wjruoxue

Use GPU to run FST decoding is not something new. The major challenge is how to run it efficiently when model cannot fit into one GPU. The reason Kaldi still uses CPU for decoding is that with a beefy CPU machine many decoders can run in parallel with just one copy of the model in memory. With enough CPUs, the RTF isn't so bad.

Depends on what the invokers wants. If the invoker only wants to log things to disk and read those logs later, surely he can do it himself. What XLog provides for the invoker under the hood is essentially performance: the invoker can log lots of things and does not need to worry about IO speed, disk space, and log safety. And of course it's cross platform. A detailed performance analysis is here (in Mandarin Chinese): http://mp.weixin.qq.com/s/cnhuEodJGIbdodh0IxNeXQ

It has a lot of overlaps with other network frameworks such as OkHttp, AFNetworking, cronet, etc., but I think it main contributions are these two components:

Xlog: in memory mmap log for fast logging and also supports encryption. Xlog log enables them to log everything fast and "safely" even in production build to quickly identify bugs.

STN: short and long TCP connection manager, which is optimized for various networks in China.

Plus other network utilities to make development easier.

Could be a win-win situation for both the Chinese and the local.

If this could bring the INTERNET to places like China (where there is only INTRANET)!

I once went to an IDEO talk and noticed that all words in the speaker's Powerpoint are capital letters. After the talk, I asked the guy why he made it like that and his answer was that since capital letters make it harder to read, you would force yourself to write concisely on the Powerpoint. He also mentioned that many people inside IDEO follow this rule.

Badly made Powerpoint is just useless while 6-page memo might be too time consuming and inconvenient. Capital letter approach could be a good compromise between those two.

Maybe you want to try Musubi, an experimental message app from Stanford's MobiSocial lab which really cares about user's privacy. It's available for Android and iOS