How do I convert the npy form of Whisper's encoder into a voice?

[changhyunkwang]

Now I'm making an autoencoder model, and I'm making an autoencoder that uses Whisper to extract voice features and restore them again. First of all, I'm done extracting it in npy form, but I can't restore it back to its original voice, so how can I approach it?
I use Griffin-Lim argorisim