huyanxin / kaldi-aslp

ASLP Kaldi Code Trunk

Feature

• nnet enhancement
  • BatchNormalizaiton
  • Standard Lstm & BLstm, without projection
  • Latency control BLstm
  • CTC in two ways(Warp-CTC vs Eesen)
  • Skip training & decode
  • Graph network(eg multi-input, multi-output, add and splice)
  • Row Convolution
  • GRU
  • FSMN
• vad
• online
• parallel(MPI based multi-card speedup)
  • BSP
  • ASGD
  • EASGD
  • BMUF
• kws

Install

kaldi-aslp compile just like the standard kaldi recipe.

First, compile kaldi dependent tools if the network is not avaliable, You can download kaldi depending tools from here.

cd tools
make

Then, go to src dir, modify aslp.mk, change the compile switch as you like

# Optional make warp-ctc
USE_CTC = true
USE_WARP_CTC = false

# Optinal make mpi parallel(aslp-paralle aslp-parallelbin)
USE_MPI = true

# Optinal make online part, Online depends on crf++, add it's lib and include
USE_ONLINE = false

CRF_ROOT = /home/zhangbinbin/zhangbinbin/kaldi-aslp/src/crf
CRF_FLAGS = /home/zhangbinbin/zhangbinbin/kaldi-aslp/src/crf/libcrfpp.a

Last, just configure and compile.

cd src
./configure
make -j 24

How to Use?

To use it, just copy aslp_scripts to your project, add kaldi-aslp bin dir to your path.sh

export KALDI_ROOT=`pwd`/../../../
[ -f $KALDI_ROOT/tools/env.sh ] && . $KALDI_ROOT/tools/env.sh 
export PATH=$PWD/utils/:$KALDI_ROOT/src/bin:$KALDI_ROOT/tools/openfst/bin:$KALDI_ROOT/src/fstbin/:$KALDI_ROOT/src/gmmbin/:$KALDI_ROOT/src/featbin/:$KALDI_ROOT/src/lm/:$KALDI_ROOT/src/sgmmbin/:$KALDI_ROOT/src/sgmm2bin/:$KALDI_ROOT/src/fgmmbin/:$KALDI_ROOT/src/latbin/:$KALDI_ROOT/src/nnetbin:$KALDI_ROOT/src/nnet2bin/:$KALDI_ROOT/src/kwsbin:$KALDI_ROOT/src/online2bin/:$KALDI_ROOT/src/ivectorbin/:$KALDI_ROOT/src/lmbin/:$KALDI_ROOT/src/nnet3bin/:$PWD:$KALDI_ROOT/src/aslp-nnetbin/:$KALDI_ROOT/src/aslp-bin/:$KALDI_ROOT/src/aslp-vadbin/:$KALDI_ROOT/src/aslp-parallelbin:$KALDI_ROOT/src/aslp-onlinebin:$PATH
export LC_ALL=C

How to Develop & Contribution

Anybody, any contribution is highly appriciated.

What's New

• 2017-03-20 SOD parallel surpport, adagrad, adadelta, rmpsprop, adam optimizer
• 2016-10-06 MPI based Parallel support
• 2016-07-10 More VAD support, ROC, boundary accuracy, and Praat format tools
• 2016-06-06 Online support
• 2016-04-30 VAD support(variable vad models, energy based, gmm based, dnn and lstm based)
• 2016-04-15 CTC support. Both Eesen and Warp-CTC(Baidu AI-Lab) is intergrated.
• 2016-03-28 Skip training & decode support.
• 2016-03-20 Phone CD-Phone prepare and training support.
• 2016-03-20 Standard Lstm BLstm support.
• 2016-03-07 Graph nnet support.
• 2016-02-20 MPI based 2-gpu speed up.
• 2016-02-20 Batch Normalizaiton support.
• 2016-01-15 Init trunk(DNN, CNN prepare and training script)