Home | News | 10.10.2024

 

Skopje, October 2024 – The University 'St. Cyril and Methodius' in Skopje (UKIM) today officially launched the new digital platform based on artificial intelligence for transcribing speech in the Macedonian language – "Buki", which enables the enhanced use of the official language when creating texts in digital form.

This model, which was created by Dejan Porjazoski, an expert in artificial intelligence technologies for speech recognition from Aalto University in Finland and Professor Dr Nikola Stikov, Professor of Biomedical Engineering at the Polytechnic School of the University of Montreal, in collaboration with Prof. Dr. Ordan Chukaliev, The head of the Centre for Advanced Interdisciplinary Research (CeNIIS) enables the transfer of spoken thought in the Macedonian language into written text that includes punctuation, lowercase and uppercase letters.

The Rector of UKI, Prof. Dr.  Biljana Angelova, at today's presentation, qualified the creation of the platform as a 'patriotic contribution' by the University and the Centre for Advanced Interdisciplinary Research (CeNIIS).

„This model can be of use to anyone who works with text transcription (interviews, lectures, field notes, reminders), as well as to people with disabilities, and we hope that this model will also find application in many digital tools that make our lives easier. Of course, we are currently presenting the first version of the digital model for the Macedonian literary language, but the team that developed the model is determined to create a second version for the transcription of the dialects of the Macedonian language, which will contribute to the scientific research work of UCM, but also preserve the dialects for future generations" – stated Rector Angelova.

She added that the development of an improved user interface and the addressing of the sustainability of the platform for transcribing Macedonian speech are also planned.

„Our Buki-w2v2 and Buki-Whisper models are based on an encoder–decoder architecture. The encoder's task is to transform the audio into a format suitable for decoding. The decoder, on the other hand, uses the transformed audio to generate a transcript. Buki-w2v2 and Buki-Whisper are adapted versions of the existing Wav2vec2 and Whisper, having been trained with our own resources in the Macedonian language," Porjazoski pointed out at the presentation.

The expert explained that, in order to verify the performance of their models, they tested them on various types of audio (read text, dialects, etc.) in comparison with the publicly available Whisper from OpenAI and MMS from Meta.

„As a metric, we used the Word Error Rate, i.e. the percentage of incorrect words in the transcript (a lower percentage = a better model). Our Buki-w2v2 and Buki-Whisper achieved scores of 9.0 and 8.1 respectively, far better than OpenAI's Whisper, which achieved 24.5, and Metal's MMS, which achieved 20.1," emphasised Porjazoski.

Professor Stikov pointed out that creating this model is of exceptional importance to him, initially for personal reasons and then because of the importance of preserving the Macedonian language and heritage.

„"Thousands of hours of recorded material are decaying in the damp corners of Macedonian institutions, waiting to be searchable and indexed. The 'Buki' model is the first step towards preserving this treasure. It has been trained with just 60 additional hours of speech, but it is already far superior to the competition," emphasised Stikov.

Prof. Ilina Jakimova, Head of the Institute for Ethnology and Anthropology at the Faculty of Natural Sciences and Mathematics – UKIМ, stated at the presentation that the Digital Archive of Ethnological and Anthropological Resources of the Institute for Ethnology and Anthropology at the Faculty of Natural Sciences and Mathematics contains over 1,250 hours of audio material, manually transcribed into Word files (mainly field interviews and oral testimonies), a large number of visual materials, a music collection and digitised manuscripts from our former professors and donors.

„We are honoured that a portion of the audio recordings and their transcriptions are included in the 'Buki' model, making up 40 per cent of its training. In selecting them, care was taken to ensure they were from different parts of Macedonia, i.e. in multiple dialects, with speakers of different genders and ages. This means that 'Buki' also indirectly incorporates archaic words and the voices of people who are no longer with us. Thus, through this product, the technology of the future meets the cultural heritage of the past. Seventeen complete issues of the Institute's international journal, 'EthnoAnthroZoom', were also used to train the model. They were recorded in an audio version of literary Macedonian and are available on the journal's website," explained Jakimova.

Minister for Digital Transformation, Mr Stefan Andonovski, He paid tribute to the project's organisational structure for the overall effort invested in creating the platform, pointing out that in a world of rapid technological change the promotion and protection of the Macedonian language is a genuine imperative, concluding with the following remark: "What you are doing is a truly patriotic act“.

„"It is now important for politics to recognise what you are doing as science and for us to find a way to apply artificial intelligence to the digital transformation of society, but also to create policies that will mean the real-world application of artificial intelligence in society, but in the right way, that is to say, not to lose our way or our vision of how we should develop, both from an ethical and a legal perspective," emphasised Minister Andonovski.

He emphasised that protecting and promoting the Macedonian language with digital tools is both a challenge and an opportunity, as it has the potential to develop across multiple spheres.

„Anyone who wishes to contribute will be able to donate their voice – to read a text in Macedonian and upload it to the Mozilla platform, for which more detailed information will be shared at the launch. To that end, UКИМ will launch "Donate a Voice" campaign„, UКИМ reported.

The name "Buki" derives from the name of the second letter of the Glagolitic alphabet Ⰱ ("buki") – meaning letter or writing. In English, it is transcribed as "Bookie.".

 

---

Namely, as of today the digital model is open for testing. Link to the platform: https://huggingface.co/Macedonian-ASR

Ss. Cyril and Methodius University in Skopje