Winning Viettel the Giant, Sun* places third in the Text To Speech VLSP 2019 competition

With a total score of 3.80, Sun* won third place in the shared task of Text to Speech organized by the 6th International Workshop on Vietnamese Language and Speech Processing (VLSP 2019).

Text to Speech is one of the 4 shared tasks of VLSP 2019 aiming at solving the challenges in natural Vietnamese processing. 

When registration is opened, the competition attracts about 40 teams, including many big tech companies in Vietnam like VNG, Zalo, Viettel...  

The challenge of the competition is that the training data provided by the organizer is extremely static. In order to provide the best training data for the model, the teams had to filter out the data, then put into machine learning. With the data of 1000 audios (a total of 45 minutes) for Southern Vietnamese accent and 15,000 audios (a total of 23 hours) for Northern Vietnamese accent, 2 weeks of processing are extremely pressing. This is the reason why many teams 'gave up'. 

Despite being the first timer in VLSP 2019, Sun* did not give in but rise up as an excellent candidate for the champion. 

The final result is Sun* won Viettel the Giant for the third place with the total point of 3.80/5, a very close defeat to the 2 'brutes' Zalo (3.94/5) and VNG (3.85/5). 

Mr. Pham Huu Quang - R&D Unit shared his feelings about this competition: "With many years of experience in speech processing, VNG winning the first place is no surprise. After this competition, we have gained lots of experience to improve for the next data batch".

Mr. Quang's biggest regret is not having enough time to filter out all training data. For that reason, he had to put in static data for machine learning. 

"Next year, Sun* will continue to improve our shortcomings to climb up the rank in VLSP 2020" - Mr. Quang said. 

Dr. Nguyen Thi Thu Trang - Hanoi University of Science and Technology said what impressed her the most about Sun*'s product is that the synthesized speech quality is very natural and identical to the real accent. With this criteria,  Sun* got 4.13/5, surpassing Zalo with 4.10/5.

 

Before that, Sun* also won First Prize for Hate Speech Detection on Social Networks - 1 of 4 criteria of VLSP 2019. 

Read more: Victory for Sun* Bear at the national online contest on natural language processing

Get the hottest updates from Sun* News every day at 8h00 and 13h00!

#Sun*

#Viettel

#VNG

#Zalo