To build a short-vocabulary 1 hour Hindi Speech Corpus which can be used for Automatic Speech Recognition, and further perform acoustic and phonemic analysis on the dataset.
These instructions describe the prerequisites and steps to get started with the project.
To setup an Anaconda environment in-built with all the prerequisite packages used to retrieve the audio files from Youtube, do the following:
-
Download Conda, and install the same using
> bash install conda_install.sh -
Create a conda environment from the included
environment.ymlfile using the following command:$ conda env create -f environment.yml -
Activate the environment
$ conda activate hscorpus
To retrieve audio files from Youtube, write the URL's of the required files in the src/links.txt alongwith a Reference ID. Now, run the script downloader.py using the command: > python downloader.py. This will create an audios directory in the current working directory containing the retrieved audio files with their reference IDs. To change the defaults for the downloader.py script, run > python downloader.py --help to receive a list of arguments accessible to the user.
The short-vocabulary Hindi speech corpus is stored in the corpus directory.
The dataset is created from the audio files in audios directory, by breaking them down into sentences using the Praat software and Audacity software. All of the files retrieved by this procedure are saved in the data sub-directory within the corpus directory.
The files required for acoustic modeling in the Automatic Speech Recognition application include speaker_to_gender.txt and text.txt. The speaker_to_gender.txt file contains information mapping speaker IDs to their genders. The text.txt file contains information mapping audio data utterance ID's to their text transcriptions.
The files required for language modeling in the Automatic Speech Recognition application include lexicon.txt and nonsilence_phones.txt. The former contains information mapping all the words in the dataset with their corresponding phoneme transcriptions, while the latter includes all the individual phonemes used in the dataset.
See Contributors file for more details.
This project is licensed with MIT License. See License file for more details.