[dataset] new io for code reuse for many speech tasks

[dataset] new io for code reuse for many speech tasks #3659