your greed decode implement is wrong.

[Duum]

I think your transducer greed decode implement is wrong.
here is my implement of pytorch.

[HawkAaron]

Its based on the assumption that for each acoustic feature frame, there is at most one corresponding label, so we just need to move one step up, then turn right.

I implemented the greedy decode several weeks ago as your way, but the PER is worse in TIMIT.

The decoding algorithm is still under developing, so any comments are welcome.

[Duum]

but in my dataset my greedy decode implement CER is 15% lower than yours.

[HawkAaron]

Which dataset did you use?

I'll check that again in TIMIT.

[Duum]

In the paper of alex grave, in a transducer path ，only the label is null, the frame will step. when the label is not null ,the U will increase,But the T will stop to wait.

[Duum]

in my private dataset，only when use the second method ，the rnn transducer will compare with ctc.

[HawkAaron]

@Duum
Yes, when non-null is predicted, u forward one step and t stop. But as what I said, at any frame t, the corresponding label will not be more than 1, so if non-null is predicted at frame t, it must move to the next frame after the next prediction. What I said is based on the physical meaning of speech feature. But according to the original transducer model definition, at any time t, there could be more than one up transitions, which is meanless in speech recognition.

I'll check your implementation and reply to you asap.

[HawkAaron]

@Duum By the way, what you mean "the second method" ? Have you ever try beam search ?

[Duum]

the second method is my implement of greedy decode，I haven't use beam search for now .
a frame corresponding to a label may be just your assumption. If the frame is between two phones, it is possible to have two phones in one frame.

[HawkAaron]

@Duum Thanks for your comments, I'll check that.