Adding `safetensors` variant of this model (#1)

end-2-end reborn model for librispeech unsupervised phoneme recognition (iter5-stage1)

initial commit