AN OVERVIEW OF THE EXISTING TEXT-TO-SPEECH(TTS) ALGORITHMS BASED ON DEEP LEARNING
DOI:
https://doi.org/10.53297/18293336-2022.2-63Keywords:
speech synthesis, deep learning, mel spectrogram, attention approach, CNN (Convolutional Neural Network).Abstract
Text-to-speech technology has been around for a long time, but it is still considered challenging. The creation of human-like synthesized speech requires the solution of many tasks. As an example, the definition of intonation, the polyphony of systems, the combination of pairs of phonemes without audible flaws, and the definition of prosody.
Various methods exist for synthesizing speech from the text, including parametric synthesis, compilation synthesis, domain-specific synthesis, and full speech synthesis by rules. With the advent of the necessary computing resources and the accumulation of sufficient data for training, machine learning methods have become widely used in speech synthesis. With significant improvements in text-to-speech algorithms, text-to-speech is now widely used in mobile devices, smart speakers, and voice assistants. In spite of this, the algorithms of this class, developed at the present time, are not always able to complete the assigned task correctly. In some cases, they are unable to correctly put stress on the right parts of the text or speak with the right intonation.
In this study, the existing text-to-speech deep learning algorithms and models are reviewed and compared. The comparison is made by taking into account the positive and negative aspects of the algorithms, the complexity of implementing the algorithms, and the availability of the existing libraries for implementation. We consider the following algorithms as the main ones: WaveNet, DeepVoice 3, Tacatron 2, and Transformer.
