Systematic analysis and experimental evaluation of speech synthesis models for the Kazakh language
Neural text-to-speech technologies have advanced rapidly in recent years, however, selecting models that provide high synthesis quality with acceptable computational cost remains a challenging task for lowresource languages, including Kazakh. This study presents a comparative analysis of modern speech synthesis models for the Kazakh language with the aim of evaluating their computational efficiency, perceptual quality, and speech intelligibility. The experiments were conducted using synthesized Kazakh speech generated by the Tacotron 2, Piper-VITS, and MMS-TTS models. The evaluation was performed using the Real-Time Factor, the Mean Opinion Score, and the Word Error Rate. The experimental results demonstrate that Piper-VITS achieves the lowest RTF values and enables faster-than-real-time speech synthesis, whereas MMS-TTS provides the highest perceptual quality while maintaining moderate computational cost. Tacotron 2 produces acceptable synthesis quality but is significantly slower than the other evaluated models. The obtained results support the use of Piper-VITS and MMS-TTS for the development of practical speech synthesis systems for the Kazakh language.
Сведения
- Опубликовано
- 2026 · Т. 101, № 3, С. 125-135
- Язык
- Английский
- ISSN
- 2709-4693 · 2709-4707
- Издатель
- National Engineering Academy of the Republic of Kazakhstan
- Аффилиации
- Akhmet Baitursynuly Kostanay Regional University