Scholarlink

<p>Standard Moroccan Amazigh (Moroccan Berber) lacks any text-to-speech system despite its rich phonemic inventory. High-quality parallel data are scarce and costly for this low-resource language. We introduced a cross-lingual transfer learning approach that leverages phonemic similarity: after ranking related languages via Pearson correlation of shared phonemes, we initialized phoneme embeddings from pretrained models fine-tuned on those auxiliaries. Our TTS comprises a spectrogram predictor and neural vocoder, trained on a small dataset of Amazigh. We evaluated four strategies:Ar-Am SpeechT5 pretrained on Arabic, then fine-tuned on Amazigh; Pr-Am SpeechT5 pretrained on Persian, then fine-tuned on Amazigh; Pr-Ar-Am SpeechT5 pretrained on Persian, subsequently fine-tuned on Arabic and finally on Amazigh and En-Am SpeechT5 pretrained on English, then fine-tuned on Amazigh. Sequential cross-language transfer outperformed all other models achieving a MOS score of 3.92 the best among all evaluated approaches, demonstrating its value for low-resource TTS.</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

Analyzing cross-language similarities to enhance low-resource text-to-speech via transfer learning, case study: the Moroccan Berber Amazigh

  • 期刊论文

  • 发表日期：

    2025年12月24日

查看全文
  • Fatim Ezzahrae Dorhmi,
  • Imane Lasri,
  • Naoufal El-Marzouki,
  • Anouar Riadsolh,
  • Mourad Elbelkacemi,
  • Hamid Ez-Zahraouy

摘要

Standard Moroccan Amazigh (Moroccan Berber) lacks any text-to-speech system despite its rich phonemic inventory. High-quality parallel data are scarce and costly for this low-resource language. We introduced a cross-lingual transfer learning approach that leverages phonemic similarity: after ranking related languages via Pearson correlation of shared phonemes, we initialized phoneme embeddings from pretrained models fine-tuned on those auxiliaries. Our TTS comprises a spectrogram predictor and neural vocoder, trained on a small dataset of Amazigh. We evaluated four strategies:Ar-Am SpeechT5 pretrained on Arabic, then fine-tuned on Amazigh; Pr-Am SpeechT5 pretrained on Persian, then fine-tuned on Amazigh; Pr-Ar-Am SpeechT5 pretrained on Persian, subsequently fine-tuned on Arabic and finally on Amazigh and En-Am SpeechT5 pretrained on English, then fine-tuned on Amazigh. Sequential cross-language transfer outperformed all other models achieving a MOS score of 3.92 the best among all evaluated approaches, demonstrating its value for low-resource TTS.