executedone / chinese-fastspeech2 Goto Github PK

在使用biaobei.py里的get_char_embedding_from_bert生成expanded_embeds.pkl时显示如下，该怎么解决呀？
size mismatch for bert.embeddings.position_embeddings.weight: copying a param with shape torch.Size([512, 1024]) from checkpoint, the shape in current model is torch.Size([512, 768]).

请问能提供韵律模型训练的训练流程和代码吗？

本人小白一枚, 录制了自己的声音集, 想知道您是怎么样进行微调的? 具体的步骤是什么? 如果方便的话, 可以加微信 598131440, 我非常想学, 您会帮了我一个大忙的, 本人正在做本科毕业项目, 感激不尽感激不尽

您好，请问有英文训练的prosody bert做韵律增强吗？例如ljspeech, libritts

Traceback (most recent call last): | 1/38 [00:01<01:06, 1.79s/it]
File "train.py", line 202, in
main(args, configs)
File "train.py", line 87, in main
output = model(*(batch[2:13]))
File "/home/LtyDD/.conda/envs/fastspeech2/lib/python3.6/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
result = self.forward(*input, **kwargs)
File "/home/LtyDD/.conda/envs/fastspeech2/lib/python3.6/site-packages/torch/nn/parallel/data_parallel.py", line 165, in forward
return self.module(*inputs[0], **kwargs[0])
File "/home/LtyDD/.conda/envs/fastspeech2/lib/python3.6/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
result = self.forward(*input, **kwargs)
File "/home/LtyDD/Chinese-FastSpeech2/model/fastspeech2.py", line 65, in forward
output = self.encoder(texts, src_masks, char_vecs=char_vecs) ##############
File "/home/LtyDD/.conda/envs/fastspeech2/lib/python3.6/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
result = self.forward(*input, **kwargs)
File "/home/LtyDD/Chinese-FastSpeech2/transformer/Models.py", line 95, in forward
enc_output += char_vecs
RuntimeError: The size of tensor a (33) must match the size of tensor b (36) at non-singleton dimension 1
请问有没有人在训练的过程中遇到这个错误？

This error occurs when running synthesize_all.py

老师你好，请问可否分享一下您的TextGrid文件以及生成方法和expanded_embeds.pkl？

如题，作者有没有多speaker场景下的韵律预测方法，尝试加过speaker信息效果一般

你好，请问韵律模型是使用什么数据训练的？

您提供的预训练模型，我推理测试后，效果很惊艳！所以我一定要深入研究和复现！

该项目支持中英文混合训练或推理吗

您好，你的项目代码对我的帮助非常大。特别是优化后的语音效果比较让我满意。非常感谢！
我想在目前的基础上做进一步的开发，比如让合成后的语音数据能流式地输出到其他地方。应用场景是，比如输入一段新闻稿，合成后，流式地输出到其他设备。主要借鉴了飞浆的流式TTS功能。但设计过程中遇到了一此问题，希望能得到您的指点或建议。
主要问题：是否支持分批次推理完整文本的其中一部分，如果可以应该如何做到？
我主要参考了飞浆(paddle)其中一个demo的代码，主要从这里开始。
结合本项目的代码，我觉得应该从这里切入。
不知道这个方向有没有什么不托，但实现起来有些让我”挠头“ :-)

您好老师，请问biaobei.py里的代码是不是没有整理哇？expand_chars函数以上和下面一直到tokenizer之前的代码是不是都是数据处理重复啦，看晕辽qaq

提示这个文件找不到，能提供一下么？expanded_embeds.pkl
Prepare training ...
Traceback (most recent call last):
File "E:\python\Chinese-FastSpeech2\train.py", line 202, in
main(args, configs)
File "E:\python\Chinese-FastSpeech2\train.py", line 28, in main
dataset = Dataset(
File "E:\python\Chinese-FastSpeech2\dataset.py", line 31, in init
with open(self.expanded_char_path, 'rb') as f:
FileNotFoundError: [Errno 2] No such file or directory: './preprocessor/expanded_embeds.pkl'

您好，感谢您公开这么好的工作。我现在有2800条的语音，想训练一个prosody 模型。您觉得是 fintune prosody 模型还是从头训练效果会更好一些呢？ 谢谢您的回复

Thank you for sharing your work. I am truly impressed by your project and have developed a keen interest in understanding it more deeply. If it's convenient for you, I have a few questions that I'd like to ask.
I noticed that you've used BERT for extracting prosodic features in your project. I've conducted some experiments on my own, but the BERT models I found on HuggingFace didn't yield results as good or as natural as yours. I've tried the WWM version and the large models, but neither seemed to work very well. This has been a point of confusion for me, and I was hoping you could help clarify. Is the BERT model you used trained from yourself or taken from Google? Is it the WWM version and did you do some modification? Also, have you fine-tuned it on datasets other than Chinese Wikipedia? I would greatly appreciate your insights on these matters.

UP主有用这个项目训练多个模型没？我再想用之前训练的模型调用synthesize.py时就会报这个异常，对这个不熟悉的我真的难到了，万望UP主试一下看看能不能使用之前的模型，先行谢过。

有和成例子吗？

你好，在训练韵律模型的时候标注是依靠什么打的呢？ 011011这种

I plan to fine-tune my own dataset based on the AISHELL3 model, but my dataset only has 6 speakers, while AISHELL3 has 218. When loading the model, an error occurred due to the size mismatch. Additionally, Baker dataset only has one speaker, which also doesn't match with AISHELL3. I wonder how the author dealt with this issue?