Wer of small dataset

[JRMeyer]

>>> anupbera
[January 18, 2018, 3:52pm]

Anyone tell the WER of switchboard when train only on switchboard
dataset. Also WER of TED when training on TED dataset. Actually I wanted
to know how the Deepspeech perform when train on small dataset (around
200hrs).

[This is an archived TTS discussion thread from discourse.mozilla.org/t/wer-of-small-dataset]

[JRMeyer]

>>> lissyx
[January 19, 2018, 11:47am]

I tend to remember that on too small datasets you could not expect that
good results: WER would not achieve good level without overfitting
enough that the network is barely usable for anything more general. So,
maybe it can work in your case, what are you trying to achieve exactly?

[Archived Post]

[JRMeyer]

>>> anupbera
[January 20, 2018, 7:09am]

I have small amount (around 200hrs) of homogeneous dataset, telephonic
conversion. I am curious which algorithm should I use, whether I should
go to deep learning or HMM based algo. Deep learning itself has CNN/RNN
based algo. How much accuracy I can achieve if my test dataset also
similar to train dataset of telephonic conversation.

[Archived Post]