经过几轮测试，相较其他几个现在火热的模型，效果已经非常棒了！我的项目已经从f5-tts迁移过来了。
但是同一个音频多次推理测试还是经常出现生成的音频超长现象（其它推理模型也一样）：
比如也就不到10来个字，参考音频语速也正常。但是生成的音频有时会超过10几甚至30秒+的音频。 有时是莫名多了一大段的静音空白，有时候是每个单词之间超长的间隔。
这个问题不是必现，但是10来次推理基本必现。在其他推理库中也经常出现，也不知道是不是推理卡住了还是怎样。
有没有好的建议？我要用于批量场景。速度慢点没关系，质量第一。
感谢贡献，期待建议，项目急需
我们infer.py 增加了一个 remove_long_silence 函数，临时解决超长静音问题，后面还是会通过优化模型提升稳定性从根本上解决这个问题。
我一般是几百次偶然出现一次超长静音区间问题。我检测到后，自动重新推理，大概率还是一样会出现长静音，就是静音的长度变化一下。这时候，大概率推出来的音频也会比较奇怪。于是我基本放弃了自动重试，而是去检测到这种wav数据，发现这类情况后，重新采用去掉标点符号的短句去推理，然后在短句之间插入延迟，最后合成一段话。 这个虽然能部分解决这类问题，但是去掉标点，分短句去推理的话，有可能导致这段话合成的不够自然，没有整体推理来的自然。
測試下來發現有些樣本固定會出現這些現象。我測試的樣本是一個15秒的遊戲語音。 更換成其他來源的語音的30秒片段後即正常，且出現此現象時，也會因為 token 中包含許多 52 而在產出期間特別久。
infer.py 更新了，应该不会出现这种结果了，大家可以试下。。
@index-tts self.stop_mel_token 未定义，是否也是 8193 ？
是的
LOOK AT THIS PICTURE. NOW, LET'S READ, SICK.
tensor([[ 52, 52, 52, 7809, 4879, 421, 5715, 3147, 3628, 4711, 3808, 2128,
7068, 101, 1744, 4446, 821, 1319, 6021, 1003, 8022, 4461, 5358, 3695,
139, 2927, 7777, 7041, 988, 4891, 7524, 1151, 5654, 7207, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
6478, 2854, 6338, 5541, 3568, 3332, 469, 4242, 2374, 2527, 4530, 6490,
4365, 4856, 224, 2818, 8134, 10, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 7342, 5183, 5343, 1011, 3358,
1592, 5930, 1784, 971, 7733, 5018, 1773, 1337, 7603, 8069, 1234, 4489,
7557, 1287, 2017, 4435, 4135, 1257, 7232, 1019, 1258, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52, 52,
52, 52, 52, 52, 52, 52, 52, 52, 52, 52]],
device='cuda:0')
codes shape: torch.Size([1, 598])有些句子替换 52后是可以 静音的，但是如上推理这种情况下，就是丢失句子了。 似乎是因为模型没有推理完成，
这个bug修复了，大家可以拉下来试试。
这个bug修复了，大家可以拉下来试试。
新代码测试了几轮，确实没发生了，继续观察。 感谢及时处理，也感谢各位建议。
又出现了，试了几次都必现，素材如下：
text="Like I said, on your body, for me",
@index-tts 生产环境急等修复。麻烦帮忙看一下
这类似于LLM重复回复的问题吧，推理时把gpt2的token采样随机性提高应该就行了吧
又出现了，试了几次都必现，素材如下： text="Like I said, on your body, for me", clip_8.zip
@index-tts 生产环境急等修复。麻烦帮忙看一下
你好 请问解决这个问题了吗