# ROSE — Acoustic feature datasets and analysis code for within-speaker cross-lingual associations with depressive symptom severity

**Enlighten: Research Data — deposit reference 2332**
**Data DOI:** 10.5525/gla.researchdata.2332
**Related code record:** 10.5281/zenodo.21620473 (Zenodo archive of https://github.com/Raneem-Knaj/ROSE-speech-analysis)

**Dataset creators:** Dr Raneem Knaj (Raneem.Knaj@glasgow.ac.uk); Dr Julien Le Kernec (Julien.LeKernec@glasgow.ac.uk)
James Watt School of Engineering, University of Glasgow, Glasgow G12 8QQ, United Kingdom
**Corresponding author for the dataset:** Dr Raneem Knaj
**Copyright:** University of Glasgow
**Deposit date:** 2026

---

## 1. Access conditions at a glance

| Component | File | Access | Licence |
|---|---|---|---|
| Documentation (this README, data dictionary, provenance note, manifest) | unzipped, top level | **Open** | CC BY 4.0 |
| Analysis and feature-extraction code, non-disclosive aggregate outputs | `ROSE_open_access.zip` | **Open** | MIT (code) / CC BY 4.0 (outputs, docs) |
| Participant-level acoustic-feature dataset — British cohort | `ROSE_closed_British.zip` | **Closed** (deposited for integrity; not shared) | n/a — not released |
| Participant-level acoustic-feature dataset — Syrian cohort | `ROSE_closed_Syrian_HOLD_pending_ethics_review.zip` | **Closed**, and **held pending RDM review** of the Syrian consent documentation | n/a — not released |
| Raw audio recordings | **not deposited** | — | — |

The British participant-level acoustic-feature dataset is deposited as **closed data** on the
advice of the University of Glasgow Research Data Management team, following their review of the
approved British consent form and participant information sheet. The Syrian participant-level
feature dataset is held separately and must not be added to the repository record until RDM has
reviewed the corresponding Syrian consent documentation and confirmed the appropriate deposit
arrangements. The raw audio recordings are **not** deposited in any form.

---

## 2. What this dataset is

The deposit supports a study of whether acoustic correlates of depressive symptom severity remain
stable when the *same speaker* changes language. Participants read short sentences; eight
pre-specified acoustic features were computed from the recordings; symptom severity was measured
with the nine-item Patient Health Questionnaire (PHQ-9). The design has three cells:

1. **Syrian speakers, Arabic (first language, L1)** — sentences S1–S3
2. **The same Syrian speakers, English (second language, L2)** — sentences S4–S6
3. **Independent British native-English speakers (L1)** — sentences S7–S9

Cells 1 and 2 support a within-speaker comparison across languages; cell 3 is an independent
external replication. PHQ-9 is a self-report screening and severity instrument, **not** a
diagnostic instrument, and no clinical diagnosis was made. The study is cross-sectional and
observational; the symptom associations are between-speaker.

## 3. Contents

### Open, unzipped documentation (this level)
| File | Description |
|---|---|
| `README.md` | This file |
| `ROSE_Data_Dictionary.md` | Every variable in the feature datasets: type, units, permitted values |
| `ROSE_Data_Provenance_and_Exclusions.md` | How the datasets were produced; participant flow and exclusions; pseudonymisation |
| `MANIFEST_SHA256.txt` | SHA-256 checksum of every deposited file |
| `LICENCE_CODE_MIT.txt` | Licence for the code |
| `LICENCE_DOCUMENTATION_CC-BY-4.0.txt` | Licence for documentation and aggregate outputs |

### `ROSE_open_access.zip` (open)
- `code/` — the two feature-extraction scripts, the confirmatory analysis script (current and
  superseded versions), the two figure scripts, the environment-capture helper, pinned analysis
  requirements, and partially documented extraction requirements — the exact extraction-time
  versions of pandas, numpy and openpyxl were not retained.
- `aggregate_outputs/` — model coefficient tables, diagnostics, the analysis run log, the output
  checksum file, and the forest figure. These are **aggregate model outputs only**: they contain no
  participant identifiers and no participant-level rows (verified). The extraction scripts in
  `code/` mention three pseudonymous identifiers in exclusion settings and a log message; they
  contain no participant-level data and no linkage information.

