search

syhw / wer_are_we

Attempt at tracking states of the arts and recent results (bibliography) on speech recognition.
1.86k stars 226 forks source link

DS2 parameter correction to 100M #13

Closed samgd closed 6 years ago

samgd commented 6 years ago

With reference to the Deep Speech 2 paper, it states in Section 6.1.1 that:

All further English DS2 results are reported with this same 100 million parameter RNN model since it achieves the lowest generalization errors.

The table giving the WER results is below this line, and hence implies 100M, not 68 as stated in Table 10.

Admittedly it is ambiguous and annoyingly also fails to state the total number of parameters used for the DS1 model.

syhw commented 6 years ago

Thanks, you're right. :)