News

Participation in radio show "Digital icons". About Serbian language and speech technologies for Radio Belgrade 2 speakers were prof. Vlado Delić from Faculty of Novi Sad and Vesna Rehner from „Alfanum" company. (link)

Participation in an event Anniversary of the .срб domain – RNIDS on 26th January 2023. It was streamed via YouTube channel (0:38:40-1:19:55) (link)

After the interviews with the AI project leaders, S-ADAPT project has been listed as an exemplary AI research project at the website of the project MISSION4.0 (link)

Participation in RTS television show “Scientific portal”. It has been recorded and aired on the Serbian Broadcasting Corporation's channel 2 (3 min about the project) (link).

S-ADAPT project brochure in Serbian has been finished and distributed

Professor Vlado Delić held a lecture "Progress in the development of speech technologies based on AI project S-ADAPT" on the Workshop of AI project Com-in-AI at the Facultz of Electronic Engineering in Niš, 31st May 2022.

Participation in RTS television show “Science in movement” on 27th May 2022. It has been recorded and aired on the Serbian Broadcasting Corporation's channel 2 (12 min about the project) (link).

Professor Vlado Delić held a lecture "Development of speech technologies and applications in smart homes and smart cities" in Serbian Chamber of Engineers on a workshop “Smart building and smart city management systems” on 26th May 2022.

A visit of a distinguished lecturer from ISCA (International Speech Communication Association) has been arranged. Prof. Reinhold Haeb-Umbach held a lecture "Far-field speech recognition" on 28th March 2022 at the Faculty of Technical Sciences in Novi Sad. The next day S-ADAPT project team have a meeting with professor Haeb-Umbach. On 31st March, prof. Haeb-Umbach held a lecture "Computational Analysis of Sound Scenes and Events" in Faculty of Electronic Engineering in Niš, after which the meeting with the Com-in-AI project team was organized.

Nikola Simić held a lecture "Practical application of neural networks in speaker and emotion recognition" in workshop of the AI project Com-in-AI in Niš, 19.11.2021.

Expressive speech database (9 hours of amateur recordings of sentences in acted emotional states) has been obtained via smartphone application, and is in its final stages of prrocessing. The subjective evaluation of the database using PHP has been completed, which is a requirement for publication in a scientific journal. The publication is being prepared, whereafter the database will be made public. It will be used for emotion/speaker recognition and adaptation at the Project.

Algorithm development is based on two new Linux GPU servers (DNN training), and the database is stored at a Windows server at AlfaNum company (SQL) and transferred to NAS storage at the Scientific-Technological Park of FTN. Activities concerning the development of models for speaker/emotion recognition based on "spectrogram images" and MEL cepstral coefficients has continued – testing of CNN configurations for speaker/emotion classification (models will also be used for speaker/emotion conversion). Research into recognition of speaker gender and age has also started, based on open corpora. Experiments include DNN models based on ResNet architecture extended by Squeeze and Excitation levels as well as transformers. Language model adjustment aimed at implementation of speech dialogue with advanced adaptation methods into a virtual speech assistant application. As to image processing, a module for semi-supervised training of CycleGAN goal function (concerning image transformation with few known pairs) has been implemented and improved by additional bootstrap of the generator statistics. Additional literature was studied, related to cost function optimization, including a specific approach concerning metrics generating geodesics on the cone of positive definite matrices.

Project S-ADAPT is presented in the article on the website of the National Geografic as one of the best projects from Serbian scientists in the area of artificial intelligence. Read the article.

Seminar on "Inclusion of persons with disabilities in the digital society", organized by the “Hendi centra Koloseum", 03.12.2020. Branislav Popović held a lecture „Speech technologies for people with disabilities“ in Serbian language on the occasion of marking the International Day of Persons with Disabilities, under the slogan "For a life without barriers" (link).

The project S-ADAPT was presented at the Science Fund exposition dedicated to the promotion of the Programme for the development of projects in the domain of artificial intelligence, which took place on October 9th 2020 at Savsko šetalište, Kalemegdan (Belgrade).

The project S-ADAPT has been presented for the first time within the lecture entitled „Human-Machine Speech Communication X & Y vs. Z & α Generation“, delivered by the S-ADAPT Project Manager prof. Vlado Delić, PhD, at the international conference MIPRO in Opatija, Croatia, on October 1st 2020.

In the first month of the project we organized kick-off meeting, made project logo, launched the website and investigated literature. Also, we are preparing necessarry software environment and started testing publicly available code for image style translation based on cycle-gan. What is more, we specified necessary equipment and expecting the tender to be finished soon. We have also started implementation of the application for the database collection.