---
name: call-cross-lingual-emotion-preservation
description: Offline experimental QA helper for CALL-E relay transcripts. Compares English urgency-marker levels in source and relay-agent text; both inputs must be English or operator-prepared English translations. Returns advisory lexical drift labels and suggested relay wording, without translating, measuring emotional state, certifying relay fidelity or placing calls.
license: MIT
---

# call-cross-lingual-emotion-preservation

> **When the message crosses a language, does the urgency survive?**

`language-bridge-call` relays a request across languages in two legs. This
skill is an experimental QA companion: it compares English urgency-marker
levels in supplied text from both legs. It does not measure actual emotion
or prove that a translation preserved it. A phrase such as "today, immediately, please"
that arrives as "sometime this week would be fine" is a failed relay even
when the words are translated correctly.

## When To Use

- after a `language-bridge-call` relay, to check the requester's urgency
  survived the second leg
- after any translated CALL-E call where emotional subtext matters
  (escalations, care requests, time-critical arrangements)
- to generate an intensity-calibrated relay goal for the next `plan_call`

## When Not To Use

- to relay or translate anything; use `language-bridge-call`
- to detect sarcasm or stated-vs-meant mismatch; use
  `call-verbal-irony-detector`
- to audit the relay callee's own emotions; only the relay agent's
  expressed intensity is measured, because the relay speaks on the
  requester's behalf
- on non-English source OR relay text; the same English-only lexicon scores
  both. Provide operator-prepared English translations when needed; this
  script does not translate. A target-language parameter does not change
  the scorer. Non-English text can produce misleading low scores.

## Workflow

### Analyze a relay

```bash
python3 scripts/emotion_preservation.py analyze --source-context path/to/source.json --relay-transcript path/to/relay.json
```

`--source-context` accepts the requester's call-result JSON (callee turns
are used) or a plain-text operator note. `--relay-transcript` accepts the
relay leg's CALL-E result (nested `get_call_run` or flat fixture shape).
Both texts must be English or separately translated into English by the
operator. The script does not detect or enforce language eligibility.
Emits a card:

- `source_intensity` / `relay_intensity`: {score, level, markers}; the
  relay side counts AGENT turns only
- `drift`: FLATTENED / PRESERVED / AMPLIFIED (level comparison)
- `parity_score`: 1.0 equal, 0.5 adjacent, 0.0 two steps apart
- `evidence`: masked spans with matched markers, from both sides
- `emotion_assessment: "unclear"` with a reason when the source context is
  empty or the relay has no agent turns
- `recommended_action`: `re_relay_with_calibrated_goal` (with the goal text
  calibrated to the SOURCE intensity) or `proceed`

These are legacy action labels for human review. Scores and low/medium/high
levels are illustrative lexicon thresholds, not empirically calibrated
emotion measures. Neither `proceed` nor a re-relay suggestion authorizes a
new call, emergency response or other consequential action.

### Craft the calibrated relay goal

```bash
python3 scripts/emotion_preservation.py craft --scenario emotion-relay --intensity high --language en
```

Emits the plan_call inputs JSON whose `goal` is the same intensity-
calibrated template the card recommends on drift (high / medium / low).

## Scientific Foundation

| Research | Relevance |
|---|---|
| ZEST: Zero Shot Audio to Audio Emotion Transfer With Speaker Disentanglement (ICASSP 2024, arXiv 2401.04511) | Zero-shot emotion transfer between speakers; motivates intensity-preserving relay design |
| EELE: Exploring Efficient and Extensible LoRA Integration in Emotional Text-to-Speech (2024, arXiv 2408.10852) | Efficient emotional TTS control; motivates mapping intensity levels to concrete phrasing guidance |

Both papers build neural emotion-transfer systems; this skill deliberately
implements a lexical intensity mapper on transcripts only and labels every
output `analysis_mode: "heuristic"` - it is informed by that research, not
an implementation of it.

## Differences from sibling skills

- `language-bridge-call` performs the relay; this skill audits the relay's
  emotional fidelity and calibrates the next attempt.
- `call-semantic-barge-in-analyzer` profiles how the callee participated;
  this skill measures what the relay agent expressed on someone's behalf.