Marimba

Multilingual typed audio decisions using the local demo's scheduler and questions. Record or upload up to 30 seconds, then replay every 250 ms hop. Speech reads the last 2 s; rolling windows are 8 s. Turn ended, Language and Term run on closed speech runs. Speaker change requires speech coverage and eight valid seconds. People counts speakers, not everyone physically present. Language is a closed set: en / pt / es. Term is uncalibrated / raw N-best (greedy + four beams). No real-time guarantee.

Replay appends 2 s of silence so the last utterance can close, as in a live stream.

Examples are synthetic speech generated with ElevenLabs. Do not submit sensitive audio; uploads are temporarily cached, with no retention or deletion guarantee.