Identification and Analysis of Assamese vowel speech signal using Formant Feature-Fusion and Feed-Forward Neural Network Model

Objective: Vowel identification from speech helps translate spoken words to their textual forms through computational linguistics. Even though the Assamese language has Eight vowel sounds, Eleven vowels are present in the Assamese script. However, four vowels (ই, ঈ, উ, ঊ) are no longer pronounced differently by typical Assamese speakers. Since pronunciation plays a vital role in translating speech to text, the present work aims to identify andanalyse Assamese vowels, which can lead to good Assamese Speech to Text translation. Methods: In this work, we have builta speech dataset by recording all Eleven vowels by 26 speakers needed for the vowel identification purpose. We used the Feed Forward Neural Network to train the model using the Formant Frequencies. Findings: The neural network model developed with the help of formant frequencies gave an accuracy rate of 89parcent for identifying the Assamese vowels. The confusion matrix demonstrates that the vowel set, i.e., [ই, ঈ] and [উ, ঊ], are pronounced in a very similar manner by the informants. Such similar pronunciation can be due to various factors like negligence, lack of linguistic knowledge, popular culture, etc., which led to the misclassification of these vowel sets. Novelty: No prior work has been done in Assamese vowel recognition using a combination of Formant Feature Fusion and Neural Networks.