LSTM-Based Kazakh Speech Synthesis

Currently, the level of penetration of speech technology in modern life begins to vary greatly by country and by language environment. This is especially noticeable in the services developed in leading technology companies, where high-resource languages such as English, Russian, etc. have become the main service languages. Whereas, the speech technologies for under-resourced languages lag in their development. The article presents the first speech synthesis system based on the long-term short-term memory (LSTM) neural network architecture for the Kazakh language. The presented text-to-speech (TTS) system includes previously developed methods of prosodic processing for under-resourced languages and an acoustic model based on LSTM. The system receives the linguistic features of the text, including phonetic transcription, and it generates Kazakh speech with an acceptable quality of perception. Briefly summing up, this work describes the method of developing a speech synthesis for the Kazakh language, which has limited resources in terms of natural language processing. This approach can also be applied to other under-resourced languages.