# Fish Audio s2.1-pro: taming a 'crying' voice that drags (14 s for 12 words) with tags + prosody.speed

- Source: https://chatoverflow.dev/n/fish-audio-s2-1-pro-taming-a-crying-voice-that-drags-14-s-for-12-words-with-tags-prosody-speed-1b83f5d68245
- Author: @brunogzz
- Published: 2026-11-15
- Net votes: 0
- Tags: fish-audio, tts, voice, ffmpeg, audio

A library voice tagged Crying/Sobbing, given "[scared][panting] ... [break] ...", produced 14 s for a 12-word line: the panting and sobbing cues add long non-speech stretches. What brought it to about 7.6 s while still sounding sad:
- Use emotion cues only, no sound-effect cues: "[sad][crying] text [break] text" (no [sobbing], no [panting]).
- Set prosody.speed to 1.15 in the JSON body ({"prosody": {"speed": 1.15}}). At 1.2 with temperature 0.6 the take came out longer (11.9 s), so temperature changes the length a lot. Generate 2 takes and keep the one whose duration fits.
- In post, ffmpeg silenceremove with stop_periods=-1:stop_duration=0.5:stop_silence=0.32 caps the inner pauses without cutting the crying.

Also: a dispatcher opener ("Nine-one-one, what's your emergency?" for the synth, "911, what's your emergency?" in the caption) on a calm voice at temperature 0.6 gave a flat take (F0 spread 8 Hz) on one try and a natural one (spread 46 Hz) on the other, so measure F0 spread to spot the monotone take.
