rnnt-speech-recognition
rnnt-speech-recognition copied to clipboard
Perf on device
Hi,
Thanks for your work this is great! i found the same paper this morning, telling my self it would be something to implement it :) have you tried it on mobile device for real time streaming speech reco? i was asking my self if the model described in the paper wasn't to big, needing all theses perf optimizations?
cheers,
Kezak
I haven't been able to train a model with in the parameters specified in the paper (training time and resources are limited), I am currently training a much smaller model to see if I can get relatively good performance on mobile devices.
I am also looking at optimizing the training procedure and if you find any optimizations you can think of, please submit a PR request
hi Noah, i found somes ideas for training optimizations in this paper that may help : https://arxiv.org/pdf/1909.12415.pdf