Rono8/kiswahili_tts

Realt-Time State-of-the-Art Speech Synthesis in Tensorflow v2 for Kiswahili Language This repository is an implementation of Text To Speech and has been possible due to original work done on Text To Speech in English and other Languages like German and Korean, which can be found here. # Requirements - Python 3.8.3 - Tensorflow 2.2 - cuda 11.4 - cudaDNN 8.2 - Tensorflow Addons >= 0.10.0 # Data Collection This project was implemented on Dataset obtained from the Word Project. Its a collection of all Bible books from Genesis to Revelation. Each of the book is written chapter-wise, and a corresposnding audio for the words written. A total of 7108 sentences was obtained, ranging for 5 different books. # Data Preprocessing For data processing, the format followed was like the ljspeech dataset. The audio files were programatically split into short audio clips based on silence. They were then combined based on a random length such that each of eventual audio file lies between 1 to 10 seconds as its for the ljspeech dataset, but still maintaining the order in which they followed each other. This was done using python 3. The audio files were saved as a single channel,16 PCM WAVE file with a sampling rate of 22.05 kHz The texts for each of the audio files were then manually mapped by listening to each of the audio file. At the end, a total of 7108 lines of sentences were obtained, corresponding to the audio clips. For the individual sentences, the format is id, text and normalised text each separated by a pipe ``` id | text | normalized text e.g. Kiswa-00001 | Mithali za Sulemani,Mlango 1, Mithali za Suleimani | Mithali za Sulemani,Mlango wa kwanza, Mitahli za Suleimani ``` An example audio clip, it's for the above text, is below: Listen to Kiswa-00001.wav: Your browser does not support the audio element. In case your browser does not support this audio format, download it here Kiswa-00001.wav # Background Research When doing research on Language Processing …