laion/moss-va-sft2-samples

One prompt per emotion, in English and German, rendered in the v2 timed-script format: per-sentence durations, explicit pauses, and every vocal burst in round brackets with its length. 80 prompts x 4 completions = 320 generated clips, each scored for emotion-band achievement, vocal-burst blend, genuineness, burst realisation, word error rate, and duration fidelity.

Prompt text comes only from sources that are already public (emolia, kartoffelphon, mls and the LAION voice profiles). All audio here is generated by the model; none of it is source recording.

Note on the prompts: this grid is the round-2 baseline. Its prompts carry the timed script -- per-sentence durations, explicit pauses, vocal bursts with their lengths -- but no inline delivery directions, because round-2 training contained none. The emotion appears only as description in the GENERAL: line. That is the measured reason the models reach percentile 0.50 against a 0.90-0.98 target on this grid, and round 3 adds the directions back.