Incorrect maintaining of states for words?

parlance / ctcdecode

PyTorch CTC Decoder bindings

MIT License

829 stars 245 forks source link

Incorrect maintaining of states for words? #33

Closed SeanNaren closed 7 years ago

SeanNaren commented 7 years ago

One of our colleagues got in touch with @willfrey and he mentioned an issue with the pytorch-ctc implementation:

He pointed out that the scorer in pytorch-ctc wasn't maintaining state between words properly, rendering it a "spell checker" (only scoring unigrams basically)

I'll do some investigation into this claim and report back!

ryanleary commented 7 years ago

This is definitively true. Fix incoming.