Why the output of the CTC decoder is not seperated by space?

PaddlePaddle / PaddleSpeech

Easy-to-use Speech Toolkit including Self-Supervised Learning model, SOTA/Streaming ASR with punctuation, Streaming TTS with text frontend, Speaker Verification System, End-to-End Speech Translation and Keyword Spotting. Won NAACL2022 Best Demo Award.

https://paddlespeech.readthedocs.io

Apache License 2.0

11.23k stars 1.86k forks source link

Why the output of the CTC decoder is not seperated by space? #244

Closed dreamibor closed 6 years ago

dreamibor commented 6 years ago

Hi,

I used logits from another model (Wav2Letter) to test the CTC decoder, however, I can't get correct results from the decoder, for example, I got :"sometimemesttheeultbotthfehinatetwhorkhasreallygod" from the decoder while the greedy decoder gave me "sometimd es the ou but of fan at work is really god". I am wondering why the CTC decoder will concatenate all the characters?

Thanks,

dreamibor commented 6 years ago

Haha, it turns out I have to put a softmax layer between the logits and the CTC decoder to normalise the input.