you can use you own audio set an train your own project
In general,your total dataset should be separate as 60%(traning ), 20%(validationset),20% (testing).
what i show you just the right works code
all code you could just change a little bit ,then you could train your own dataset.
if you still want to talk with me , just cue me ,i would like to share my idea .
i just dont have time to learn use GitHub , but i would like to learn if i can .
I am so Sorry ,My English was so poor or i would like to talk more.Thanks for your view！
My project was base on :
https://github.com/llSourcell/tensorflow_speech_recognition_demo
If you got questions, Feel free to contact me ,my FaceBook was :
https://www.facebook.com/profile.php?id=100005220045328