Training with small sample set


(Philofellow) #1

Is it possible to train deepspeech with very small sample set, e.g., 30 audio clips with each in 5 to 10 seconds? I tried with different parameters including n_hidden, learning_rate, etc. and the testing on the same sample set gives very bad results, e.g., WER 0.98.

If I want to train and test on the same small sample set, what parameter should I use and how many epochs are needed for it to converge to a good result?

thanks!


(Philofellow) #2

Update: I use n_hidden = 128 and learning_rate = 0.001 and was able to get a good model in 30 minutes with hundreds of epochs.