Kaldi Customization

This is the main repository of the IT-Project "Missing Title". The report with additional information is stored in the documentation repository.

Quick start & initial setup guide

Requirements

• Docker
• Docker Compose
• Git to download this repository
• Python3 for the initialization
• Prerequisites for the python package mysqlclient also for the initialization

Start the compose

• Open a shell
• Clone this repository to your local system: git clone https://git.informatik.fh-nuernberg.de/kaldi/kaldi-customization.git and switch into the repository folder (first time only)
• Use the env.cmd or env.sh script in your shell to setup the environment variables for docker-compose
• Build or import missing docker images (first time only)
  • kaldi-base: See kaldi/base/README.md. Note: Make sure the name of the image matches the name in the corresponding Dockerfile.
  • Create all other images by starting the docker-compose or by executing docker-compose build
• Start the customization service:
  • Load the compose with docker-compose up and have a cup of tea or coffee
  • Wait until the service is online (website is reachable: localhost:8080)
  • To scale the amount of workers, use the --scale parameter for docker-compose up, available workers are text-preparation-worker, data-preparation-worker, kaldi-worker and decode-worker
• Use the initialization script initialization/init.py (first time only):
  • A few modules are required to execute the script: For example use pip and pipenv:
    • Open another shell
    • pipenv install and pipenv shell to activate the pipenv shell
    • pip install -r ./initialization/requirements.txt to install the requirements
    • Note: If an error occurs, verify that the prerequisites for the python package mysqlclient are installed
  • Execute python ./initialization/init.py to prepare the database and upload default model data

The customization service is now available

• Web Interface: localhost:8080
• Web API: localhost:8080/api

Stop the service

• Make sure that there are no running jobs like a training
• Use docker-compose stop in the repository folder or press Ctrl + c in the shell where you startet the compose
• All data (database, files) are stored persistantly on the local disk
• Use docker-compose down to shut down the service and delete the database

Structure of the repository

/docker-compose.yml

This file defines the service. It is used by docker to build and run the images/containers.

/api

Definition of the public API. See api/README.md for further information.

/config

Contains some global settings for the docker-compose.

/dfs

Persistent storage for database (/dfs/mariadb) and file serivce (/dfs/data).
Do not touch manually!
Use a SQL explorer (e.g. MySQL Workbench) and the MinIO web client at localhost:9001 instead.

/initialization

As the name indicates: Preparation for the first usage. See initial setup guide.
Contains also the pretrained acoustic models.

/kaldi

Our docker image with a kaldi installation. Use the base image and see the README there.

/server

The server components to run the kaldi customization web service.

/server/api

This is the API backend. It provides access to the features of the kaldi customization web service and handles authentication.
See the README.

/server/web

This is the web frontend for users. It offers a user interface to train and test user defined ASR.

/shared

Scripts and resources which are used by several components.

/worker

The worker directory contains the workers used in the backend to process the user requests via the API.
See the directories for further information about the workers:

• text-preparation-worker: Extract text from uploaded resource files.
• data-preparation-worker: Prepares the training process.
• kaldi-worker: This is the general kaldi-worker to process ASR testing.
• decode-worker: Decodes audio to text.

Further Docker Images

MariaDB Server

A SQL Server for the persistent data.

Redis Server

An in memory Redis Server for the task queue.

API Functions

See localhost:8080/api/v1/ui.