contextual_speech_correction

This is a ROS package which listens for raw text (produced from a speech recognizer) on the topic speech_raw and publishes the corrected text onto the topic speech_corrected. This performs a contextual, phoneme based correction, where the context is provided specifying a grammar (CFG) in the init_ros.py file.

Prerequisites

• Ubuntu 20.04 LTS

To check OS version: Press Super key ⟶ Enter 'About' in the search ⟶ Scroll down and see OS Name.

• ROS 1, noetic

To check ROS version: Enter ls /opt/ros in a terminal window. You should see the directory noetic.

NOTE: If you get this error ls: cannot access '/opt/ros': No such file or directory, this means you do not have ROS 1 installed. You can follow the ROS installation steps here.

• Python 3.8

To check your python version, enter python --version. You should see Python 3.8.x where x is any number. In case this produces Command 'python' not found error, you should try python3 --version.

If this also produces an error it means you do not have Python 3 installed. You should run

sudo apt update
sudo apt install python3.8

• Virtualenv

You will need to create a python virtual environment. To do this you will need the virtualenv utility. Check if installed by running virtualenv --version.

If not installed, do a simple pip install: pip install virtualenv

Installation

1. Sourcing setup.bash: source /opt/ros/noetic/setup.bash. Note you will need to do this everytime you open a new terminal window.
   For convenience you can add this line at the end of your ~/.bashrc file so that it is automatically executed.

2. Create a catkin workspace

mkdir -p ~/catkin_ws/src
cd ~/catkin_ws/
catkin_make
source devel/setup.bash

Note you will need to source devel/setup.bash everytime you open a terminal. So it is advisable to add it after the last line in ~/.bashrc

3. Clone the repository to fetch the contextual_speech_correction package. run git -C /src https://github.com/Aradhye2002/contextual-speech-correction.git

Or if you are facing authentication problems, you can download a zip file from the github page. Make sure to extract the package in the directory ~/catkin_ws/src

4. Create a python virtual environment:

virtualenv -p /usr/bin/python3.8 <your-env-name>

5. Change current environment to :

source <your-env-name>/bin/activate

5. Install the dependencies:

cd /home/<user-name>/catkin_ws/src/contextual_speech_correction
pip install -r requirements.txt

Documentation

• contextual_speech_correction implements a rosnode "speech_corrector" which serves to contextually correct raw text from a speech recognition system.
• The context is provided by specifying the CFG in the init_ros.py file. This is done by initializing the grammar-eng and grammar-hin variables.
• There are two modes: hindi and english. The mode is passed as a command line argument to init_ros.py.
• You also need to specify the depth of derivations upto which you will search in the grammar. NOTE: The user specified CFG is internally represented in CNF and hence is in general much larger than the entered CFG.
• A good rule of thumb for the depth is 30. Large values of depth also slow down the correction, so you have to balance between the two factors.
• The speech_correcter rosnode will listen for sentences of the type std_msgs/String on the channel speech_raw and output the corrected sentence onto the channel speech_corrected.

Usage

• Make sure you have sourced all the mentioned files, and are in the required virtual environment.

• Initialize ros master: run roscore

• Start listening on the speech_corrected channel for the corrected text: rostopic echo speech_corrected

• Start the rosnode speech_correcter: rosrun contextual_speech_correction init_ros.py <lang> Here the placeholder takes values "english" or "hindi" (without quotation marks).

• Write to speech_raw channel: rostopic pub std_msgs/String <raw_speech> Where <raw_speech> is any string (with quotation marks) to be corrected.