语言复杂度与语音识别准确率:正字法复杂度有害,音法复杂度无关
计算与语言
2024-06-14 v1 人工智能
摘要
我们研究语言因素如何影响自动语音识别(ASR)模型的性能。我们假设正字法和音法复杂度都会降低准确率。为此,我们在 25 种语言(15 种书写系统)上微调了 multilingual self-supervised pretrained model Wav2Vec2-XLSR-53,并比较了这些语言的 ASR 准确率、 grapheme 数量、unigram grapheme entropy、logographicity(词/ morpheme 级信息编码在书写系统中的程度)以及 phoneme 数量。结果表明,正字法复杂度与低 ASR 准确率显著相关,而音法复杂度显示无显著相关性。
引用
@article{arxiv.2406.09202,
title = {Language Complexity and Speech Recognition Accuracy: Orthographic Complexity Hurts, Phonological Complexity Doesn't},
author = {Chihiro Taguchi and David Chiang},
journal= {arXiv preprint arXiv:2406.09202},
year = {2024}
}
备注
11 pages, 5 figures, 5 tables, submitted to ACL 2024