Phoneme based Kannada Speech Corpus for Automatic Speech Recognition System

Abstract

Automatic speech recognition is a challenging task in languages which are not having many resources for research. The state-of-the-art technologies like Deep Neural Network approach require a standard dataset for training and testing the approach for a particular language. In this paper, development of a novel dataset for Kannada speech corpus is described. The speech dataset developed is useful for implementation of Automatic Kannada speech recognition system. The speech data for this dataset is crowd-sourced by using a website developed for the purpose. Since the data collection process is active in the internet the size of the dataset grows as more number of people contributes their voice. The system asks the user to enter their details when they open it for the first time. The user-data collected here is helpful to categorize the data in the dataset based on different parameters, which is the requirement for some types of ASR system. The phonetic representation of the word is stored in the database along with the numeric representation which optimizes the processing of speech to text conversion.