只有 CPU的情況, 感覺 web 效果比 toolbox 好一點

web 和 toolbox 都可以運行
但感覺 web 效果比 toolbox 好一點

toolbox 的電流音比較明顯而多
而且有時候輸入的是男聲, 他輸出的變成女聲了

已有下載 dataset, 是用最新版本, 用 pretrained的 data.

相似度仍是有很大的距離,是要自己 train 嗎?

輸入的 wav 再長一點有用嗎? 通常多久最好?

Start with the web and toolbox CPU execution paths and reproduce the reported electrical noise and voice-gender mismatch using the pretrained dataset and WAV input described. Compare their outputs and document whether the difference is reproducible, including input duration and whether training is required; the issue is done when the cause or a concrete reproduction and next step is established.

Written by the indexing model from the issue text.