Abstract:
"
Speech recognition has been a hot topic as the intelligent era has developed. Even
though numerous automated speech recognition (ASR) programs have been available,
a significant number of them do not support Tamil with full features. And when it
comes to practical usability, just supporting only Tamil language words will not be an
option since the English words are often used in the day-to-day conversation.But still
considering a mixed language scenario will not totally satisfy the scenario without
considering the different accent in language. So that the research has been done by
considering both Tamil-English mixed language and accented speech.
Data has been collected from two group of people with different Tamil accent (Sri
Lankan Tamil , Indian Tamil) and the ASR models have been created using the open sourcetool “Kaldi”.“Mel Frequency Cepstral Coefficient”(MFCC) has been used for
feature extracting and Hidden Markov Model (HMM) and Gaussian Mixture Model
(GMM) based monophone and triphone models were created and tested. From the
triphone model, the best model has been selected and used for the hybrid model
creation by replacing the GMM with neural network.
Model accuracy has been compared based on the WER and SER value for each model
and also benchmarked with the previous systems. The results showedaccuracy
improvement for the hybrid modelcompared to triphone models.
Key words :Automatic speech recognition, Neural network , Acoustic models, Kaldi,
Mixed language ASR"
