Breathless Exile's Whisper — all 832 conditions

Four sensors produce the inputs, all automatic:

• genuineness — does it sound felt rather than performed
• blend — how naturally non-speech sounds (breaths, laughs, sobs) sit inside the line
• target — this is the part that changes per dimension; see below
• WER — word error rate of an ASR pass against the script, so an unintelligible take cannot win by being dramatic

Why multiply by (1 − WER) rather than divide by (1 + WER)? The project manual specifies the dividing form. Measured over 28,698 candidates it is too lenient: a take that rambles past the written line keeps accumulating emotional content, so genuineness, blend and strength all rise while ÷(1+0.79) only scales it by 0.56. One 28-second take of a 20-word line — whose transcript wanders into "meant in staying in the dist…" — outranked an 8-second take of the same line with a quarter of the word errors. Switching to ×(1−WER) moved 28 of 808 winners.

An earlier version z-scored each term within the group. That sounds fairer but has a flaw: normalising inside a group means a group whose takes are all badly blended still hands its best one a good z-score. Blend then stops competing across groups, and the whole corpus drifts toward flat delivery.

Switching to raw values, measured on the same 28,698 candidates:

The genuineness loss is a real trade, not an oversight: blend and strength are bought with it. It is recorded here rather than buried, because a reader comparing this corpus against another on genuineness alone would otherwise reach the wrong conclusion.

The straightforward case: the emotion classifier's reading of that emotion, and more is better. Condition A is intense and freely expressed, B is intense but held in.

B additionally gets the containment treatment described under moderate/contained below — being intense and being contained are independent axes here.

This is the one people find counter-intuitive. A "moderate sadness" group is not scored for being as sad as possible — it is scored for landing in a band. Maximising is what the intense conditions are for. A take that overshoots into full grief is as wrong for this cell as one that is flat.

Two adjustments, both because containment is the opposite of letting it out:

• Blend is weighted down to 0.6. An audible burst is the emotion escaping, which is precisely what a contained take should not do. Rewarding blend at full weight here would select against the condition.
• A containment term is added: how far below this group's own mean the take's vulnerability sits. This one stays on the z-scale deliberately — "below the group mean" has no meaning as a raw number.

Here the target is the dimension being steered, not an emotion. The sign flips for the low end, so an "extremely low tempo" cell is rewarded for being slower, using the same formula as the high end rather than a separate rule.

Without the sign flip, every low cell would be ranked by how much of the dimension it had — i.e. ranked backwards.

A character group's target is a voice, not a feeling, and there is no per-character classifier to score against. If the emotion term were left on, the twelve character clusters would be ranked by whichever emotion happened to leak into each take — the dragon would be represented by its angriest reading rather than its most draconic one.

So the term is zeroed and the ranking asks only: clean, in character, long enough, and not accidentally emoting.

These rank like intense emotions, with one important difference in how WER is measured: inline burst tags such as <sobs> are stripped from the reference before WER is computed.

This was a real bug. The code emitted <sobs> while the stripper matched only round brackets, so the tag stayed in the reference text and every burst-carrying take was charged word errors for failing to pronounce the word "sobs". The ranker was selecting against exactly the takes the bursts were meant to improve. After the fix, bursts win: emotion 2.532 → 2.788, vulnerability 4.666 → 5.085.

Subset A is 32 different sentences; subset B is 16 shared ones rendered across all four sub-conditions. They answer different questions, so a single ranking would let whichever subset happened to draw easier sentences crowd the other out of the top slot the page shows. Ranks are computed within subset; the z-scores stay group-wide so both remain on one scale.

Every number above comes from automatic sensors. Their agreement with a human listener on spoken material is only about ρ ≈ +0.21. Treat the ranking as useful, not authoritative.

This project has produced four separate cases of a scorer confidently rewarding broken audio: a genuineness gain that turned out to be a generator running to its length cap; an aesthetics head scoring a model 4.12 while human-facing judges scored it 0/30; the WER stripper described above; and a caption model whose "improvement" tracked a label-noise artefact. Where a number here surprises you, listening is the tie-break, not the score.