陳宜昌 Yi-Chang Chen (YC)

It's only lowercase currently. (https://github.com/GitYCC/crnn-pytorch/blob/master/src/dataset.py#L14) You can change dataset to train your case sensitive model. (https://github.com/GitYCC/crnn-pytorch/blob/master/src/dataset.py#L13)

@BarryKCL Our model is trained on the hyper-parameter `window_size = 32`. It might slightly affect performance after changing this hyper-parameter.

@yt605155624 請問具體是怎麼做的？

Wow~ Great work

@BarryKCL Thanks for your effort! And please reference [GitYCC](https://github.com/GitYCC)/[g2pW](https://github.com/GitYCC/g2pW) in https://github.com/PaddlePaddle/PaddleSpeech/blob/develop/docs/source/reference.md Thank you!

You can reference UNIFORM DEQUANTIZATION in [this paper](https://arxiv.org/pdf/1902.00275.pdf).

> 简单增加超时时间似乎不能解决问题，测试了下是卡在logging step上了，应该是其他rank等待rank 0计算loss时卡死了，暂时把logging steps 设置成1e9。另外运行日志看起来很奇怪，会有多个进度条。logging step设置为20的时候进度条为： > > ```shell > 0%|▏ | 1/1170 [00:18 06/25/2023 07:35:18 - INFO - torch.nn.parallel.distributed - Reducer buckets have been rebuilt in this iteration....

目前可下載的模型是基於繁體中文資料所訓練而成，其資料品質較好，而資料不開源。 而資料的格式採用和g2pM相同的格式，因此你可以自行使用g2pM訓練一個簡中模型，訓練方法如README所示。

please refer: https://github.com/GitYCC/g2pW#load-offline-model

你可以從pretrained model進行微調