Robust Tamil Alphabet Recognition–Multiple Features and Machine Learning Paradigms

Abstract

The novel method of identifying Tamil alphabets from their sound patterns has potential uses in language learning and teaching. Given the script's various letters, the work automatically classifies Tamil alphabets through audio processing, presenting a distinct set of procedures. The audio files with Tamil alphabet pronunciations are divided into smaller segments according to the suggested methodology. The frequencies found in each sample are then examined using Fourier transforms, which serve as the foundation for the following feature extraction process. Features are extracted through the Perceptual Linear Prediction Cepstral Coefficients, which help capture the perceptual characteristics, and Mel – Frequency Cepstral Coefficients, which capture information on the distribution of frequencies. Two models were used to train and test the data: the Gaussian Mixture Model (GMM), which helps identify patterns and distinguish between features, and the Random Forest (RF) Classifier, which utilizes an ensemble of decision trees to find complex relationships between features and labels. The result of the testing process is the confusion matrix, which tells us about the model's accuracy and helps assess how well the system can correctly classify each letter. This work gives the best model for classifying the Tamil alphabet based on the confusion matrix results.