JIJIN CHEN

` cum_prob` cannot be initialized to 0.

Dear sir, Should I add natural noise in not-wake-word folder ? I find my train result will be waked up by natural or speech.

可以先用sisnr预训练几步,初始化参数,然后在换成mixloss进行训练

Code has error in 'rnn_train.py' so that loss is nan. ``` rb = targets[:,:,:34] gb = targets[:,:,34:68] ``` but in 'denoise.cpp': ``` fwrite(g, sizeof(float), NB_BANDS, f3);//gain fwrite(r, sizeof(float), NB_BANDS, f3);//filtering...

There is a new question for 'loss nan'. The feature of pitch correlation could be 'nan'. The value 'error' could be zero in the file named 'celt_lpc.cpp', which make pitch...

我现在用的是mixloss，一个epoch，幅度损失在12左右，相位损失幅度相当大，180-40不等吧

我觉得aishell还挺干净的,我的SiSNR差不多13左右,但效果还是不太好.我再试试其他数据集吧.

> > 我觉得aishell还挺干净的,我的SiSNR差不多13左右,但效果还是不太好.我再试试其他数据集吧. > > 你好，请问您最后换了其他数据集测试了嘛？最后的效果怎么样？ 降噪最好还是用wsj数据集

这个问题已解决。我是将Tag以后的进行分类，wetext里的Char类型维持原有映射，其他类型组成一个组合，进行时间戳的映射。