CTC along with RNNs (LSTM/GRU) is currently state of the art in ASR enough to say it is the technique used currently by Google [http://googleresearch.blogspot.com.eg/2015/09/google-voice-search-faster-and-more.html] and Baidu [http://arxiv.org/abs/1512.02595v1]
The technique also won competitions in handwriting recognition and are very successful in OCR I am currently only a user : D and didn't delve enough in the theory but there are a number of Cpp implementations that should greatly simplify the job: 1) wrap-ctc released as a fast Cpu/Gpu CTC library from Baidu 4 days ago [https://github.com/baidu-research/warp-ctc] 2) RNNLIB, the reference CTC implementation by Alex Graves [http://sourceforge.net/p/rnnl/wiki/Home/] 3) CLSTM, an OCR library based on LSTM+CTC by Thomas Breuel [https://github.com/tmbdev/clstm] ________________________________________ From: [email protected] <[email protected]> on behalf of Wang Wei <[email protected]> Sent: Wednesday, January 13, 2016 12:41 PM To: [email protected] Cc: 钟鸣 Subject: Re: New features to be included in v0.3 Hi Mohamed Yes. Only GRU has been added in v0.2. I think Zhongming (cc'ed) is implementing LSTM. It would be interesting to provide an example for acoustic modeling using CTC. If you have experience of using CTC for acoustic modelling , it would be appreciated if you can help to add it in v0.3. regards, Wei On Wed, Jan 13, 2016 at 5:41 PM, Mohamed Yousef <[email protected] > wrote: > I would suggest implementing an LSTM layer and a CTC layer (LSTM was > planned for v0.2, but only GRU was added) > a CTC layer would greatly simplify acoustic modeling and many other > applications > ________________________________________ > From: [email protected] <[email protected]> on behalf of Wang Wei < > [email protected]> > Sent: Wednesday, January 13, 2016 8:51 AM > To: [email protected] > Subject: New features to be included in v0.3 > > Hi all, > > I have proposed some features for v0.3, which is scheduled to be released > in March 2016. > The major goals include usability and performance improvements. > > I have assigned some of the tasks to some developers/committers. > If you are interested in any of them, please join the development by > replying to that ticket. > > I will add new tickets continuously. > You are also welcomed to create new tickets for other features about > usability and performance improvements. > Let's get started! > > Thanks! > > regards, > Wei > SINGA-136 <https://issues.apache.org/jira/browse/SINGA-136> > > Support cuDNN v4 <https://issues.apache.org/jira/browse/SINGA-136> > Haibo Chen > <https://issues.apache.org/jira/secure/ViewProfile.jspa?name=seaokcs> > *Actions* > < > https://issues.apache.org/jira/rest/api/1.0/issues/12929715/ActionsAndOperations?atl_token=A5KQ-2QAV-T4JA-FDED%7Cdab6635ed244de6f2d2e77e2220bfc953beea22a%7Clin > > > SINGA-135 <https://issues.apache.org/jira/browse/SINGA-135> > > Improve hybrid partitioning and benchmark performance > <https://issues.apache.org/jira/browse/SINGA-135> > Sheng Wang > <https://issues.apache.org/jira/secure/ViewProfile.jspa?name=wangsh> > *Actions* > < > https://issues.apache.org/jira/rest/api/1.0/issues/12929708/ActionsAndOperations?atl_token=A5KQ-2QAV-T4JA-FDED%7Cdab6635ed244de6f2d2e77e2220bfc953beea22a%7Clin > > > SINGA-134 <https://issues.apache.org/jira/browse/SINGA-134> > > Extend SINGA to run over a GPU cluster > <https://issues.apache.org/jira/browse/SINGA-134> > *Unassigned* > *Actions* > < > https://issues.apache.org/jira/rest/api/1.0/issues/12929371/ActionsAndOperations?atl_token=A5KQ-2QAV-T4JA-FDED%7Cdab6635ed244de6f2d2e77e2220bfc953beea22a%7Clin > > > SINGA-133 <https://issues.apache.org/jira/browse/SINGA-133> > > Implement message passing using MPI > <https://issues.apache.org/jira/browse/SINGA-133> > *Unassigned* > *Actions* > < > https://issues.apache.org/jira/rest/api/1.0/issues/12929370/ActionsAndOperations?atl_token=A5KQ-2QAV-T4JA-FDED%7Cdab6635ed244de6f2d2e77e2220bfc953beea22a%7Clin > > > SINGA-132 <https://issues.apache.org/jira/browse/SINGA-132> > > Optimize training on a single node with GPUs > <https://issues.apache.org/jira/browse/SINGA-132> > Haibo Chen > <https://issues.apache.org/jira/secure/ViewProfile.jspa?name=seaokcs> > *Actions* > < > https://issues.apache.org/jira/rest/api/1.0/issues/12929368/ActionsAndOperations?atl_token=A5KQ-2QAV-T4JA-FDED%7Cdab6635ed244de6f2d2e77e2220bfc953beea22a%7Clin > > > SINGA-131 <https://issues.apache.org/jira/browse/SINGA-131> > > Implement and optimize hybrid training using both CPU and GPU > <https://issues.apache.org/jira/browse/SINGA-131> > *Unassigned* > *Actions* > < > https://issues.apache.org/jira/rest/api/1.0/issues/12929354/ActionsAndOperations?atl_token=A5KQ-2QAV-T4JA-FDED%7Cdab6635ed244de6f2d2e77e2220bfc953beea22a%7Clin > > > SINGA-130 <https://issues.apache.org/jira/browse/SINGA-130> > > Implement a layer subclass for data prefetching > <https://issues.apache.org/jira/browse/SINGA-130> > *Unassigned* > *Actions* > < > https://issues.apache.org/jira/rest/api/1.0/issues/12929190/ActionsAndOperations?atl_token=A5KQ-2QAV-T4JA-FDED%7Cdab6635ed244de6f2d2e77e2220bfc953beea22a%7Clin > > > SINGA-129 <https://issues.apache.org/jira/browse/SINGA-129> > > Add new SGD updaters <https://issues.apache.org/jira/browse/SINGA-129> > Jinyang Gao > <https://issues.apache.org/jira/secure/ViewProfile.jspa?name=jinyang> > *Actions* > < > https://issues.apache.org/jira/rest/api/1.0/issues/12929184/ActionsAndOperations?atl_token=A5KQ-2QAV-T4JA-FDED%7Cdab6635ed244de6f2d2e77e2220bfc953beea22a%7Clin > > > SINGA-128 <https://issues.apache.org/jira/browse/SINGA-128> > > Update compilation options and scripts for installing SINGA on MacOS > <https://issues.apache.org/jira/browse/SINGA-128> > Xie Zhongle > <https://issues.apache.org/jira/secure/ViewProfile.jspa?name=zhongle> > *Actions* > < > https://issues.apache.org/jira/rest/api/1.0/issues/12929182/ActionsAndOperations?atl_token=A5KQ-2QAV-T4JA-FDED%7Cdab6635ed244de6f2d2e77e2220bfc953beea22a%7Clin > > > SINGA-127 <https://issues.apache.org/jira/browse/SINGA-127> > > Clear optimization level of compilation for debugging > <https://issues.apache.org/jira/browse/SINGA-127> > Xie Zhongle > <https://issues.apache.org/jira/secure/ViewProfile.jspa?name=zhongle> > *Actions* > < > https://issues.apache.org/jira/rest/api/1.0/issues/12929180/ActionsAndOperations?atl_token=A5KQ-2QAV-T4JA-FDED%7Cdab6635ed244de6f2d2e77e2220bfc953beea22a%7Clin > > > SINGA-126 <https://issues.apache.org/jira/browse/SINGA-126> > > Improve Python Binding for interactive training > <https://issues.apache.org/jira/browse/SINGA-126> > *Unassigned* > *Actions* > < > https://issues.apache.org/jira/rest/api/1.0/issues/12928828/ActionsAndOperations?atl_token=A5KQ-2QAV-T4JA-FDED%7Cdab6635ed244de6f2d2e77e2220bfc953beea22a%7Clin > > > SINGA-125 <https://issues.apache.org/jira/browse/SINGA-125> > > Improve Python Helper - additional features > <https://issues.apache.org/jira/browse/SINGA-125> > *Unassigned* >
