Recurring speakers
Participants return across sessions with stable voices, creating an identity thread through the history.
A BENCHMARK FOR MULTI-PARTY SPOKEN MEMORY
VoxPolyBench tests long-term memory in spoken histories where the same people return, facts change, and the answer depends on who said what to whom.
Controlled, synthesized speech across recurring participants and sessions.
01 / THE BENCHMARK
Each scenario follows a recurring group through 8 to 12 sessions. A system must connect voices and participants over time, track changing information, and answer questions grounded in the shared history.
Participants return across sessions with stable voices, creating an identity thread through the history.
Plans, preferences, and facts are updated, contradicted, and resolved over multiple conversations.
Questions may depend on the source speaker, the addressee, or who should receive a response.
The same shared history can support different answers for different participants.
02 / DATASET PROFILE
These figures describe the accepted 18-scenario catalog. The speech is synthesized from fixed participant reference voices; the benchmark evaluates controlled spoken-memory conditions.
03 / CASE STUDIES
These examples are drawn from two accepted histories. Questions and reference answers are reproduced from the benchmark, not written as illustrative prompts.
A family and travel contacts coordinate trains, hotel rooms and activities as plans evolve across the trip.
Reference answerTessa Kim
Reference answera quiet room away from the elevator; a 3 p.m. walking tour
Reference answera 3 p.m. walking tour
Reference answera written itinerary
Checking the audio package…
A shared renovation history where decisions, dates and responsibilities change across conversations.
Reference answerThe kitchen island outlet count was changed from six to eight outlets.
Reference answerThe kitchen island has eight outlets, and the change order ID is CO-17.
Reference answerCarla Rossi
Reference answerI said I would handle the notifications.
The complete benchmark has 18 histories. This page highlights two; the family-trip history is the selected audio example with dialogue and ten synthesized personalized queries.
04 / DATA ACCESS
This project repository hosts the selected audio demo. For dataset files and release updates, visit VoxPolyBench on Hugging Face. The Arrow release separates dialogue content and QA samples by case, with audio embedded in the dataset files.
GitHub Pages / ├─ index.html ├─ assets / figures └─ assets / audio / selected-story ├─ demo_manifest.json ├─ S1–S8.mp3 └─ q01–q10.mp3 Full dataset release / └─ cases / 18 case directories ├─ full_case.json ├─ audio / turn-level WAV └─ audio_persona / QA WAV
The compressed case demo does not replace the canonical turn-level audio.