### `ROSE_closed_British.zip` (closed — deposited for integrity, not shared)
- `ROSE_British_native_features_FINAL.xlsx` — 30 speakers × 3 recordings = 90 rows

### `ROSE_closed_Syrian_HOLD_pending_ethics_review.zip` (closed — held)
- `ROSE_Syrian_bilingual_features_FINAL.xlsx` — 103 speakers × 6 recordings = 618 rows

The Research Data Management team's advice to deposit the participant-level data as closed was given
on the basis of the **British** cohort ethics paperwork. The Syrian consent and participant
information documents are in Arabic and are still under review by the RDM team. The Syrian file is
prepared separately and clearly marked; **it is not included in the current transfer and should not
be added to the record until that review is complete.** The record can be finalised with the open
subset and the British closed file alone.

These two files contain pseudonymous speaker identifiers, PHQ-9 scores, age,
participant-reported sex/gender, language, sentence and stimulus condition, and the eight derived
acoustic features. They are **pseudonymised**, not anonymised, and include special-category
(health) information.

## 4. Ethics and consent

| Cohort | Approval | Reference | Date |
|---|---|---|---|
| Syrian | Tartous University Institutional Review Board | No. 1113 | 4 May 2025 |
| British | University of Glasgow College of Science and Engineering Ethics Committee | 300250128 | 17 May 2026 |

All participants gave informed consent. Participants were informed of their PHQ-9 scores, and those
with scores indicating high symptom severity were advised to seek assessment from an appropriate
health professional in accordance with locally available services. A blank consent form and the
participant information sheet have been supplied privately to the Research Data Management team;
they are not published under an open licence unless RDM confirms this is appropriate.

## 5. Methods summary

**Recording.** All Syrian recordings were made in the same sound-isolated university laboratory
using the same Dell university laptop and its built-in microphone. British participants used their
personal laptops after instructions to record in a quiet environment; device models and native
recording formats varied and were not systematically documented. British data were collected during
June 2026, after ethics approval on 17 May 2026.

**Feature extraction.** Audio was converted to 16 kHz mono, 16-bit signed format with FFmpeg 8.1.1;
the extended Geneva minimalistic acoustic parameter set (eGeMAPSv02) was computed at the Functionals
level with the openSMILE toolkit (`opensmile` Python package v2.6.0, bundled `eGeMAPSv02.conf`,
88 features), under Python 3.11. The eight features retained for analysis are listed in the data
dictionary. Features were computed from the audio; PHQ-9 and demographic metadata were joined
afterwards by pseudonymous speaker identifier.

**Analysis.** Linear mixed-effects models with participant random effects, covariate adjustment
(age, participant-reported sex/gender, stimulus condition) and Benjamini–Hochberg false-discovery-rate
control within pre-specified families; a by-speaker (cluster) bootstrap and leave-one-speaker-out
refitting as sensitivity analyses. Python 3.12.3 with statsmodels 0.14.6, numpy 2.4.4, pandas 3.0.2,
scipy 1.17.1; fixed random seed (20260712), so the analysis is deterministic. Full code is in the
open subset and in the linked code record.

## 6. Reuse and citation

The code and documentation may be reused under the licences stated above. The participant-level
feature data are **not** available for reuse. Please cite both records:

> Knaj, R., & Le Kernec, J. (2026). *ROSE: Acoustic feature datasets and analysis code for
> within-speaker cross-lingual associations with depressive symptom severity* [Data set].
> University of Glasgow. https://doi.org/10.5525/gla.researchdata.2332

> Knaj, R., & Le Kernec, J. (2026). *ROSE: feature-extraction and analysis code* (v1.0.0)
> [Software]. Zenodo. https://doi.org/10.5281/zenodo.21620473

An associated research article is under review; its citation will be added to this record on
publication.

## 7. Contact

Questions about this deposit should be directed to Dr Raneem Knaj
(Raneem.Knaj@glasgow.ac.uk), James Watt School of Engineering, University of Glasgow.