sciminds/same-words-different-judgements

Same Words, Different Judgments: How Preferences Vary Across Modalities

Code, data, and stimuli for our study of how people judge LLM responses delivered as text versus audio (TTS). This repository lets you:

• reproduce every result in the paper from the anonymized data shipped here;
• reuse the human and AI ratings, stimuli, and audio for your own analyses;
• deploy the same survey apps to collect new preferences.

Aaron Broukhim, Nadir Weibel, Eshin Jolly — University of California San Diego.

The studies

Study	Question	Participants	Data
Main study	Do text and audio raters judge the same response pairs differently?	94 analyzed (48 text, 46 audio); 100 PRISM prompt pairs, 20 per participant	data/main_study/
Voice study (follow-up)	Does the audio–text difference survive a change of TTS voice? Three new voices: Kokoro am_michael, Chatterbox, Qwen3-TTS Serena	48 analyzed; 30 prompts	data/voice_study/
Audio-context study (follow-up)	Does it survive when the conversation history is heard rather than read? One-voice (B1) and two-voice (B2) history	22 + 23 analyzed; 30 prompts	data/audio_context_study/

All three studies ran on Prolific with US participants, under the same UC San Diego IRB exemption, consent procedure and pay rate. The data also includes AI ratings from three model families (GPT-4o, Gemini 3.6 Flash, Qwen 3.5 Omni Flash):

• every main-study stimulus, in text and in audio;
• every main-study prompt in each new voice, rated by Gemini 3.6 Flash and Qwen 3.5 Omni Flash;
• the full unguided PRISM corpus (10,800 prompt pairs), rated once per pair by Gemini 3.6 Flash in text and in audio (data/main_study/synthetic_ratings_fullcorpus_gemini_36_flash.csv).

Layout

├── analysis/        scripts that produce every table and number (see analysis/README.md)
├── data/            anonymized inputs, one folder per study, each with a data dictionary
├── results/         tables written by the analysis scripts
├── figures/         paper figures and interface screenshots
├── stimuli/         audio and stimulus metadata, one folder per study
├── survey_apps/     the four React + Firebase apps used to collect the data
├── tools/           Firestore export scripts for new collections
├── croissant.json   Croissant 1.1 + RAI metadata for the dataset
├── requirements.txt Python dependencies (tested versions)
└── package.json     Node dependency for the export tools (firebase-admin)

Reproduce the results

pixi install     # Python, R and every package; pip setup is in analysis/README.md

pixi run python analysis/02_main_study/01_omnibus_bias_model.py      # Table 1
pixi run python analysis/03_synthetic_raters/01_ai_human_alignment.py # Table 3
pixi run python analysis/08_cross_study/02_condition_contrasts.py    # Tables 17–18

pixi run bash analysis/run_all.sh    # everything, about 3.5 hours

analysis/README.md maps every table in the paper (Tables 1–21) and the in-text numbers to the script that produces them. Everything runs offline on the shipped data. The reliability analysis takes about three hours of that, and two acoustic scripts need the voice-study WAVs (their outputs already ship).

Use the data

Each data folder has a README describing every file and column: data/main_study/, data/voice_study/, data/audio_context_study/. The stimuli folders do the same for the audio: stimuli/main_study/, stimuli/voice_study/, stimuli/audio_context_study/.

All audio is in this repository:

• the main-study audio (200 MP3s) and its PRISM provenance, in stimuli/main_study/audio/;
• the voice-study WAVs, in stimuli/voice_study/audio/ (about 1.3 GB);
• the audio-context WAVs, in stimuli/audio_context_study/audio/ (about 0.6 GB);
• a 1,000-clip sample of the full Kokoro corpus, in stimuli/main_study/audio_full_sample/ (about 1.6 GB);
• the full 21,603-clip Kokoro corpus rendered during stimulus selection, in stimuli/main_study/audio_full_zips/ (14 zips, about 26 GB).

To download everything except the 26 GB corpus:

hf download sciminds/same-words-different-judgements --repo-type=dataset \
  --exclude "stimuli/main_study/audio_full_zips/*" --local-dir same-words-different-judgements

Drop --exclude to get the full corpus too.

Participant IDs are pseudonyms (P###, V###, AC1-###, AC2-###). Prolific IDs, Firebase IDs, timestamps and post-survey feedback are not included. Every free-text justification was manually reviewed before release.

Collect new data

survey_apps/README.md explains how to run and deploy each app on your own Firebase project and export what it collects. data/your_study/ walks through running the main-study analyses on a new collection.

cd survey_apps/main_study
cp .env.example .env && $EDITOR .env    # your Firebase web config
npm install && npm start

Citation

@inproceedings{broukhim2026samewords,
  title     = {Same Words, Different Judgments: How Preferences Vary Across Modalities},
  author    = {Broukhim, Aaron and Weibel, Nadir and Jolly, Eshin},
  booktitle = {Advances in Neural Information Processing Systems: Evaluations and Datasets Track},
  year      = {2026}
}

Please also cite PRISM (Kirk et al., 2024), the source of the conversations.

License

• Code (analysis/, survey_apps/, tools/): MIT.
• Data and stimuli (data/, results/, stimuli/): CC BY-NC 4.0.

The non-commercial term comes from PRISM, which licenses its human-written text under CC BY 4.0 and its model responses under CC BY-NC 4.0. This release contains model responses and TTS renderings of them, so the more restrictive license applies. Under PRISM's terms, reuse of the model responses must also follow the original model providers' terms.

The TTS models allow redistribution of their output: Kokoro-82M (Apache-2.0), Chatterbox (MIT), and Qwen3-TTS (Apache-2.0). Chatterbox audio carries Resemble AI's inaudible PerTh watermark. See LICENSE.

Contact

Open a discussion in this repository's Community tab, or email aabroukh@ucsd.edu.