| # Data dictionary: `data/audio_context_study/` |
|
|
| Follow-up study: does the audio–text difference survive when the |
| conversation history is heard rather than read? In the main study the |
| history was shown as text even in the audio condition. Here the history |
| and the current query are played as audio before the two responses. The |
| responses are the main study's Kokoro `af_heart` clips. |
|
|
| | Slug (file suffix) | History rendering | Analyzed raters | Label in the paper | |
| |---|---|---:|---| |
| | `audiocontext_onevoice` | one voice (`af_heart`) for both speakers, with an on-screen speaker indicator | 22 | B1 | |
| | `audiocontext` | two voices: user turns `am_michael`, assistant turns `af_heart` | 23 | B2 | |
|
|
| - **Stimuli:** 30 prompts that have conversation history |
| (`stimuli/audio_context_study/prompts.csv`). 15 of them overlap with |
| the voice study. |
| - **Session:** each participant rated 15 trials plus one attention |
| check, about 47 minutes. |
| - **Coverage:** 9–14 raters per clip. |
| - **Ethics and pay:** same IRB protocol, Prolific recruitment (US |
| residents), consent flow and pay rate as the main study. |
| - **Pseudonyms:** `AC1-###` (one-voice) and `AC2-###` (two-voice), |
| numbered in assignment order. These are different people from the |
| main study and the voice study. |
|
|
| Results built from this folder: the audio-context ablation in Appendix H |
| and Tables 13–18 (see `analysis/README.md`). |
|
|
| ## Files |
|
|
| Each table exists once per slug: `<table>_audiocontext_onevoice.csv` |
| (B1) and `<table>_audiocontext.csv` (B2). The schemas match |
| `data/voice_study/`: |
|
|
| | Table | Contents | |
| |---|---| |
| | `long_<slug>_raw.csv` | every rated clip, attention checks included; same columns as `long_voice_raw.csv` | |
| | `long_<slug>_normalized.csv` | analysis input, with the same normalization columns as `long_voice_normalized.csv` | |
| | `participants_<slug>.csv` | one row per participant record, same columns as `participants_voice.csv`, with demographics blank for records with `Analyzed` = False. `assignedTrialIDs` lists PromptIDs, and Arm is blank for records never assigned | |
| | `justifications_<slug>.csv` | optional free-text justifications, manually reviewed before release (102 for B1, 100 for B2) | |
| | `trial_durations_<slug>.csv` | time on each trial (`TaskDurationMs`), used for the B1 duration split | |
|
|
| Two differences from the voice study: |
| - `Arm` is `audio_context_onevoice` or `audio_context`. |
| - The attention check (PromptID `attention_0`) plays clips |
| `37_0_0.mp3` / `37_0_1.mp3`, where the voice study plays |
| `attention_0_0.mp3` / `attention_0_1.mp3`. In both studies, clip `_0` |
| is the coherent response and `_1` the nonsensical one. |
|
|
| No synthetic ratings were collected for these arms. |
|
|
| ## Provenance |
|
|
| Built with `analysis/01_data_prep/04_convert_followup_json_to_csv.py` |
| from the two studies' Firestore exports and anonymized as described in |
| `data/voice_study/README.md`. |
|
|
