# ROSE — Data Dictionary

Documentation for the British participant-level acoustic-feature dataset deposited under
**doi:10.5525/gla.researchdata.2332** as `ROSE_closed_British.zip`, and for the Syrian dataset
prepared separately as `ROSE_closed_Syrian_HOLD_pending_ethics_review.zip` pending RDM review.
This dictionary is openly available so that the structure of the data is fully documented even
though the data files themselves are not released.

| File | Rows | Speakers | Recordings per speaker | Columns |
|---|---|---|---|---|
| `ROSE_Syrian_bilingual_features_FINAL.xlsx` | 618 | 103 | 6 (S1–S6) | 15 |
| `ROSE_British_native_features_FINAL.xlsx` | 90 | 30 | 3 (S7–S9) | 16 |

**Unit of observation:** one recording (one speaker reading one sentence).
**Missing values:** none in either file (0 missing cells).
**File format:** Microsoft Excel (.xlsx), one worksheet, one header row, with no formulas or external links.

---

## 1. Metadata variables

| Variable | Type | Values / range | Description |
|---|---|---|---|
| `SpeakerID` | text | Syrian: `ROSE001`–`ROSE###`; British: `UK001`–`UK###` | Pseudonymous speaker identifier. Not derived from any personal attribute. The linkage between these identifiers and the original participant records is held separately by the research team and is **not** deposited. |
| `Language` | text | Syrian: `Arabic`, `English`; British: `English` | Language of the recording. For Syrian speakers, `Arabic` = first language (L1) and `English` = second language (L2). |
| `Nativeness` | text | British file only: `native_L1` | Nativeness of the speaker in the recorded language. Present only in the British file; for the Syrian file, nativeness is implied by `Language` (Arabic = native L1, English = non-native L2). |
| `Sentence` | text | Syrian: `S1`–`S6`; British: `S7`–`S9` | Sentence identifier. Syrian: `S1`–`S3` Arabic, `S4`–`S6` matched English translations. British: `S7`–`S9`, the same English sentences. |
| `Valence` | text | `positive`, `negative`, `neutral` | Intended affective category of the sentence. Mapping: S1/S4/S7 = positive; S2/S5/S8 = negative; S3/S6/S9 = neutral. **Interpretation caveat:** each category is realised by a single sentence, so the category is confounded with that sentence's wording, phonetic content and length. In the associated article this factor is therefore reported as *stimulus condition* rather than as a pure valence manipulation, and no arousal claim is made. |
| `PHQ9` | integer | Syrian 2–26; British 1–25 (instrument range 0–27) | Total score on the nine-item Patient Health Questionnaire. **Special-category (health) data.** A self-report screening and severity measure, not a clinical diagnosis. One value per speaker, repeated across that speaker's rows. |
| `Gender` | text | Syrian file: `F`, `M`; British file: `Female`, `Male` | Participant-reported sex/gender. **Coding differs between the two files** (abbreviated in the Syrian file, spelled out in the British file); the analysis code harmonises them to a single two-level variable. The exact questionnaire item did not distinguish biological sex from gender identity, so the variable is described throughout as *participant-reported sex/gender*. |
| `Age` | integer | Syrian 18–25; British 19–25 (years) | Participant age in whole years. One value per speaker, repeated across that speaker's rows. |

## 2. Acoustic feature variables

All eight features come from the extended Geneva minimalistic acoustic parameter set
(**eGeMAPSv02**) computed at the **Functionals** level with the `opensmile` Python package v2.6.0,
after conversion of the audio to 16 kHz mono, 16-bit signed format with FFmpeg 8.1.1. Column names
are the exact openSMILE identifiers. All are continuous (floating point), one value per recording.

| Variable (openSMILE identifier) | Domain | Description | Unit |
|---|---|---|---|
| `VoicedSegmentsPerSec` | Temporal (timing) | Number of voiced segments per second; relates to speaking/articulation rate | per second (s⁻¹) |
| `MeanUnvoicedSegmentLength` | Temporal (timing) | Mean duration of unvoiced segments (unvoiced/silent intervals; pausing) | seconds |
| `StddevUnvoicedSegmentLength` | Temporal (timing) | Standard deviation of unvoiced-segment duration (variability of pausing) | seconds |
| `loudness_sma3_amean` | Energy / amplitude | Arithmetic mean of frame-level perceptual loudness | arbitrary units (perceptual loudness estimate; no absolute calibration) |
| `loudness_sma3_stddevNorm` | Energy / amplitude | Normalised standard deviation of loudness (standard deviation ÷ arithmetic mean) | dimensionless (ratio) |
| `F0semitoneFrom27.5Hz_sma3nz_amean` | Frequency (pitch) | Arithmetic mean of fundamental frequency on a semitone scale referenced to 27.5 Hz | semitones (re 27.5 Hz) |
| `F0semitoneFrom27.5Hz_sma3nz_stddevNorm` | Frequency (pitch) | Normalised standard deviation of the semitone fundamental frequency | dimensionless (ratio) |
| `HNRdBACF_sma3nz_amean` | Voice quality (spectral) | Arithmetic mean of the harmonics-to-noise ratio (autocorrelation-based) | dB |

**Notes on units.** Units are stated only where documented by the toolkit. The loudness functional
is an uncalibrated perceptual estimate and is sensitive to recording gain; the two `stddevNorm`
functionals are dimensionless ratios and are described as *normalised standard deviation* rather
than as a coefficient of variation. The harmonics-to-noise ratio is sensitive to microphone, noise,
device and recording environment, and should not be compared across recording set-ups without
caution.

## 3. Derived variables used in the analysis (not stored in the files)

These are generated in code (see the open subset) and are documented here for transparency:

| Derived variable | Definition |
|---|---|
| `StimulusCondition` | Three-level factor mapped from `Sentence` (positive / negative / neutral), reference level *neutral* |
| `Context` | Three-level factor: `Syrian_Arabic_L1`, `Syrian_English_L2`, `UK_English_L1` |
| `Sex_or_Gender` | Harmonised two-level factor from `Gender`, reference level *female* |
| `*_z` variables | Standardised (z-scored) feature, PHQ-9 and age values. Syrian features are standardised across both languages pooled; British features are standardised within the British cohort. Cross-context comparisons therefore concern association strength, not raw acoustic levels. |

## 4. Integrity

| File | SHA-256 |
|---|---|
| `ROSE_Syrian_bilingual_features_FINAL.xlsx` | `1b7b3a0660662b641a68d2105abc017cbce9cd85f7408b4196301aa31e198495` |
| `ROSE_British_native_features_FINAL.xlsx` | `7388c56de41c253287561c92138930cf27654314a618b46fd64b3d0dd0254f3c` |

The analysis script verifies these checksums before use, so any alteration of the input files would
be detected. See `MANIFEST_SHA256.txt` for checksums of every deposited file.