An Arabic dialect identification "post processor" after full result of the kaldi-gstreamer-server

This project is a part of QMDIS "QCRI-MIT Advanced Dialect Identification" system. The purpose of the system is to identify between five Arabic dialects, namely Egyptian Arabic, Levantine Arabic, Gulf Arabic, North African Arabic and Modern Standard Arabic (or MSA).

This project is just the post processor script that is called after the full transcript result of the ASR system. You need to install Kaldi ASR and the its Gstreamer server first before you can use it.

As QMDIS saves audio and json files within each session, It is usful to store the local time as well. To adjust the timezone according to the wanted place, we can do that through one of the following commands,

tzselect

which opens a gui in terminal, or

sudo cp /usr/share/zoneinfo/Europe/London /etc/localtime

NB!: Please don't use anaconda to install the ASR becasue bindings for gobject-introspection libraries will not be installed in its site-packages.

NB!: If you want to use virtual environment it is recommended to create it with --system-site-packages flag so that it can use Python bindings

NB!: The server doesn't work quite correctly with ws4py 0.3.5 because of a bug that has been reported here: Lawouach/WebSocket-for-Python#152.

• Python 2.7
• In addition, you need Python 2.x bindings for gobject-introspection libraries, provided by the python-gi package on Debian and Ubuntu.
• Tornado 4.5.2
• ws4py (0.3.0 .. 0.3.2)
• YAML
• JSON

sudo apt-get update
sudo apt install virtualenv
sudo apt install python-gobject
sudo apt install python-dbus

Use sudo -H only when you installing in the global environment. Check requirements.txt in repo for libraries versions.

virtualenv --system-site-packages <your_env_name> -p <path/to/python2.7>
source <path/to/your/env>/bin/activate

[sudo -H] pip install tornado==4.5.2
[sudo -H] pip install ws4py==0.3.2
[sudo -H] pip install pyyaml
[sudo -H] pip install numpy # numpy dropped python2 try [sudo apt-get install python-numpy]
											# scipy dropped python2 try [sudo apt-get install python-scipy]
[sudo -H] pip install tensorflow
[sudo -H] pip install soundfile
[sudo -H] pip install librosa

Clone the Kaldi as follows:

git clone https://github.com/kaldi-asr/kaldi.git

Change directory to

cd kaldi/tools/

and then run the following line to check the requirments to install Kaldi tools are all available

extras/check_dependencies.sh

.. and install them if needed

run make; IT WILL TAKE TIME

make [-j n]  # please change n to the number of cores you want to use

After building kaldi tools, change directory to kaldi/src/ and do the same thing. Note that kaldi tools need to be built first

cd kaldi/src/
./configure --shared
make clean [-j n] 
make depend [-j n] 
make [-j n]  # please change n to the number of cores you want to use

The DNN-based online decoder requires a newer GStreamer plugin that is not in the Kaldi codebase and has to be compiled seperately. It's available at https://github.com/alumae/gst-kaldi-nnet2-online.

Clone it and then change directory to the src directory under the cloned directory

git clone https://github.com/alumae/gst-kaldi-nnet2-online.git
cd gst-kaldi-nnet2-online/src

Install the following

sudo apt-get install gstreamer1.0-plugins-bad  gstreamer1.0-plugins-base gstreamer1.0-plugins-good  gstreamer1.0-pulseaudio  gstreamer1.0-plugins-ugly  gstreamer1.0-tools libgstreamer1.0-dev
sudo apt-get install libjansson-dev

Edit the Makefile in the directory and change the variable KALDI_ROOT?= poinring to the kaldi home directory as follows

vim Makefile
KALDI_ROOT?=/path/to/kaldi/home/directory

and then save and run

make depend
KALDI_ROOT=/home/disooqi/kaldi make # IT WILL TAKE TIME 

Download the model as

wget -O /tmp/model.tar.gz  crowdsource.cloudapp.net/models/Arabic/20180304/nnet3sac.tar.gz
wget -O /tmp/model.tar.gz  https://qcristore.blob.core.windows.net/public/asrlive/models/arabic/nnet3sac.tar.gz

and then untar it to /opt/model

sudo mkdir -m 777 /opt/model
tar xzvf /tmp/model.tar.gz -C /opt/model

create the following dir

sudo mkdir -p -m 777 /var/spool/asr/nnet3sac

0. Before cloning this repo please make sure Git LFS is insalled, use the following command to install it

curl -s https://packagecloud.io/install/repositories/github/git-lfs/script.deb.sh | sudo bash
sudo apt-get install git-lfs

1. clone this repo

git clone https://github.com/disooqi/qmdis-post-processor-full.git

2. make sure that dialectid_post_processor.py module and its parent directories have execution permissions
3. edit /opt/model/model.yaml and assign the path to dialectid_post_processor.py to the variable full-post-processor: as follows and append it to the file:

full-post-processor: /the/path/to/dialectid_post_processor.py
or 
post-processor: /the/path/to/dialectid_post_processor.py  #depending on yaml  

clone

git clone https://github.com/alumae/kaldi-gstreamer-server.git

change directory to the directory you just cloned, and the run server as follows:

python kaldigstserver/master_server.py --port=8888 [--certfile=] [--keyfile=]
sudo python kaldigstserver/master_server.py --port=8888 --certfile=/home/disooqi/qcri_certificate.pem --keyfile=/home/disooqi/qcri_key.pem

export GST_PLUGIN_PATH=~/gst-kaldi-nnet2-online/src #change path to the respective path on your system
cd kaldi-gstreamer-server
python kaldigstserver/worker.py -u ws://localhost:8888/worker/ws/speech -c /opt/model/model.yaml

1. pass server parameters
2. run worker with wss:// instead of ws://
3. run client with wss:// instead of ws://

sudo python manage.py runserver --insecure 0.0.0.0:80

git clone https://github.com/Kaljurand/dictate.js.git