/kaldi

Documentation and resources for MLA's implementation of Kaldi speech-to-text software

Kaldi Speech-to-Text at WGBH's Media Library and Archives

This repo contains documentation and resources for MLA's implementation of Kaldi speech-to-text software.

Getting Started

  • Get MLA Amazon Web Services Console Login Credentials

  • Login URL: https://wgbh-mla.signin.aws.amazon.com/console

  • Click on EC2 under AWS Services

  • Go to Launch Instance

  • Select Amazon Linux

  • Select t2.xlarge

  • Configure Instance Details

  • Shutdown Behavior: Stop

  • Enable Termination Protection

  • Click Next: Add Storage

  • For Root Volume Type

  • Size: 64

  • Volume Type: General Purpose

  • Click Next: Add Tags

  • Kaldi

  • Step 6: Configure Security Group

  • SSH/Protocol/Port Range 22/ Anywhere

  • Launch!

  • Create a new key pair: kaldi-aws-ec2-user

  • Save pem file

  • Launch Instance

  • chmod key file 600

Installation

  • ssh in to instance

  • sudo yum update

  • install git: sudo yum install git-all

  • git clone ...kaldi.git kaldi --origin upstream

  • cd in to kaldi/tools directory

  • extras/check_dependencies.sh

  • sudo yum install zlib-devel gcc automake autoconf patch libtool

  • sudo yum install atlas.x86_64

  • sudo yum install gcc-c++

  • run 'make -j 4'

  • snapshot image

  • cd kaldi/src

  • run ./configure --shared

  • run make depend -j 4

  • run make -j 4

  • PUA Installation (https://github.com/popuparchive/american-archive-kaldi)

  • cd $HOME

  • git clone https://github.com/popuparchive/american-archive-kaldi

  • Download the latest PUA Models, final result should be (re-)named exp

  • curl -O 'https://cytranet.dl.sourceforge.net/project/popuparchive-kaldi/exp2.tar.gz'

  • NOTE: THIS WAS PROBLEMATIC, REQUIRING MANY HOURS AND "CONTINUE" SESSIONS LIKE THIS:

  • Unpack the fully-formed exp2.tar.gz to become directory 'exp'

    • tar -xzf exp2.tar.gz exp

  • Move the PUA Models into the american-archive-kaldi/sample_experiment

    • mv exp american-archive-kaldi/sample_experiment/

  • Create symlinks between structures in kaldi and american-archive-kaldi

    • ln -s kaldi/egs/wsj/s5/steps american-archive-kaldi/sample_experiment/
    • ln -s kaldi/egs/wsj/s5/utils american-archive-kaldi/sample_experiment/
    • ln -s kaldi/egs/wsj/s5/steps american-archive-kaldi/sample_experiment/exp/
    • ln -s kaldi/egs/wsj/s5/utils american-archive-kaldi/sample_experiment/exp/

Install system dependencies for perl - sudo yum install perl-core.x86_64 - sudo yum install perl-JSON.noarch - sudo yum install perl-CPAN.noarch # optional unless the following statement fails - cpan JSON # initializes CPAN and does confirm/install JSON - perl -MCPAN -e shell - install File::Slurp::Tiny - install Data::Dump - exit

  • Install SOX in root directory

  • Run ./configure in sox-[version] directory

  • Run make -s

  • Run sudo make install

  • Run make installcheck

  • Install FFMPEG in root directory

  • Sudo yum install yasm

  • Run ./configure in ffmpeg directory

  • Run make

  • Run sudo make install

  • Ensure CMUSeg is installed

  • cd american-archive-kaldi/sample_experiment/

  • ./install-cmuseg.sh

  • Install IRSTLM

  • cd kaldi/tools/extras

  • run install ./install_irstlm.sh

  • vi .bash_profile

  • Add command the following command to .bash_profile: . ./kaldi/tools/extras/env.sh

  • Install sclite

  • cd kaldi/tools/sctk-2.4.10

  • configure CFLAGS variable in makefiles to compile sclite for a 64-bit environment, adding -m64 to:

