Deep Speech v0.4.0 Released


(kdavis) #1

We just released Deep Speech 0.4.0!

We are looking forward to your feedback!


(Silenter) #2

Can’t wait to test
" example for Python streaming from mic with VAD" :smiley:


(Ken Yeung) #3

thanks it’s great. might i ask how long did it take to train the release model this time?


(kdavis) #4

27 hours on a machine with 8 Titain X Pascal GPU’s


(Yv) #5

there’s a new release 0.4.1 with a note "(The incorrect model was uploaded this will be fixed in 0.4.1) " in 0.4.0 - which model was incorrect (lm, pb, checkpoints)?


(kdavis) #6

The pb was incorrect.

Normally we run many numbered jobs to optimize the results. Embarrassingly, there was a typo in the job number we identified as optimal. 3206 was the optimal job but the model was pulled from job 3306.

Well, that’s life.


(Yv) #7

k, so all the other artifacts (pbmm, checkpoints etc.) were pulled from the incorrect job as well I presume


(kdavis) #8

checkpoints are the only thing pulled from the job, then pb, pbmm, rounded.pb, and rounded.pbmm are all derived.

So having checkpoints incorrect causes pb, pbmm, rounded.pb, and rounded.pbmm to also be incorrect.

But the correct pb, pbmm, rounded.pb, rounded.pbmm, and checkpoints are all already uploaded to 0.4.1


(Yv) #9

ok, thanks for clarification


(Ken Sinclair) #10

Just a single 8-GPU machine? Not the cluster?