SeanNaren / deepspeech.torch

Speech Recognition using DeepSpeech2 network and the CTC activation function.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

Questions about training AN4 dataset

Yizhen-Z opened this issue · comments

I followed the installation guide and tried to run Train.lua on AN4 dataset.
But the results I got was quite strange, the Average Loss turned out to be a huge number, while WER and CER were also larger than 100. (In my understanding, it should be a percentage number which is less than 100).
screenshot

I didn't make any change on code. So I am really confused about it...

Another question is that, from the DS2 paper CER refers to 'Characer Error Rate' for the Mandarin system, so what is CER here refers to?

Thanks for any advise!

Could you let me know what OS/GPU you are using?

The OS is Ubuntu 16.04.1.
GPU is NVIDIA GTX 1080.

Thanks for reply!

Yeah I had the same issue as you, you need to install warp-ctc a bit differently on pascal cards, check here and try installing this branch instead, and restarting training!

Yes, it works now. I finally got WER = 18, and CER = 3.88 on AN4 dataset.
Thank you!
loss

Great stuff!