Recent advances in model designs, training methods, and generalization across speakers and languages have accelerated deep learning-based TTS systems. These advances have not been fully tested for Indian language voice synthesis. This study is computationally expensive because to the large number of Indian languages, limited resources, and untested neural TTS technologies. This research employs an attention sequence-to-sequence neural network to turn text input into natural-sounding voice to detect Dravidian languages including Tamil, Malayalam, Tamil Code Mix, and Malayalam Code Mix. It is used to map a sequence of text tokens (e.g. characters, phonemes, or words) to a sequence of acoustic features (e.g. mel-spectrograms or waveforms) that can be used to generate speech. The language-agnostic model surpasses all other models in the task of language identification for Dravidian languages. The results of the proposed model, namely Att_Seq2Seq, have been analyzed empirically with performance measures and found that it achieves 12.4 dB of mel-cepstral distortion, 11.4% of root-mean-square error and 32 s of computatational time.