Scholarlink

Automatic speech recognition (ASR) systems have become increasingly popular due to their ability to enhance business communication and promote inclusivity. This article provides a concise overview of the general architecture of ASR systems and summarizes the latest scientific advancements in the field. Additionally, it reviews some of the most widely used ASR products and evaluates their performance in both Ukrainian and English languages. The comparison considers several criteria, including the number of model parameters, word error rate (WER), and character error rate (CER) for both languages and dataset processing time. Our findings suggest that ASR systems perform better in English than in Ukrainian. Notably, Facebook’s wav2vec2-large-960h-lv60 achieved the lowest WER (2.1588%) and CER (0.5708%) for English, while Nvidia Citrinet Large demonstrated the best performance for Ukrainian, with WER and CER values of 11.7691% and 8.1250%, respectively. The study also underscores the trade-off between model complexity and processing efficiency, as more complex models offer greater accuracy but require significantly more processing time.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Performance Evaluation of Automatic Speech Recognition Systems for Ukrainian and English Languages

  • 章节

  • 发表日期：

    2026年

查看全文
  • Andrii Dumyn,
  • Oleh Basystiuk,
  • Iryna Dumyn,
  • Nataliya Shakhovska

摘要

Automatic speech recognition (ASR) systems have become increasingly popular due to their ability to enhance business communication and promote inclusivity. This article provides a concise overview of the general architecture of ASR systems and summarizes the latest scientific advancements in the field. Additionally, it reviews some of the most widely used ASR products and evaluates their performance in both Ukrainian and English languages. The comparison considers several criteria, including the number of model parameters, word error rate (WER), and character error rate (CER) for both languages and dataset processing time. Our findings suggest that ASR systems perform better in English than in Ukrainian. Notably, Facebook’s wav2vec2-large-960h-lv60 achieved the lowest WER (2.1588%) and CER (0.5708%) for English, while Nvidia Citrinet Large demonstrated the best performance for Ukrainian, with WER and CER values of 11.7691% and 8.1250%, respectively. The study also underscores the trade-off between model complexity and processing efficiency, as more complex models offer greater accuracy but require significantly more processing time.