    • sctk-2.4.10/src/asclite/core/makefile: CFLAGS = -g -Os -m64
    • sctk-2.4.10/src/asclite/test/makefile: CFLAGS = -Os -Wall -Wconversion -m64
    • sctk-2.4.10/src/sclite/makefile: CFLAGS = -Os -m64
    • NOTE: the README cites src/rfilter1/makefile as well, but it contained no CFLAGS variable

  • Run make config

  • Run make all

  • Run make check

  • Run make install

  • Run make doc

  • Deal with python errors thrown when using american-archive-kaldi/run_kaldi.py

  • Upgrade pip and install ftfy

    • NOTE: the following assumes use of python v2
    • sudo pip install --upgrade pip
    • sudo ln -s /usr/local/bin/pip /usr/bin/pip
    • sudo pip install ftfy==4.4.1
    • sudo pip install PY3
    • sudo pip install -U six

  • More idiocy about inadequate permissions:

  • list inventory of now-executable files [ec2-user@ip-172-30-0-129 sample_experiment]$ find . -type f -name '*sh' -executable ./install-cmuseg.sh ./set-kaldi-path.sh ./run.sh

  • list non-executable files ending with '.sh' and make them executable [ec2-user@ip-172-30-0-129 sample_experiment]$ find . -type f -name '*sh' -not -executable ./path.sh ./tools/CMUseg_0.5/src/lib/sphere/src/scripts/install.sh ./tools/CMUseg_0.5/src/lib/sphere/src/scripts/gen_make.sh ./tools/CMUseg_0.5/src/lib/sphere/src/scripts/h_header.sh ./tools/CMUseg_0.5/src/lib/sphere/src/scripts/raw2nist.sh ./tools/CMUseg_0.5/src/lib/sphere/src/scripts/gen_setv.sh ./tools/CMUseg_0.5/src/lib/sphere/src/scripts/gen_lord.sh ./tools/CMUseg_0.5/src/lib/sphere/bin/h_header.sh ./tools/CMUseg_0.5/src/lib/sphere/bin/raw2nist.sh ./tools/CMUseg_0.5/src/lib/sphere/bin/gen_setv.sh ./tools/CMUseg_0.5/src/lib/sphere/bin/gen_lord.sh ./add_grammar.sh ./prep_lang_local.sh ./scripts/create_big_lm.sh ./scripts/qq.sh ./scripts/create_lm.sh [ec2-user@ip-172-30-0-129 sample_experiment]$ find . -type f -name '*sh' -not -executable -exec chmod +x {} ;

find /home/ec2-user/american-archive-kaldi/sample_experiment/tools/CMUseg_0.5/bin/linux -type f -exec chmod +x {} ; chmod +x american-archive-kaldi/sample_experiment/scripts/*

  • Begin idiocy involving perl -Use CPAN to install perl4 core lib for 'getopts.pl' reference in CMUseg.pl script -perl -MCPAN -e shell cpan > install Perl4::CoreLibs

  • Steps after following PUA install guide • In american-archive-kaldi/sample_experiment/run.sh, set the following option to reduce the number of simultaneous jobs to 1: nj=1 • In kaldi/egs/wsj/s5/utils/run.pl, set the following option: $max_jobs_run = 10; • In kaldi/egs/wsj/s5/steps/decode_fmllr.sh and kaldi/egs/wsj/s5/steps/tandem/decode_fmllr.sh, set the following options in the section where they appear together: nj=1 max_active=1000 • In decode.sh, set the following option: nj=2 in these locations:

  • kaldi/egs/wsj/s5/steps/nnet/decode.sh

  • kaldi/egs/wsj/s5/steps/online/decode.sh

  • kaldi/egs/wsj/s5/steps/online/nnet2/decode.sh

  • kaldi/egs/wsj/s5/steps/online/nnet3/decode.sh

  • kaldi/egs/wsj/s5/steps/decode.sh

  • kaldi/egs/wsj/s5/steps/tandem/decode.sh

  • kaldi/egs/wsj/s5/steps/nnet3/decode.sh

  • kaldi/egs/wsj/s5/steps/nnet2/decode.sh

  • kaldi/egs/aspire/s5/local/multi_condition/decode.sh

Usage