Cantonese needed!

[huangkun1985]

Please vote for Cantonese!!!

[SWivid]

b（￣▽￣）d　

[wong813]

Yes Cantonese needed!!!

[ringolam]

Strongly request to have Cantonese language

[elbartohub]

Yes, one of the most powerful language on Earth

[pebblehack]

It's not an issue but- Cantonese would be neat.

[indiejoseph]

+1

[chau9ho]

I have fine tuned a Cantonese model with F5-TTS

Dataset Details

• Dataset: Common Voice Corpus 19.0
• Validated Hours: 109 hours
• Training Steps: ~190K steps

  Current Results

  The model can generate Cantonese speech, but has following issues:

• Strong Mandarin (Putonghua) accent
• Not natural Cantonese pronunciation
• Some characters still pronounced in Mandarin despite being in vocab.txt
• Numeric characters not pronounced

Training Configuration

{ "exp_name": "F5TTS_Base", "learning_rate": 7.5e-05, "batch_size_per_gpu": 12000, "batch_size_type": "frame", "max_samples": 64, "grad_accumulation_steps": 4, "max_grad_norm": 0, "epochs": 50, "num_warmup_updates": 2000, "save_per_updates": 20000, "last_per_steps": 500, "finetune": true, "file_checkpoint_train": "", "tokenizer_type": "char", "tokenizer_file": "", "mixed_precision": "fp16", "logger": "tensorboard" }

Questions

Character/Vocab Issues:

Why are some characters skipped despite being in vocab.txt? How to improve numeric character handling? Best way to enforce Cantonese pronunciation for all characters?

Training Concerns: Would more training steps help with pronunciation? Is the dataset quality affecting accent/pronunciation? How to reduce the Mandarin accent influence?

Looking for: Solutions for character skipping issues Methods to improve numeric character handling Community experience with similar issues