aabroukh's picture
Add files using upload-large-folder tool
b96b491 verified
|
Raw History Blame Contribute Delete
5.68 kB
# Data dictionary: `data/voice_study/`
Follow-up study: does the audio–text difference survive a change of TTS
voice? The response text is held fixed and only the voice changes.
Participants were assigned to one voice arm and heard it on every trial.
The main study's audio condition (Kokoro `af_heart`) is the baseline and
was not re-run.
| Arm | Engine / voice | Analyzed raters | Label in the paper |
|---|---|---:|---|
| `kokoro_am_michael` | Kokoro-82M, `am_michael` | 15 | A1 |
| `chatterbox` | Chatterbox, default voice | 16 | A2 |
| `qwen3tts_serena` | Qwen3-TTS-12Hz-1.7B-CustomVoice, `Serena` | 17 | A3 |
- **Stimuli:** 30 of the main study's 100 prompts
(`stimuli/voice_study/prompts.csv`), presented in six fixed blocks of
five.
- **Session:** each participant rated four blocks (20 trials) plus one
attention check.
- **Coverage:** 8–12 raters per clip.
- **Ethics and pay:** same IRB protocol, Prolific recruitment (US
residents), consent flow and pay rate as the main study.
- **Pseudonyms:** `V001..V066`, numbered in assignment order. These are
different people from the main study's `P001..P094`.
Results built from this folder: the voice ablation in Appendix G and
Tables 9–12 and 15–18 (see `analysis/README.md`).
## Human ratings
### `long_voice_raw.csv`
One row per (participant × rated clip), every arm, including attention
checks and participants who later failed them.
| column | type | description |
|---|---|---|
| PromptID | str | `<conv_id>_<turn_no>`, same ids as the main study; `attention_0` for the attention check |
| ParticipantID | str | `V001`..`V066` |
| AudioClip | str | `<PromptID>_<0|1>.mp3`; the same clip id names the same response text in every arm and in the main study |
| TrialNumber | int | 0 = attention check, 1–20 = real trials |
| IsAttentionCheck | bool | |
| PresentedFirst | bool | this clip was played first on the trial |
| Rating | int | 1–100 response-quality rating |
| qualityRating | int | 1–100 audio-quality rating |
| overallPreference | str | A / Tie / B, where A is the first-played clip |
| Arm | str | voice arm (table above) |
### `long_voice_normalized.csv`
The analysis input: real trials only, with participants who failed the
attention check removed. It adds the main study's normalization columns,
computed within each arm:
| column | description |
|---|---|
| L | shrunk logit of Rating, `log(p / (1 − p))` with `p = (Rating + 0.5) / 101` |
| L_adjusted | L with the arm's mean position bias removed from second-played clips |
| position_bias | that arm's mean position bias (second-played minus first-played L) |
| user_mean, user_sd | participant's mean and SD of L_adjusted |
| L_centered | L_adjusted − user_mean |
| z | L_centered / user_sd (0 when user_sd is 0) |
| Pref_Num | 1 = this clip preferred, 0.5 = tie, 0 = the other clip |
### `participants_voice.csv`
One row per participant record, including people who started and never
finished. Demographic columns (Age through Education) are blank for
records with `Analyzed` = False; their status and session columns are
kept so completion rates can be reproduced.
| column | description |
|---|---|
| ParticipantID, Arm | pseudonym and assigned voice. Arm is blank for 6 records that signed in but were never assigned |
| Started, Completed | reached the first trial / finished the session |
| PassedAttention | True / False, or blank if the participant never reached the attention check |
| Analyzed | contributes rows to `long_voice_normalized.csv`. Five partial sessions that passed the attention check are included, as in the main study |
| TotalSessionDuration | ms |
| Age, Gender, Race, Ethnicity, EnglishProficiency, Country, Education | self-reported, same categories as the main study |
| postSurveyDifficulty | 1–5 |
| assignedTrialIDs | comma-separated `<arm>__<PromptID>` trial ids |
### `justifications_voice.csv`
Optional free-text justifications (183): `PromptID`, `ParticipantID`,
`Arm`, `TrialNumber`, `overallPreference`, `Justification`. Each was
manually reviewed before release.
### `trial_durations_voice.csv`
`ParticipantID`, `PromptID`, `TaskDurationMs`: time on each trial, used
for the feasibility and cost analyses.
## Synthetic (AI) ratings: `synth/`
Gemini 3.6 Flash and Qwen 3.5 Omni Flash rated every one of the 100 main
study prompts rendered in each new voice, with the main study's audio
protocol: 10 generations per pair, order counterbalanced, waveform in
directly. The judges were not run on text here, since the text is
unchanged from the main study. The baseline voice's ratings are the main
study's own (`data/main_study/synth_long_audio_table_<model>.csv`).
- `synthetic_ratings_<arm>_<model>.csv`: raw wide output, one row per
prompt. Same columns as `data/main_study/synthetic_ratings_10x*.csv`,
plus `arm`. The audio blocks are `response_rating_{a,b}_<i>`,
`audio_quality_rating_{a,b}_<i>`, `justification_<i>` and
`audio_order_a_first_<i>`. The `text_*` columns are empty.
- `synth_long_audio_table_<arm>_<model>.csv`: the long form (`PromptID`,
`AudioClip`, `PresentedFirst`, `Rating`, `qualityRating`, `synthNo`,
`model`), produced by `analysis/01_data_prep/03_synth_to_long_csv_multi.py`.
The judges' inputs (the prompt rows sent per voice) are in
`stimuli/voice_study/synth_inputs/`.
## Provenance
Built from the study's Firestore export with
`analysis/01_data_prep/04_convert_followup_json_to_csv.py`, then
anonymized. Dropped fields: Firebase UIDs, Prolific IDs, every
timestamp, post-survey free-text feedback, and demographics the main
release doesn't carry (birth country, timezone, employment, languages,
"Other" free text).