How we built a custom voice for a kids device with Fish Audio drama-3

NERDDISCO · x · 2026-10-29

• After learning ElevenLabs' terms didn't cover their computer for kids aged 6–12, the team benchmarked open-weight TTS models and picked Fish Audio's drama-3.
• Their voice-creation pipeline: ① prompt-design a voice persona and generate 26s samples until one clicks; ② clean a 10–30s "golden clip" (30s cloned best) with EQ if needed; ③ clone it into a 30KB voice.json — weights unchanged, the file is the voice; ④ serve it on an RTX PRO 6000, where drama-3 reads the voice codes and continues writing new ones, decoded back to streamed audio.
• An illustrated infographic walks through the full design→clone→serve pipeline.

Original post →