Deep learning for text-to-speech

Pretvorba besedila v govor je uporabna na različnih področjih. Z globokim učenjem lahko za glas take pretvorbe uporabimo poljubno osebo, če le imamo nekaj minut posnetkov njenega govora. Pretvorba posnetkov v nabor podatkov za učenje modelov je zamudno, zato smo izdelali programsko opremo, ki ta postopek olajša. Nato smo izdelali modele z uporabo implementacije Tacotrona in dveh vokoderjev: Griffin-Lim in WaveRNN. Na koncu smo izvedli primerjavo teh dveh vokoderjev in ugotovili, da je Griffin-Lim veliko hitrejši pri sintetiziranju govora kot WaveRNN, a je kvaliteta govora bistveno slabša.Text-to-speech (TTS) is useful in a variety of areas. With deep learning we can use any person\u27s voice for TTS, if only we have a few minutes of recordings of their speech. Converting the recordings into a dataset useful for model training is time consuming, so we created software that makes this process easier. We then created models using Tacotron and two vocoders: Griffin-Lim and WaveRNN. In the end we performed a comparison of these two vocoders and found that Griffin-Lim is much faster at synthesizing speech than WaveRNN, but the quality of speech is significantly worse