Scholarlink

<p>In this work, we explore text-to-speech (TTS) synthesis for the Tamazight language, a low-resource and underrepresented language in the field of speech technology. Leveraging SpeechT5, a transformer-based encoder-decoder model pre-trained by Microsoft, we fine-tuned it on a carefully curated Tamazight dataset. The dataset combines contributions from community-sourced corpora, including the Mozilla Common Voice project, to ensure linguistic diversity and speaker variability. A specialized text normalization pipeline was introduced to convert Tifinagh script into Latin transcription, improving model compatibility and phonetic consistency. We further incorporated speaker embeddings using SpeechBrain’s x-vector extractor to enable high-quality speaker-conditioned synthesis. Our experiments demonstrate that SpeechT5 can be successfully adapted to the Tamazight language, producing intelligible and natural-sounding speech. The results support the viability of TTS systems in low-resource and indigenous language contexts. The findings also underscore the importance of transfer learning and data augmentation techniques in bridging linguistic gaps and enabling inclusive speech technologies for communities with limited digital resources.</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

Fine-tuning SpeechT5 for tamazight text-to-speech: a foundation for educational technology in low-resource languages

  • 期刊论文

  • 发表日期：

    2026年2月20日

查看全文
  • Youssef Amkrane,
  • Fatima Amounas,
  • Mohamed Badiy,
  • Mourade Azrour,
  • Abdulatif Alabdulatif

摘要

In this work, we explore text-to-speech (TTS) synthesis for the Tamazight language, a low-resource and underrepresented language in the field of speech technology. Leveraging SpeechT5, a transformer-based encoder-decoder model pre-trained by Microsoft, we fine-tuned it on a carefully curated Tamazight dataset. The dataset combines contributions from community-sourced corpora, including the Mozilla Common Voice project, to ensure linguistic diversity and speaker variability. A specialized text normalization pipeline was introduced to convert Tifinagh script into Latin transcription, improving model compatibility and phonetic consistency. We further incorporated speaker embeddings using SpeechBrain’s x-vector extractor to enable high-quality speaker-conditioned synthesis. Our experiments demonstrate that SpeechT5 can be successfully adapted to the Tamazight language, producing intelligible and natural-sounding speech. The results support the viability of TTS systems in low-resource and indigenous language contexts. The findings also underscore the importance of transfer learning and data augmentation techniques in bridging linguistic gaps and enabling inclusive speech technologies for communities with limited digital resources.