# Hard delivery conditions (2026-09-10)

Evidence for `182.C.18.08`, which names nine: fast and slow speech, whisper,
shouting, laughter, singing, emotional delivery, background music, accent
variation, and overlapping speech. Those nine phrases expand to ten conditions
in code, because "fast/slow speech" is two and they behave differently.

**Four of the ten were measured. Six were not, and could not be.**

## Why six could not be built

The four that were built are TRANSFORMATIONS of a signal — a tempo change, a
mix, an offset copy — and each produces audio that genuinely has the property.
The six are different phonation or different speakers:

| Condition          | Why a filter will not do it                                         |
| ------------------ | ------------------------------------------------------------------- |
| whisper            | whispered speech is unvoiced; filtering a voice gives a quiet voice |
| shouting           | shouting changes formants and phonation, not only level             |
| laughter           | not speech; cannot come out of a speech clip                        |
| singing            | pitch-shifting speech to a melody is not singing                    |
| emotional delivery | one Kokoro voice is cached and it has no emotion control            |
| accent variation   | an accent is another person                                         |

A fixture that merely looks like the condition proves the detector reads the
filter. All six need recorded speech — the `182.C.33` corpus.

## What the four measured, and the two holes they found

Onset errors are against the KNOWN onsets **scaled by the tempo factor**, so a
correct aligner on a 1.25x clip reads near zero rather than 25% late. Onsets
carry the `182.C.18.02` calibration of −48.288 ms.

| Fixture          | max onset error | stretch slope | minGop   | est. SNR |
| ---------------- | --------------- | ------------- | -------- | -------- |
| normal           | 32.9 ms         | 0.994         | 0.00     | 81.3 dB  |
| fast-1.25x       | 52.7 ms         | 0.992         | −4.99    | 81.6 dB  |
| fast-1.60x       | 68.4 ms         | 0.992         | −4.84    | 81.2 dB  |
| slow-0.80x       | 63.5 ms         | 0.990         | −10.20   | 81.5 dB  |
| music-under-20db | 52.9 ms         | 0.991         | 0.00     | 31.2 dB  |
| music-under-10db | 52.9 ms         | 0.991         | 0.00     | 21.5 dB  |
| music-under-5db  | 52.9 ms         | 0.991         | 0.00     | 16.8 dB  |
| **overlap-0.5s** | **513.8 ms**    | 1.016         | **0.00** | 81.5 dB  |
| **overlap-2.0s** | **2014.4 ms**   | **1.406**     | −14.02   | 81.5 dB  |

### Hole 1 — time stretch produces false mispronunciation findings

The fast and slow clips contain **identical words from an identical speaker**.
Only the tempo changed. Yet `182.C.18.03`'s goodness-of-pronunciation falls to
−4.99 at 1.25x and −10.20 at 0.80x, past its −3.0 floor, so the gate would
report a mispronunciation that did not happen. A phase vocoder's transient
smearing looks like bad pronunciation to a frame-level acoustic model.

`assertGopMeasurable` now refuses a pronunciation score on time-stretched audio.
Refuses, not discounts: a discounted wrong number is still a wrong number.

The aligner, by contrast, held — 68.4 ms at 1.6x is well inside the ITU +45/−125
ms reject boundary — so timing claims are still allowed on it.

### Hole 2 — overlapping speech breaks the aligner silently

A second voice starting 0.5 s in gives a **513.8 ms** onset error, four times
the ITU reject boundary, and the aligner reports success. At 2.0 s it is 2014.4
ms with a 1.406 stretch slope. Forced alignment is given the words and asked
where they are; it will always find them somewhere.

Worse, at the 0.5 s offset the GOP is **exactly 0.00** — the take looks
perfectly clean while its timing is half a second wrong. A clean-looking
evaluator beside a broken one is how a bad take gets through.

`assertAlignmentMeasurable` and `assertGopMeasurable` both refuse on overlapping
speech.

### Not a hole, but worth knowing — music trips the SNR gate

The aligner and the GOP are **completely unaffected** by a harmonic bed even 5
dB under the speech. But the percentile-split SNR estimator reads 16.8 dB there
and the `182.C.18.04` low-snr gate fires at 20. Music under speech is not noise,
and that estimator cannot tell the difference. The finding is not suppressed;
`snrFindingCaveat` attaches the reason it may be the estimator rather than the
take.

## Reproducing

```
python tools/human-video/evaluator/make_delivery_fixtures.py --out /tmp/delivery
python tools/human-video/evaluator/ctc_align.py --audio /tmp/delivery/overlap-0.5s.wav \
  --transcript "engine ships tuesday costs forty dollars"
```

The fixtures are seeded and not committed; the generator reproduces them.

## Per-item status this evidence supports

| Item          | Was           | Is                                                                                                       |
| ------------- | ------------- | -------------------------------------------------------------------------------------------------------- |
| `182.C.18.08` | `not-started` | `blocked-corpus`: 4 of 10 conditions measured with real fixtures, 6 need `182.C.33` recorded speech      |
| `182.C.18.03` | done          | now guarded — GOP refuses on time-stretched audio, after measuring false findings at every tempo factor  |
| `182.C.18.02` | done          | now guarded — word timing refuses on overlapping speech, after a 513.8 ms silent error at a 0.5 s offset |
