Bangla Text to Speech With Emotion

Abstract

Spoken language technology improved a lot. There are many text to speech models like tacotron, tacotron 2, deep voice, Fastspeech, and wavelet are used for synthesizing speech. Tacotron 2 has a 4.58 mean opinion score, mostly human-like speech generated by a computer. There is no work done for an emotional Bangla text-to-speech. This paper proposes a transfer learning approach for generating emotional speech with the respective tacotron two models. We have created a Bangla Emotional Text To speech web app. It generates Bangla speech for a text or audio input with a specific emotion. Initially, it gives speeches for three types of emotions and the poem. Users can choose which kind of emotions they want in the speech. Then their text will go to one of our tacotron two models. We have three models for creating sad, neutral, and happy speech. We have created another model for reading poems. For training our sad and happy model, we have built our dataset.