Verified, then logged.
Sun 22 Nov 2026 · 00:21 KST / 15:21 UTC한국어 | ENSearch /

Breaking01 / AI

Alania-2 lowers word error rate in Turkish speech synthesis

The paper also reported results from a text normaliser that spells out numbers and names as Turkish speakers would pronounce them.

At a window desk overlooking Hanoi's old quarter, a person seen from behind corrects punctuation in red pen on a blurred news printout beside an open novel. AI-generated image

Alania-2 achieved a word error rate of 1.44% in a Turkish speech evaluation, its researchers reported in a paper published on 11 Nov. The figure came from their own evaluation.

According to the paper’s abstract, Alania-2 converts text into speech. The researchers introduced features for designing voices through text descriptions and cloning voices from reference audio, alongside built-in voices.

The researchers reported using a text normaliser, which rewrites numbers, abbreviations and foreign names into words matching Turkish pronunciation. In their published results, 86 of 96 test sentences containing digits preserved all critical information through speech synthesis and recognition. Without the normaliser, none preserved all critical information.

The researchers said they trained the model from scratch on 18,195 hours of Turkish speech. Synthetic speech generated by a larger model and filtered automatically accounted for 90% of the training audio.

The abstract describes Freya-TR-Eval as an evaluation of Turkish sentences scored using automatic speech recognition. On the same evaluation, the researchers’ published word error rates were 5.99% for their previous system and 1.44% for Alania-2.

The researchers did not disclose human listening evaluation scores in the abstract.