Fish Audio: finding a scared/sad voice — search by tag, not title, and drop sobbing voices that blow up duration
Title search on GET https://api.fish.audio/model (title="scared", "crying") mostly returns character or celebrity clones, which you should avoid, or unrelated popular voices. Filtering by tag works much better: tag=scared|anxious|worried|nervous|emotional|sad with language=en and sort_by=task_count, keeping only items whose tags include "female". Each item's description often states the delivery (e.g. "Near-panic, breathless, adrenaline surging").
Preview with the real line, and keep the emotion tags in square brackets for s2.1-pro (e.g. "[scared][panting] ... [break] ..."). Then compare candidates objectively if you can't listen: duration vs word count, median F0, F0 spread, voiced ratio (lower means more breath or panting), longest silence, peak. A voice tagged Crying/Sobbing took 14 s for a 12-word line, which is unusable in a tight edit. A "near-panic, breathless" voice gave 5 s with a voiced ratio around 0.8, which read as scared. A general young-female library voice with a high F0 still came across as happy, so the tags matter more than pitch.
Cache the TTS by sha256(model + payload) so rebuilds never respend credits, and count requests against a hard budget.