Skip to content

Latest commit

 

History

History
245 lines (163 loc) · 6.65 KB

INSTALL.md

File metadata and controls

245 lines (163 loc) · 6.65 KB

Note:

install pyaudio install pavucontrol

mkdir /home/lyncis/proj/voco/data/staging/audio_data/

install i3blocks link i3 blocks to config

change terminal name to mainterm

Starting point

Start here to install and configure Voco on a fresh installation of Ubuntu 17.10.

i3wm

I'm using i3wm which is a great option for a keyboard only window manager. if you decide to use it the tutorial is located here.

To install i3wm:

sudo apt install i3 i3status suckless-tools

After a restart there should be an i3 option at the login screen.

  • To open a termainal press Mod+ENTER where mod is the modifier key you chose when logging into i3 for the first time.
  • To run a program press Mod+D to access dmenu and then type the name of the program (e.g. firefox or nautilus).

A good tutorial on display management (e.g. a dual screen setup) is located here.

Emacs with Spacemacs

Emacs is a text editor / IDE (depending who you ask) that is well suited to voice operation since all commands are accessible via the keyboard. Spacemacs is an addon layer for Emacs that makes it prettier and easier to use.

To install EMACS in ubuntu or debian execute:

sudo apt install emacs

Then to install Spacemacs:

git clone /~https://github.com/syl20bnr/spacemacs ~/.emacs.d

When you start emacs the first time it will ask you to choose a kayboard style, choose "EVIL mode".

You can find more information about Spacemacs at http://spacemacs.org/

Installing and configuring Voco

Final directory structure

note: replace with tree command output

  • Kaldi [kaldi_root]
    • tools
    • src
  • Silvius
    • grammar (where the grammar parser lives)
    • stream
  • Voco [voco_root]
    • data [where all data related to your model lives]
      • audio_data [audio files that will be trained on]
      • audio_records [supporting files, in Kaldi format, describing the above audio files]
      • data [training data conforming to structure required by Kaldi]
      • staging [audio files and records still to be reviewed from live recording]
    • data_creation [module that creates your training set]
      • commands.csv [list of the commands you want to use]
      • create_recording_list.py [converts commands.csv to recording list]
      • record.py [does the actual recording]
      • create_dataset.sh [creates the voco\training\data directory]
    • training [Kaldi training recipe]
      • run.sh [runs the training]
      • exp\tri1_ali [final trained model]
    • main [decoder module]
      • record_decode.py [the script that runs the decoder]
    • parse_log
      • log [logfile you want to parse]
      • parse_log.py
      • parse_counter.txt [saves the linenumber of the last processed entry in log]

Setting environment variables

First we add variables defining the root directory for the various subcomponents. I'm assuming the project will located in a subdirectory called in your home direcotry.

echo 'export KALDI_ROOT=~/ASR/kaldi' >> ~/.bashrc
echo 'export VOCO_ROOT=~/ASR/voco' >> ~/.bashrc
echo 'export VOCO_DATA=~/ASR/voco/data' >> ~/.bashrc
. ~/.bashrc

echo 'export KALDI_ROOT=~/proj/kaldi' >> ~/.zshrc
echo 'export VOCO_ROOT=~/proj/voco' >> ~/.zshrc
echo 'export VOCO_DATA=~/proj/voco/data' >> ~/.zshrc
. ~/.zshrc

Compiling Kaldi

For a crash course on Kaldi check out: Kaldi for Dummies. But be warned, Kaldi is more of a research project than a finished user friendly program. Dont delve too deep unless you need to. Below is the process I followed.

Clone the Kaldi repository:

git clone /~https://github.com/kaldi-asr/kaldi.git kaldi --origin upstream

Set up the /tools directory:

cd $KALDI_ROOT/tools
sudo apt-get install libatlas3-base
sudo apt-get install zlib1g-dev automake autoconf libtool subversion
make

Set up the /src directory:

cd $KALDI_ROOT/src
./configure --shared

The config file complained about not finding ATLAS. I tried sudo apt-get install libatlas-base-dev but it didnt help so I installed openblas:

cd $KALDI_ROOT/tools
sudo apt install gfortran
extras/install_openblas.sh

Now compile Kaldi: This step takes a long time. The -j 2 command sets the number of CPU's to be used for the make command.

cd $KALDI_ROOT/src
./configure  --openblas-root=../tools/OpenBLAS/install
make depend -j 2
make -j 2

Setting up Voco

Clone the Voco repository:

git clone /~https://github.com/lyncisdev/voco.git

[todo] need to remove all the actual directories from GIT and add files to .gitignore

Create symlinks for the steps and utils directories in the WSJ recipe:

ln -s $KALDI_ROOT/egs/wsj/s5/steps $VOCO_ROOT/training/steps
ln -s $KALDI_ROOT/egs/wsj/s5/utils $VOCO_ROOT/training/utils

Create symlink: training/data --> data/data

ln -s $VOCO_ROOT/data/data $VOCO_ROOT/training/data

Create symlink: main/decode/model --> training/exp/tri1_ali

ln -s $VOCO_ROOT/training/exp/tri1_ali $VOCO_ROOT/main/decode/model

make output directory

mkdir $VOCO_ROOT/main/decode/output/scoring

Create symlink: main/decode/data --> data/staging

ln -s $VOCO_ROOT/data/staging $VOCO_ROOT/main/decode/data

[link parse_log/log to data/staging/log]?

You will need to get the VoxForge phone dictionary (which maps words to their phonetic representation) from the VoxForge github repository (/~https://github.com/VoxForge/develop/lexicon)

curl https://raw.githubusercontent.com/VoxForge/develop/master/lexicon/VoxForgeDict.txt > $VOCO_ROOT/data_creation/VoxForgeDict

Install SRILM

download srilm from:

and copy to []

rename file to srilm.tgz

run $kaldi_root/tools/install_srilm.sh

sudo apt-get install gawk

voco_data directory: should this just be taken from the voco_data

Silvius

Currently Silvius is packaged with Voco, the plan is to push these changes back to Silvius and then just use silvius directly.

Keynav

Keynav is the Keyboard emulation program that actually executes the keystrokes on your computer.

To install Keynav in Ubuntu or Debian execute:

sudo apt install keynav
sudo apt install xdotool

You can find more information on Keynav at: http://www.semicomplete.com/projects/keynav

Rofi

Task switcher that is started with the "switch" command. "Switch window" just presses "alt + tab"

To install rofi in Ubuntu or Debian execute:

sudo apt install rofi

You can find more information on Rofi at: /~https://github.com/DaveDavenport/rofi