Fine-tuning of multiple speakers

[todochenxi]

你好，我想咨询一下使用多个speaker的数据去微调模型的流程和结果，我疑惑的点如下：

1. 数据集的trin_list.txt 的格式是什么样的，
2. config_ft.yml 中root_path要如何配置呢
3. 如果我是用三个speaker的数据去微调，得到的模型可以指定speaker去推理嘛
4. 假设可以的话，那是否可以在单个speaker的基础上去扩展，添加另外的声音呢。。。

Hello, I would like to ask about the process and results of using data from multiple speakers to fine-tune the model. My doubts are as follows:

1. What is the format of trin_list.txt of the data set?
2. How to configure root_path in config_ft.yml
3. If I use the data of three speakers for fine-tuning, can the resulting model specify the speaker for inference?
4. If possible, can it be expanded based on a single speaker to add additional sounds. . .

[todochenxi]

Oh, I found OOD_texts.txt, I think 1 and 2 have been solved :)

[todochenxi]

Oh, for 3, I suddenly thought that it can be achieved by just changing the reference audio. . .

[todochenxi]

For 4, can we continue to fine-tune based on the previously fine-tuned model? Or do you have to retrain every time a new character is added? . .

[todochenxi]

Thank you for your reply. Last Friday I did an experiment to fine-tune the data sets of two speakers. By switching reference audio, I achieved the purpose of switching speakers. The results are stunning and the work has so much potential. Thank you for your work, salute ~ :)

[junylee11]

How long was the data used for fine tuning?

[todochenxi]

30m

[todochenxi]

I just tried it briefly, you can customize your data

[junylee11]

Thank you!