This study used a multi-talker database containing intelligibility scores for 2000 sentences (20 talkers, 100 sentences), to identify talker-related correlates of speech intelligibility. We first investigated "global" talker characteristics (e.g., gender, F0 and speaking rate). Findings showed female talkers to be more intelligible as a group than male talkers. Additionally, we found a tendency for F0 range to correlate positively with higher speech intelligibility scores. However, F0 mean and speaking rate did not correlate with intelligibility. We then examined several fine-grained acoustic-phonetic talker-characteristics as correlates of overall intelligibility. We found that talkers with larger vowel spaces were generally more intelligible than talkers with reduced spaces. In investigating two cases of consistent listener errors (segment deletion and syllable affiliation), we found that these perceptual errors could be traced directly to detailed timing characteristics in the speech signal. Results suggest that a substantial portion of variability in normal speech intelligibility is traceable to specific acoustic-phonetic characteristics of the talker. Knowledge about these factors may be valuable for improving speech synthesis and recognition strategies, and for special populations (e.g., the hearing-impaired and second-language learners) who are particularly sensitive to intelligibility differences among talkers.

中文翻译：

---

正常语音的清晰度 I：全局和细粒度的声学语音说话者特征。

本研究使用包含 2000 个句子（20 个说话者，100 个句子）的可懂度分数的多说话者数据库来识别与说话者相关的语音清晰度相关性。我们首先调查了“全局”说话者特征（例如，性别、F0 和语速）。调查结果表明，女性说话者作为一个群体比男性说话者更容易理解。此外，我们发现 F0 范围倾向于与较高的语音清晰度分数呈正相关。然而，F0 均值和语速与可懂度无关。然后，我们检查了几个细粒度的声学语音说话者特征作为整体可懂度的相关性。我们发现元音间距较大的说话者通常比间距较小的说话者更易懂。在调查两种一致的听众错误（片段删除和音节关联）的情况下，我们发现这些感知错误可以直接追溯到语音信号中的详细时序特征。结果表明，正常语音清晰度的很大一部分变化可追溯到说话者的特定声学-语音特征。有关这些因素的知识对于改进语音合成和识别策略以及对于对说话者之间的可懂度差异特别敏感的特殊人群（例如，听力受损和第二语言学习者）可能很有价值。结果表明，正常语音清晰度的很大一部分变化可追溯到说话者的特定声学-语音特征。有关这些因素的知识对于改进语音合成和识别策略以及对于对说话者之间的可懂度差异特别敏感的特殊人群（例如，听力受损和第二语言学习者）可能很有价值。结果表明，正常语音清晰度的很大一部分变化可追溯到说话者的特定声学-语音特征。有关这些因素的知识对于改进语音合成和识别策略以及对于对说话者之间的可懂度差异特别敏感的特殊人群（例如，听力受损和第二语言学习者）可能很有价值。