Telugu Dialect Speech Dataset Creation and Recognition using Deep Learning Techniques

According to India’s 2011 demography, there seem to be approximately 8 crore Telugu communicators. Apart from that, the Telugu language has many dialects spread across the states Telangana and Andhra Pradesh. Telangana, Rayalaseema, and Coastal accents are the most common. The main concern is to understand the language irrespective of the dialects to have good communication near border areas of these states. Availability of data for analysis of Telugu speech dialects is of high scope for recognition. So, the creation of data is done for Telugu dialects with a total of 9 speakers, 3 speakers for each dialect. Once the data is created, analysis and recognition can help direct our needs. Classifying dialects cannot only solve this problem but also can act as a subset for solving bigger problems like machine translation, sentiment analysis, etc. We have used four RNN models viz. LSTM, GRU, BiLSTM & BiLSTM with attention layer for classification using speech data as input. Maximum test accuracy of 99.11% was obtained using the BiLSTM model with attention layer.