Chatterbox VC — expressivity×quality reward, best-of-N & SIDON

Target voice Measured Slavic Historian (emolia_c0542). 52 sources (40 intense·free EmoNet emotions + edge-case bursts) → 32 seed-varied Chatterbox conversions, each SIDON-refined and scored on emotion (Empathic-Insight), quality (DNSMOS-distilled) and speaker similarity (ECAPA + Orange timbre). A reward then re-ranks candidates by how expressive AND clean they are.

1 · Reward ranking — listen

Headline row per card = reward winner (Chatterbox vs. +SIDON ★), then best-by-Quality and best-by-emotion-MAE for reference.

40 intense · free-flowing emotions

Edge cases — screams, whimper, crying, groan, shiver, laughter

2 · Diminishing returns — how many candidates?

Expected best-of-k reward (+SIDON) across 52 groups, as % of the full 1→32 gain. This reward deliberately chases the most intense take, so — unlike quality, which saturates by k≈8 — it keeps climbing into the tail: k=4 → 51%, k=8 → 71%, k=16 → 87%. Recommended operating point k=8 (marginal gain drops off after); the strict 90%-of-gain knee only lands at k=32.

3 · Cost — GPU-hours for 1,000,000 samples (SIDON + ranking)

Measured per-candidate (ref clip 5.1s output): generation 182 ms, SIDON 255 ms, scoring 9 ms (all 40 emotion + quality experts, batched). SIDON is the dominant cost, so the efficient design ranks first, then SIDON only the winner (×1 not ×N).

4 · Global reward leaderboard (top 15, +SIDON pool)

5 · Best-of-k, SIDON effect & throughput (reference)