Personalized-Hearing: A Benchmark for Long-Term Personalization in Hearing

Benchmarking use only. The data in this dataset is intended for benchmarking purposes only.

Personalized-Hearing is a benchmark that tests whether an assistant for a wearable hearing device can recall what its user heard, saw, and asked for across many interactions. The benchmark has 104 synthetic users. Each user has 60 sessions spread over one to three years. Every session is paired with a ten-second audio-visual clip from AudioSet and a dialogue between the user and the assistant. The benchmark has 7,965 four-option multiple-choice questions on top of these sessions.

synthetic users	104
sessions	6,240 (60 per user)
distinct AudioSet clips	3,948
questions	7,965
question types	11, in four families: audio, visual, personalization, hearing adjustment

No real users took part. All user profiles, histories, and dialogues are synthetic. Users 001–100 were sampled with hearing profiles at population prevalence. Users 901–904 were added to cover audiogram shapes that are too rare to appear among 100 sampled users; they are marked "cohort": "coverage" in persona.json.

What is in this dataset

path	contents
data/<user_id>/	one directory per synthetic user (see Data format)
data/profiles/persona.json	profile, hearing profile, and speaking style of all 104 users
data/profiles/sound_profile.json	the everyday sounds of users 001–100
scripts/fetch_clips.py	rebuilds the referenced AudioSet clips (frames and audio) from YouTube
eval/run_eval.py	runs the benchmark against a model behind an OpenAI-compatible API
eval/score.py	accuracy per question type, per category, and overall

This dataset does not contain AudioSet audio, video, or frames. Sessions refer to clips by their AudioSet clip ID. To obtain the media, use the scripts in scripts/ together with the AudioSet segment CSVs, which give each clip's YouTube ID and start time. Some clips may no longer be available on YouTube.

Running the benchmark

pip install openai yt-dlp            # ffmpeg must also be on PATH

# 1. Rebuild the clips (download the three segment CSVs from the AudioSet site first)
python scripts/fetch_clips.py --data-dir data --out-dir media \
    --segments eval_segments.csv balanced_train_segments.csv unbalanced_train_segments.csv

# 2. Ask the model every question
export OPENAI_API_KEY=...
python eval/run_eval.py --model <model> --base-url <endpoint> \
    --data-dir data --media-dir media --out-dir results/<model>

# 3. Score
python eval/score.py results/<model>

For each question the model receives every earlier session of that user as a multi-turn conversation, plus the current session up to the question's turn cutoff. The first user turn of each session carries the session's clip: 10 frames spread evenly across the clip (scaled to 384 pixels wide) followed by the audio (16 kHz mono MP3). The four options are shuffled with a fixed seed per question, and the model gives its choice after the token <final_answer>. Accuracy is correct / (correct + wrong); replies that name no single choice and failed requests are reported separately. These are the settings used for the full-context results in the paper.

Data format

Each data/<user_id>/ directory holds the outputs of the construction pipeline:

file	contents
1_general_history.json	60 life events (30 long-term, 30 short-term)
2_preference.json	four listening preferences
2.1_preference_grounding.json	the life events each preference is anchored to
3_preference_history.json	30 preference events (establish, reinforce, drift, exception)
4_session_plan.json	per-session scene and activity plan
5.2_paired_captions.json	the AudioSet clip and captions paired with each session
6_grounded_session_plan.json	session plans rewritten to match the paired clip
7.1_dialogue_enhanced.json	user–assistant dialogue for each session
8_av_facts.json	audio and visual facts extracted from each clip
10_benchmark_qa.json	the multiple-choice questions for this user

Each question records its type, the correct answer, three incorrect answers, the session and turn at which it is asked (question_location), and the sessions its answer depends on (Evidence). The dialogue history before question_location is the question's context.

Question types

family	type	Type in the data	count
Audio	Sound Presence	audio_presence	1,040
Audio	Audio Scene Recall	audio_only_recall	945
Audio	Audio Occasion	audio_occasion	818
Visual	Visual Detail Recall	visual_detail_recall	621
Personalization	Past Preference Recall	preference_state_at_time	998
Personalization	Preference Change Count	preference_change_count	992
Personalization	Preference Replacement	preference_replaced	622
Personalization	Preference Hallucination	preference_never_asked	386
Personalization	Exception Recall	exception_recall	311
Personalization	Exception Expiry	exception_expiry	192
Hearing adjustment	Adjustment Suggestion	hearing_adjustment	1,040

Source data and attribution

• AudioSet clip identifiers and labels, © Google LLC, licensed under CC BY 4.0. The AudioSet ontology is not used.
• Sound-VECaps captions (Yuan et al., arXiv:2407.04416; Zenodo 12606207), licensed under CC BY 4.0. Captions were filtered to remove personal names and identifying text.
• PersonaHub (Ge et al., arXiv:2406.20094) occupation descriptions seeded the synthetic user profiles.

Citation

License

This dataset is licensed under the Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0), as found in the LICENSE file.