spicysama
> @AnyaCoder 我现在有数据集，我应该怎么做预训练呀 将音频打好标，每个音频旁边放`.lab`后缀的标注文件，里面仅含有标注文本。 然后用我们文档里的训练命令即可，去除+lora选项。注意调整学习率等参数。
@mst-rajatmishra Any update?
Well...
Hi, everyone. We can fix this issue by modifying the file `tools\generate.py`. ```python # line ~ 244 - empty = torch.empty((codebook_dim, max_new_tokens), dtype=dtype, device=device) + empty = torch.empty((codebook_dim, model.config.max_seq_len), dtype=dtype,...
We have written a tutorial for this. https://speech.fish.audio/finetune/
In `tools/generate.py` ```python # line ~ 608 @click.option("--max-new-tokens", type=int, default=1024) # change here to 1024 ``` It will be fixed soon.
Yes, certainly. :)
> >  使用的是: 参考音频的方式生成语音, 但是一直卡在processing.. 控制台窗口加载到一下信息后就没有后续了 To create a public link, set `share=True` in `launch()`. You are using the latest version of funasr-1.1.2 2024-07-19 12:28:59,203 - modelscope - WARNING...