Test - WER vs CER vs loss


(Deepspeech04) #1

I’ve trained my own model and need to know what my results describes in term of % WER, CER and loss. do i need to train with more epochs ?

Preprocessing done
Computing acoustic model predictions…
96% (171 of 177) |################################################################################################# | Elapsed Time: 0:00:13 ETA: 0:00:00^
100% (177 of 177) |#####################################################################################################| Elapsed Time: 0:00:14 Time: 0:00:14
Decoding predictions…
81% (145 of 177) |################################################################################## | Elapsed Time: 0:03:59 ETA: 0:00:46^
100% (177 of 177) |#####################################################################################################| Elapsed Time: 0:04:54 Time: 0:04:54
Test - WER: 0.213015, CER: 3.542373, loss: 25.164192


(kdavis) #2

What’s your end use case?


(Deepspeech04) #3

I am just doing research and trying to find out that how much % accuracy/error rate i’ve by using deepspeech for ASR.
it will be very helpful if you can elaborate these results
Test - WER: 0.213015, CER: 3.542373, loss: 25.164192
mean what WER, CER and loss represents in the above results and what does it mean. can we say it WER: 21.30%, CER: 354.23%, loss: -----


(kdavis) #4

I think it’d be profitable to do some background reading, for example maybe the Wikipedia article on WER would be a good start. Then maybe you could follow on reading some of the articles referenced there.


(Deepspeech04) #5

I already go-through the details of WER and CER. I just want to know how these results can be represented in percentage(%)
Test - WER: 0.213015, CER: 3.542373, loss: 25.164192
can we say it WER: 21.30%, CER: 354.23%, loss: ??


(Reuben Morais) #6

It depends on what version of the code you’re running, as we recently changed the behavior of the CER calculation to be less surprising. If you’re on a the latest master, WER and CER can be interpreted as error rates over the entire corpus (for example, CER = sum of all edit distances / sum of all lengths). The loss is not really interpretable outside of the training context.


(Deepspeech04) #7

i am using this version
DeepSpeech: v0.4.0-alpha.0-82-g6177da9
So according to this version how can i elaborate my results in percentages?
After training more epochs my results are:
Before:
Test - WER: 0.213015, CER: 3.542373, loss: 25.164192
After:
Test - WER: 0.097534, CER: 1.385030, loss: 13.236552