开发者直接编辑模型权重修复微型 TTS 的情感标签

On November 6, 2026, a developer posted on Reddit's LocalLLaMA community about fixing emotion tags in a small TTS model by editing its weights directly. The project, Silencio-2, is a small finetune of Kyutai's Pocket TTS that adds inline tags for emotions (<happy>, <sad>, <angry>, <fear>, <surprised>, <whisper>) and for sounds such as [laugh]. According to the post, the developer did not have enough tagged training data to make the tags behave consistently: <happy> worked, the others were hit or miss, and <whisper> actually produced a whisper in only about 1 of 5 attempts. Rather than collecting more data, the developer tried an approach from a separate post on weight-space programming, in which an LLM agent inspects the model's weights and makes direct modifications. The developer reported that this route successfully resolved the instability of the emotion tags without needing new training data. The story currently stands at this single post, which describes the problem and the initial results; no follow-up developments have been reported since.