Jev Audio

Sammyliu's audio decision corpus for reinforcement learning and evaluation: causal audio windows paired with runtime action questions, each receiving a categorical label distribution. Train holds 168,494 decisions over 71,952 source windows; test, dev, calibration and an external diagnostic add about 122,600 more.

A Python sidekick that listens to a call (FaceTime, Zoom, Discord or a stream) and fires a soundboard clip when Jev thinks the moment is funny. Jev answers two questions per 1.2-second audio slice, a Noul for react-now and a Choice for which clip, at about $0.30-0.40 per hour of chatter.

A Tern plugin that gives coding-agent notifications a tiny robot voice: Jev reads the end of the agent's last message and answers six typed questions, including mood and pitch, that a pure-Luau synthesizer turns into a short beep.

An audio-native decision model shaped like Jev: audio is encoded once, several typed questions branch over that single encoding, and answers come back as calibrated probability distributions — no transcript, no ASR, the Whisper decoder never runs.

A podcast chaptering tool that labels transcript windows and marks boundaries as separate local judgments, classifying content, sponsor, and housekeeping segments without generating prose for the titles.

Private TypeScript proof of concept that flags Spanish profanity with Jev Boolean and Score questions and replaces the matching word intervals with a beep via ffmpeg, keeping the original duration.