April 17, 2023
By Zuzana Ďaďová in Blog
Several months have passed since Phonexia expanded its portfolio with new speech-to-text languages. Therefore, we are excited to announce our latest addition of Ukrainian and Serbian Speech to Text!
Serbian Speech to Text
Creating a model that works great on Serbian speech has been a huge challenge for us. Why? Mainly due to the relatively recent standardization of the Serbian language and its historical and cultural contact with neighboring languages.
The Serbian Language from a Linguistic Point of View
In the last decades, there have been processes of standardization of Serbian, Croatian, Bosnian, and Montenegrin that have undeniable relevance for these countries.
However, we are also aware that in linguistics, it is usually considered to be one (macro)language that is pluricentric (i.e., one language with different national standards, like the situation with English or Spanish). It is called Serbo-Croatian, Serbo-Croat-Bosnian, and even Serbo-Croat-Bosnian-Montenegrin.
The point is that each country has its typical way(s) of speaking with its own specificities, characteristics, and standard norms, but, at the same time, they share an enormous amount of features in pronunciation, vocabulary, and grammar.
Up-to-date, Versatile, and Robust Speech-to-Text Model
Not only have we used diverse data sets for training the acoustics, but we also paid special attention to the language model. It consists mostly of Serbian text data, contains up-to-date language in terms of topics and vocabulary, and is extremely robust.
On the other hand, we couldn’t ignore the fact that Serbo-Croatian is a huge dialectal continuum (a series of language varieties spoken across various geographical areas where neighboring varieties are mutually intelligible). We thought of it as an advantage and considered it to be a nice challenge to make our speech to text work well even in the areas where Serbia borders Croatia, Bosnia, and also Montenegro.
For that purpose, we enriched the Serbian model also with Croatian and Bosnian acoustic and text data, making it extremely robust to handle the transcription of Serbian dialects well.
The final evaluation of the Serbian Speech to Text has shown a word accuracy of 68% to 86%, depending on the chosen data set.
Why the Latin Alphabet?
Even though we are aware of the importance of the Cyrillic alphabet – it being the official script of the Serbian administration – Phonexia’s new Serbian Speech to Text uses the Latin alphabet.
We decided to use Latin for two practical reasons:
- It is used in Serbia not only by individual speakers but also in newspapers, on social media, etc.
- It makes it user-friendly for people of different linguistic backgrounds.
Ukrainian Speech to Text
Spoken by more than 30 million people, the Ukrainian language is the third biggest Slavic language in the world. That, and the current geopolitical situation, made us decide it deserves its place in our Speech to Text portfolio.
A Quick Look at the Ukrainian Language
Although Ukraine’s only official language is Ukrainian, there are approximately 13 million native Russian speakers in the country. That may be the reason why, sometimes, Ukrainian is incorrectly considered to be the same language as Russian.
Yes, both are Eastern Slavic languages, and yes, they are similar, but they are definitely two distinctive languages (like Swedish and Norwegian or Spanish and Portuguese). And despite their similarities and slight overlap of vocabulary (typical for neighboring languages), we can observe their differences on all possible levels: pronunciation, words, morphology, and grammar.
We trained the Ukrainian Speech to Text model on a huge language model containing words that include recent topics like Covid-19 and the latest geopolitical events.
Because of this, Ukrainian Speech to Text is very robust. During our internal testing, it has achieved a word accuracy of between 61% and 85%, depending on the chosen evaluation data set.
Try Them Out
In total, the Ukrainian and Serbian languages are spoken by tens of millions of speakers in several European countries.
Convert their speech into text easily with our brand-new Ukrainian and Serbian Speech to Text!