about WFST decoder for AISHELL-1

[Alex-Songs]

Hello, when I use WFST for AISHELL decoding, the predicted label is one less than the real label, what is the reason for this?
such as:
predictions: tf.Tensor([[2994 717 206 1404 1487 1741 3702 77 1392 84 524]
labels: [[2995 718 207 1405 1488 1742 3703 78 1393 85 525]]

But the beam search is correct.

[liubin3702]

你好，我想问问他这个解码器的输入分数是怎么来的 我用端到端声学模型怎么链接到这里面使用LG.FST解码呢