Viettel AI won the first prize 2 times in a row in speech recognition

The contest is part of the 9th International Conference on Language Processing and Vietnamese Voice, organized by VLSP Club, a branch of the Vietnam Informatics Association.

As an annual conference in the field of Language Processing and Vietnamese Speech, VLSP 2022 is a gathering place for leading researchers, experts and technology developers.

Speech recognition (Automatic Speech Recognition) is one of the important problems of Speech Processing in order to convert input speech signals into corresponding text. Exam teams are required to identify online lectures to meet the needs of automatically classifying, indexing, searching for lecture content, etc. from the data warehouse.

Viettel AI approaches the problem in the direction of making effective use of unlabeled raw data instead of just focusing on labeled data as usual.

Accordingly, the solution of Viettel AI, the AI ​​product ecosystem developed by Viettel Cyberspace Center, has applied some important improvements such as the method of masking voice signals in both time and frequency domains. or replace the Transformer model with a more advanced Conformer model, etc.

These improvements have helped Viettel AI effectively solve both data sets in the Speech Recognition category with an accuracy rate of up to 92.03%, while the accuracy of the remaining groups is from 67, 24 - 89.79%.

Viettel AI won for the second time in a row in the Speech recognition category, VLSP contest (Photo: Viettel AI).

This is the third and second year in a row Viettel AI won at the VLSP contest. Besides the 2 first prizes in the Speech recognition category, Viettel AI also won the second prize in the Emotional Speech Synthesis category.

Pioneering in the development and application of leading technologies in Speech Processing, products of Viettel AI artificial intelligence ecosystem such as virtual assistants, virtual switchboards, etc. can interact with more than 2,600 conversation scenarios with more than 96% accuracy, inspirational voice, 95% naturalness of real human voice.

Currently, these products have been widely applied in many businesses, agencies and departments in provinces and cities across the country. A representative of Viettel AI said that the unit will continue to develop and continuously upgrade products to increase accuracy, the ability to understand user intent and work performance.

Trong những năm gần đây, hội thảo VLSP tập trung tổ chức các cuộc thi về xử lý ngôn ngữ, nhằm thúc đẩy phát triển nghiên cứu cũng như tạo ra các bộ dữ liệu chung chia sẻ cho cộng đồng nghiên cứu VLSP.

Năm 2022, cuộc thi thu hút đông đảo các đội dự thi đến từ các trường đại học nổi tiếng trong và ngoài nước như Đại học Stanford - Mỹ, Viện Khoa học và Công nghệ Tiên tiến Nhật Bản (JAIST), cũng như các đội đến từ các doanh nghiệp công nghệ lớn như Viettel, Vin Group, FPT.