Sirvi Autor "Zuppur, Hain" järgi

Nüüd näidatakse 1 - 2 2

Aligning contextual vector spaces between independent neural translation systems
(Tartu Ülikool, 2023) Zuppur, Hain; Fišel, Mark, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Numerous pre-trained machine translation models are available for translating between different languages. However, these models are limited to a fixed set of languages they were trained for. When there is no translation model available for a specific language pair, we need to translate to one or more intermediate languages, which can result in reduced translation quality. We investigate the possibility of combining two translation models by aligning the vector spaces between them using a simple regressor. We explore the effectiveness of various regression methods for achieving this alignment and evaluate their performance. We show that combining two different translation models is possible, although doing so leads to a decrease in translation quality.
Creating a Voice Conversion Model for Estonian
(Tartu Ülikool, 2021) Zuppur, Hain; Rätsep, Liisa, juhendaja; Tartu Ülikool. Loodus- ja täppisteaduste valdkond; Tartu Ülikool. Arvutiteaduse instituut
Voice conversion has a variety of uses, like enhancement of impaired speech or entertainment purposes. The main challenge in voice conversion is extracting speaker-independent linguistic features from speech. To date, one of the most promising solutions is the Cotatron model. Estonian has some high-quality speech synthesis models, but there are no voice conversion models for Estonian. This thesis aims to take the Cotatron model and train it using Estonian Text-to-Speech datasets to produce a voice conversion model for the Estonian language.