Scholarlink

Two speech recognition systems, SeamlessM4T-Large v1 and SeamlessM4T-Large v2, are evaluated in this study for the task of speech-to-text in Moroccan Arabic, a dialect which is poorly resourced. We wish to evaluate the performance of this model when transcribing Moroccan Arabic spoken sentences into text. Both models show comparable accuracy rates on the held-out test data, with a rate of 96.42% for the model v1 and 95.27% for the v2. In results, the performance of SeamlessM4T Large version 1 was better than version 2. In version one, the mean accuracy is 78.31%. Meanwhile the mean accuracy for version two is 76.65%. The results indicate that for the transcription of Moroccan Arabic, the v1 model is more reliable than the v2 model.

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

Evaluating Multilingual Speech Models for Low-Resource Dialects: Moroccan Arabic with SeamlessM4T-Large

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yassine El Kaneb,
  • Mohcine Kodad

摘要

Two speech recognition systems, SeamlessM4T-Large v1 and SeamlessM4T-Large v2, are evaluated in this study for the task of speech-to-text in Moroccan Arabic, a dialect which is poorly resourced. We wish to evaluate the performance of this model when transcribing Moroccan Arabic spoken sentences into text. Both models show comparable accuracy rates on the held-out test data, with a rate of 96.42% for the model v1 and 95.27% for the v2. In results, the performance of SeamlessM4T Large version 1 was better than version 2. In version one, the mean accuracy is 78.31%. Meanwhile the mean accuracy for version two is 76.65%. The results indicate that for the transcription of Moroccan Arabic, the v1 model is more reliable than the v2 model.