I'm learing from your project. I want sone help about custom dataset, vocabulary, and retrain from zero.

Checks

• This template is only for usage issues encountered.
• I have thoroughly reviewed the project documentation but couldn't find information to solve my problem.
• I have searched for existing issues, including closed ones, and couldn't find a solution.
• I am using English to submit this issue to facilitate community communication.

Environment Details

Kaggle notebook: Python 3.12.2 Torch version: 2.8.0+cu126 I use GPU P100.

Steps to Reproduce

Sorry if I'm annoying. I want to understand your project so I'm trying to clone your project in Kaggle for english language only since i don't have any local computer or any server for training. I clone most of your modules and function. I only adjust dataset module and vocabulary. My vocabulary: symbols = [ ' ', '!', ',', '-', '.',
';', '?', 'a', 'b', 'c', 'd', 'e', 'f',
'g', 'h', 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q',
'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', "'", '“', '”'
] print(len(symbols))

vocab_char_map = {} for i, char in enumerate(symbols): vocab_char_map[char[:]] = i vocab_size= len(symbols)

My config: based on your small config.

mel_spec_kw = dict( target_sample_rate= 24_000, n_mel_channels= 100, hop_length= 256, win_length= 1024, n_fft= 1024, mel_spec_type= 'vocos', )

model_arch = dict( dim= 768, depth= 18, heads= 12, ff_mult= 2, text_dim= 512, text_mask_padding=False, conv_layers= 4, pe_attn_head= 1, attn_backend= 'torch', attn_mask_enabled= False, checkpoint_activations= False )

My dataset is Librispeech-100h with normalized text file. I also removed prepare dataset functions and logging.

You can review my notebook here: https://www.kaggle.com/code/nguyenquoctuan12/f5tts-small

I'm still learning. I really appreciate your help.

Sorry for my bad English. Have a nice day!

✔️ Expected Behavior

A good generated wav and checkpoint.

❌ Actual Behavior

I'm currently reaching 130k step. The loss is around 0.6. I think the generated wav had ref wav's accent, vocal but i cant hear any word from it. I know you said the minimum step is 200k+ but i don't have much resource for training. Is there anything i did wrong.