AI Enthusiast

@ujscjj Sorry, I haven't tried keras. You know, Tensorflow can be backend in Keras but Tensorflow must be used when many underlying operations are involved.

both versions are okay

The actual structure of the dataset is as what you say. "run.sh" is in my programs but not uploaded. This file is including all processes.

@shreyasose Hi, I'm sorry for sharing the dataset because the dataset WSJ0 needs paid license.

@soliloquy1983 Yeah, randomly.

@ujscjj 是的。之所以这么做是因为如果按照帧进行训练输出，在测试时因为没有先验信息就不知道一个T-F bin属于哪一个说话人。与此同时，实验证明句子级别的置换不变训练)的方法相对于多类回归的模型也更加容易收敛，在没有见过的说话人甚至语言上也有很好的泛化性能。

@Yangjie55 最近我正在看并实现DANet。TIMIT数据集可以用这个代码的，简单点的话就是把TIMIT文件的组织形式改成WSJ0的文件组织形式，不然就修改代码

@nagasaibharath Hi, you just need to save the mixed audio to folders in the same format as WSJ0's.

This model is speaker-independent and language-independent. @CLCLpis

@ujscjj The code has been sent to you.