and 12 voiced stops, each repeated 5 times. The repeats line up pulse for pulse, so they can be averaged to show exactly how each sound moves.
4. Eight voiced sounds VoxProbe 3 didn't cover (NGE, NGO, LO, VV, ZZ, ZH, DH, JH), as single voicing pulses at Pitch 40.
5. Sixteen pairs of sounds (AW IY, IY AW, MM AW, AW SE, TT AW and others), to measure how the chip moves from one sound to the next. This is the last part of connected speech that's still a guess.
6. Twelve hiss sounds, four times each, again to average out the noise.
7. A few effects at Speed 90 and 127. These are last, and no Speed goes below 90, because one unit has occasionally reset part way through long, slow sequences.


Andrew Davie

Update while we wait for #4 test...

---

**SpeakJet emulation: working from the worst sounds**

After the rebuild from the recordings, I tried something different. Instead of tuning the averages, I took the sounds the library reconstructs *worst* and asked what they have in common. A sound that already matches can't tell you what the model is missing; the worst ones can.

**Method.** I scored every allophone and effect against both of my units (the original AtariVox and the newer one). Then I ranked them not by raw error but by how much worse the library is than the two chips are from *each other*. That matters because some sounds vary between real chips. The hiss sounds looked worst in the raw scores, but the two chips disagree on them just as much, so they're essentially done. Nearly all the remaining headroom was in the sound effects: buzz effects 4.3 dB beyond chip-to-chip, tone effects 2.6, vowels about 1.

**What the worst sounds turned up:**

- **Voicing starts too fast.** In 39 of 42 voiced sounds the chip takes longer to come up from silence. Measured period by period, the chip's first three pulses sit at −17, −6 and −1.5 dB; the library's were at −8, −3 and 0. The cause was in the library's mixer: the volumes did glide in over 20–50 ms, but the mixer's level limit hit full output within about two pulses. A 40 ms ramp after the mixer matches the chips within about 1 dB. It's a fixed time, not a number of pitch periods: about 30 ms at every pitch from 60 to 250. Nasals come up faster, about 25 ms.
- **NGE and NGO had never been measured.** They were still guesses. Both chips agree on every harmonic of them within 0.5 dB, so I fitted them: NGE 87% → 97%, NGO 93% → 96%. On the chip NGE's upper two formants are a single peak near 2.3 kHz.
- **TS was backwards.** The chip's TS is a hiss that starts at full level around 3.5 kHz and dies away over about 120 ms. The library had a long broad burst followed by a short hiss. Now 13% → 37%, where the two chips match each other at 35%.
- **C8 and C9 are noise, not tones.** On both chips they're flat noise from 400 Hz to 2.8 kHz, gated at 52 Hz about 15 dB deep. The old trace had taken the strongest noise peaks for three tones. The library could only gate noise 6 dB, so gate depth is now adjustable. When I fitted it freely it came out at −15 dB, the same as I'd measured. 34% → 58%, 8.4 → 2.9 dB.
- **The chip's tones aren't clean sines.** Every tone sits on a skirt of noise and stray lines 15–30 dB below it, well above the recording's background. Several C-series effects also have broad noise under their tones. Adding that noise, fitted against both chips: M0 67% → 97%, A0 66% → 86%, C1 42% → 67%, B6 6.1 → 0.7 dB.
- **The tone traces were clipped.** The tracer capped volumes at the oscillator maximum (31), and 30 effects go past that, so their loudest stretches were flat: A0's swell, B6's fall, the whole body of the DTMF tones. Each effect is now scaled so its loudest point sits at the maximum, which keeps its shape. B2 4.9 → 0.6 dB, R0 3.1 → 1.0 dB.
- **A4 was cut short.** On both chips A4 sweeps on to 3.9 kHz; the trace had stopped at 1315 Hz. 40% → 68%. M1 also lasts longer, as the chips do.

**Things that looked like findings but weren't.** These are worth knowing if you do similar analysis:

- An apparent energy excess near 170 Hz in every sound was a high-pass filter in my own processing of the recordings, not the chip.
- Nasals looking "late" to start, and a quiet lead-in on TS, were both the recording's background noise in the 30 ms before each sound. It shows up more against quieter sounds.
- A C-series gating rate of ~160 Hz was the second harmonic of the real rate.

**One curiosity.** B7's 3.9 kHz tone has a mirror image at 4.12 kHz only 1 dB down. That puts the chip's output sample rate at roughly 8.03 kHz. So that tone is beating against its own image, and the junk under it appears only with that tone. I couldn't add it without making B7 worse in other ways, so B7 stays as it was for now.

**Numbers.** Over both units, since the start of this round:

|          | Before | After |
|---|---|---|
| Effects | 80.5% / 3.45 dB | 85.3% / 1.90 dB |
| Allophones | 82.2% | 83.3% |
| Voiced sounds | 92.1% | 93.6% |

Connected-speech timing is unchanged. Compared with before the rebuild, allophones have gone from 75% to 83% and effects from 70% to 85%.

**Still open:**
- B7 (above).
- C3, where added noise made one unit match worse.
- Vowels played on their own come out about 15 ms short, even though their length is right in connected speech, which is what they're fitted to.