Scholarlink

This paper explores the feasibility of synthesizing Norwegian dialects using two state-of-the-art Text-to-Speech (TTS) models, Matcha-TTS and IMS-Toucan, trained from scratch on limited data. Participants evaluated audio clips from Trøndelag and Vestlandet for authenticity, naturalness, and intelligibility on a Likert scale (1–5). IMS-Toucan was consistently rated higher across all evaluation criteria and both dialects, with Mean Opinion Scores (MOS) of 3.39 for authenticity, 3.41 for naturalness, and 3.12 for intelligibility, significantly outperforming Matcha-TTS, which scored 2.08, 1.90, and 2.13, respectively. These findings show that current TTS systems can be used to generate dialects, even for low-resource languages like Norwegian.

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

Synthesizing Norwegian Dialects in Low-Resource TTS

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Victoria Langø,
  • Zohaib Hassan,
  • Steven Hicks

摘要

This paper explores the feasibility of synthesizing Norwegian dialects using two state-of-the-art Text-to-Speech (TTS) models, Matcha-TTS and IMS-Toucan, trained from scratch on limited data. Participants evaluated audio clips from Trøndelag and Vestlandet for authenticity, naturalness, and intelligibility on a Likert scale (1–5). IMS-Toucan was consistently rated higher across all evaluation criteria and both dialects, with Mean Opinion Scores (MOS) of 3.39 for authenticity, 3.41 for naturalness, and 3.12 for intelligibility, significantly outperforming Matcha-TTS, which scored 2.08, 1.90, and 2.13, respectively. These findings show that current TTS systems can be used to generate dialects, even for low-resource languages like Norwegian.