Scholarlink

<p>This paper presents a&#xa0;systematic comparative analysis of open Russian-language audio datasets used in automatic speech recognition, speech synthesis, and audio deepfake detection. The study covers corpora containing real human speech (Golos, Mozilla Common Voice, Russian LibriSpeech, M‑AILABS Speech Dataset, RUSLAN, Dusha), as well as datasets containing synthetic speech or paired real-fake recordings (Open STT, XMAD-Bench, MLAAD, SpeechFake, PolyGlotFake). For each resource, we analyze its size, speech type (read, spontaneous, emotional), speaker diversity, recording conditions, annotation structure, and suitability for anti-spoofing tasks. The results show that read-speech corpora provide a&#xa0;high degree of textual alignment and acoustic control but lacks the variability of natural communication. In contrast, large, crowdsourced datasets offer greater acoustic and stylistic diversity while introducing annotation heterogeneity. Specialized anti-spoofing datasets enable the study of artifacts produced by modern TTS and voice conversion models; however, Russian-language data remain unevenly represented. We conclude that reliable synthetic speech recognition systems require the integration of many complementary datasets, and we outline key areas for the development of Russian-language speech resources in “Far Speech”.</p>

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

Russian-language voice datasets: a comparative analysis for audio deepfake detection tasks

  • 期刊论文

  • 发表日期：

    2026年4月21日

查看全文
  • Ivan Palkin,
  • Alexey Ruchay

摘要

This paper presents a systematic comparative analysis of open Russian-language audio datasets used in automatic speech recognition, speech synthesis, and audio deepfake detection. The study covers corpora containing real human speech (Golos, Mozilla Common Voice, Russian LibriSpeech, M‑AILABS Speech Dataset, RUSLAN, Dusha), as well as datasets containing synthetic speech or paired real-fake recordings (Open STT, XMAD-Bench, MLAAD, SpeechFake, PolyGlotFake). For each resource, we analyze its size, speech type (read, spontaneous, emotional), speaker diversity, recording conditions, annotation structure, and suitability for anti-spoofing tasks. The results show that read-speech corpora provide a high degree of textual alignment and acoustic control but lacks the variability of natural communication. In contrast, large, crowdsourced datasets offer greater acoustic and stylistic diversity while introducing annotation heterogeneity. Specialized anti-spoofing datasets enable the study of artifacts produced by modern TTS and voice conversion models; however, Russian-language data remain unevenly represented. We conclude that reliable synthetic speech recognition systems require the integration of many complementary datasets, and we outline key areas for the development of Russian-language speech resources in “Far Speech”.