Seeking Guidance for Finetuning on One Voice (small dataset 100min)

[Balauruu]

I've been getting good results with finetuning, but I have some questions, especially regarding the training parameters. I am getting some artefacts (a wet sound, gargling like) and I'm not sure if the model is undertrained or my reference audio files are bad and it's picking too much background noise. Also on my machine (4070 laptop) it takes 1 day for around 20k-25k steps depending on parameters and I can't figure out if it's performing normally or slow. (I see ppl here training for >100k steps).

Project Details:
I have 60-120 minutes of quite clean audio (barely noticeable background noise, only when listening at 100 volume) . Language: English.

GPU: Nvidia 4070 Laptop (8GB VRAM)

Database:
825 samples (1 hours and 15 min of audio

My Parameters + Questions:
-learning_rate: 0.00001,
-batch_size_per_gpu: 2200 ( I found gives me the best training time per epoch)
-max_samples: 64,
-grad_accumulation_steps: 1,
-max_grad_norm: 1,
-epochs: 3425 ( I left it this high, because I thought it doesn't affect training. Just the slope in results when using the tensorboard to analyze?)
-num_warmup_updates: 100
-save_per_updates: 2000,
-finetune: true,
-mixed_precision: "bf16",
"logger": "tensorboard",

8 bit optimizer: false (not enabled) (from what I researched online it should be a straight speed upgrade, but I saw people in other discussion boards are not using it when training)

I actually trained my model for only 150 epochs, 30k steps for now.

[cyberbol]

Think This repo dies as no ones replying here anymore