VITS fine-tuning

[madey83]

hello,

i have started my VITS fine-tuning based on pretrained model "tts_models--pl--mai_female--vits" with 660000 steps.

My best model was marked at 728037 step. wav file sample:
https://mega.nz/file/MBxxmTwD#aOEoqW9WfLbyr7OGzl0pz41OqSmLSPOF_MkVNgUlu9Y

After 764469 steps audio quality does not improve at all. Sample:
https://mega.nz/file/pFZkATRS#-QAsv1t5tCQPSnKu0okqhtLty4afDWO6RX4wEiquMxA

this is my VITS config:

config = VitsConfig(
	audio=audio_config,
	run_name="Vits-TTS-PL",
	batch_size=12,
	eval_batch_size=6,
	batch_group_size=5,
	num_loader_workers=4,
	num_eval_loader_workers=2,
	run_eval=True,
	test_delay_epochs=-1,
	epochs=1000,
    save_step=5000,
    save_checkpoints=True,
    save_n_checkpoints=2,
    save_best_after=10000,
    add_blank=True,
    #max_audio_len=        
	text_cleaner="multilingual_cleaners",
	use_phonemes=True,
	phoneme_language="pl",
	phoneme_cache_path=os.path.join(output_path, "phoneme_cache"),
	compute_input_seq_cache=True,
	print_step=25,
	print_eval=True,
	mixed_precision=False,
	output_path=output_path,
	datasets=[dataset_config],
	cudnn_benchmark=False,
	test_sentences=[""
	],
)

i have 18hrs audio in over 18k waves.

What i need to change to improve quality over steps.

in this https://github.com/coqui-ai/TTS/blob/dev/recipes/multilingual/vits_tts/train_vits_tts.py i have found these two settings:

min_audio_len=32 * 256 * 4,
max_audio_len=160000,

How to calulate length for my audio samples? I have waves length <0s and waves length < 30s....

Thank you in advance...