laion/emotional-transitions-listening

Emotional transitions — listen and judge

399 clips from a study of how to make a synthetic voice turn from one feeling into another inside two sentences, rather than switching between them at the sentence break.

Fourteen conditions per text, including two references (the clip held entirely at the first feeling, and entirely at the second) and a control (the same amount of pushing on a shuffled schedule with no A→B story). Ten emotion pairs, three texts each, one voice, one seed — only the method changes.

The headline is a measurement problem, not a method result. The scoring model used to judge emotion can barely separate the two references: the same sentence briefed as intense sadness versus intense amusement differs by 0.12 of its own scatter, and 0 of 30 texts reach one full unit. Four methods beat the shuffled control, the multi-rate method wins, and none beats today's step behaviour significantly. All of that sits on an instrument that is nearly blind, which is why this page exists: to find out whether human ears agree.

The page explains every condition in plain language and lets you rate each clip. Ratings stay in your browser; a button copies them out as text.

Full write-up, raw per-cell records, schedules and code: LAION-AI/Voice-Acting-Pipeline-WIP, research-log-2026-08/emotional-transitions/.

Base model: `laion/moss-tts-local-transformer-4.55b-voice-acting-v2-sft3`.