Recurrent Neural Network based approach to recognize assamese vowels using experimentally derived acoustic-phonetic features

Author

Sharma, Mukesh ; Sarma, M. ; Sarma, Kandarpa Kumar

Author_Institution

Dept. of Electron. & Commun. Eng., Gauhati Univ., Guwahati, India

fYear

2013

fDate

13-14 Sept. 2013

Firstpage

140

Lastpage

143

Abstract

Vowels are the phonemes with greatest intensity and low frequencies. Assamese, which is considered as the lingua-franca of the entire north-east India, has eight vowel phonemes namely /i/, /e/, /ε/, /a/, /0/, /?/, /o/ and /u/. A Recurrent Neural Network (RNN) based algorithm is described in this paper for the recognition of the vowel sounds from Assamese speech. The feature vector is generated by considering the acoustic phonetic features of vowels like duration, fundamental frequency (F0) and the four formant frequencies (F1, F2, F3 and F4). From the experimental results a recognition rate of 84 % is obtained which can be considered to be satisfactory in comparison to the current phoneme recognition strategy.

Keywords

feature extraction; natural language processing; recurrent neural nets; speech processing; speech recognition; Assamese speech; Assamese vowel recognition; North-East India; RNN; acoustic-phonetic features; feature vector; formant frequencies; fundamental frequency; lingua-franca; recurrent neural network based approach; vowel sound recognition; Acoustics; Frequency measurement; Recurrent neural networks; Speech; Speech recognition; Training; Vectors; Acoustic Phonetic Features; Formants; Fundamental Frequency; Recognition; Recurrent Neural Network (RNN); Vowels;

fLanguage

English

Publisher

ieee

Conference_Titel

Emerging Trends and Applications in Computer Science (ICETACS), 2013 1st International Conference on

Conference_Location

Shillong

Print_ISBN

978-1-4673-5249-9

Type

conf

DOI

10.1109/ICETACS.2013.6691411

Filename

6691411