如何微调xtts_v2模型，使它克隆的声音更接近我的声音，我有30分钟自己的数据集

[SMK1311]

请一起讨论。

[eginhard]

How to fine tune the xtts_v2 model so that it clones voices closer to mine, I have 30 minutes of my own dataset

See https://coqui-tts.readthedocs.io/en/latest/models/xtts.html#training

[SMK1311]

微调数据集，我可以将配置的采样率调大吗？对原模型有影响吗

[eginhard]

微调数据集，我可以将配置的采样率调大吗？对原模型有影响吗
When fine-tuning a dataset, can I increase the sampling rate of the configuration? Will it affect the original model?

No, you should not change the sampling rate.

[SMK1311]

我想请问微调的时候，我的输出 grad_norm: 0 (0.0)，avg_loss: 也在缓慢升高，是什么原因导致的

[SMK1311]

EPOCH: 10/1000
--> /aia-storage/root/liyingjun/TTS/run/training/GPT_XTTS_v2.0_LJSpeech_FT-May-27-2025_07+19PM-dbf1a08a

TRAINING (2025-05-27 19:29:05)

--> TIME: 2025-05-27 19:29:06 -- STEP: 0/100 -- GLOBAL_STEP: 1000
| > loss_text_ce: 0.033316127955913544 (0.033316127955913544)
| > loss_mel_ce: 1.8905107975006104 (1.8905107975006104)
| > loss: 1.9238269329071045 (1.9238269329071045)
| > grad_norm: 0 (0)
| > current_lr: 5e-06
| > step_time: 0.5222 (0.5221898555755615)
| > loader_time: 0.3728 (0.3727602958679199)

--> TIME: 2025-05-27 19:29:38 -- STEP: 50/100 -- GLOBAL_STEP: 1050
| > loss_text_ce: 0.02736746519804001 (0.02858249008655548)
| > loss_mel_ce: 1.6650636196136475 (1.6495712041854858)
| > loss: 1.692431092262268 (1.6781536984443663)
| > grad_norm: 0 (0.0)
| > current_lr: 5e-06
| > step_time: 0.267 (0.4024742269515991)
| > loader_time: 0.0072 (0.008863186836242674) 有解决办法吗