Emotions on a cloned voice: a 25 MB graft, a steering vector, and a lot of dead ends

How we got sad / happy / angry / fearful speech to work on cloned voices in a pure-C Qwen3-TTS engine — a 25 MB 'graft' clone that keeps the emotion levers alive, a steering + fine-tune recipe hard-won after many dead ends, and — because emotion is a vector — blended 'dyad' emotions plus switching emotion mid-sentence from a single prompt.

Raccontata dadev.to