Current voice recognition approaches use multi-task, multilingual models for
speech tasks like Automatic Speech Recognition (ASR) to make them applicable to
many languages without substantial changes. However, broad language coverage
can still mask performance gaps within languages, for example, across genders.
We systematically evaluate multilingual ASR systems on gendered performance
gaps. Using two popular models on three datasets in 19 languages across seven
language families, we find clear gender disparities. However, the advantaged
group varies between languages. While there are no significant differences
across groups in phonetic variables (pitch, speaking rate, etc.), probing the
model's internal states reveals a negative correlation between probe
performance and the gendered performance gap. I.e., the easier to distinguish
speaker gender in a language, the more the models favor female speakers. Our
results show that group disparities remain unsolved despite great progress on
multi-tasking and multilinguality. We provide first valuable insights for
evaluating gender gaps in multilingual ASR systems. We release all code and
artifacts at this https URL

多任务、多语种模型在语音识别中的广泛应用，虽然适用于多种语言，但仍存在在性别间表现差异的问题。我们通过系统评估多语种 ASR 系统在性别表现差异上的问题，并发现了明显的性别差异。在 19 种语言的三个数据集上，我们使用两种流行的模型，跨越了七个语系。尽管在声调、说话速度等语音学变量方面没有显著差异，并且模型内部状态的分析显示探针性能与性别差异之间存在负相关。也就是说，在一个语言中更容易区分说话者性别时，模型更倾向于女性说话者。我们的研究结果表明，尽管多任务和多语种方面取得了很大进展，但性别间的差异问题仍未解决。我们提供了有价值的首批关于多语种 ASR 系统性别差距评估的见解。我们将所有代码和相关资料发布在此 https URL

多语言语音模型在自动语音识别中表现出性别性能差距

Multilingual Speech Models for Automatic Speech Recognition Exhibit  Gender Performance Gaps

While universalism is a foundational principle of science, a growing stream
of research finds that scientific contributions are evaluated differently
depending on the gender of the author, with women tending to receive fewer
citations relative to men, even for work of comparable quality. Strikingly,
research also suggests that these gender gaps are visible even under blinded
review, wherein the evaluator is not aware of the gender of the author. In this
article, we consider whether gender differences in writing styles -- how men
and women communicate their work -- may contribute to these observed gender
gaps. We ground our investigation in a previously established framework for
characterizing the linguistic style of written text, which distinguishes
between two sets of features -- informational (i.e., features that emphasize
facts) and involved (i.e., features that emphasize relationships). Using a
large, matched sample of academic papers and patents, we find significant
differences in writing style by gender; women use more involved features in
their writing, a pattern that holds universally across fields. The magnitude of
the effect varies across fields, with larger gender differences observed in the
social sciences and arts humanities and smaller gaps in the physical sciences
and technology. Subsequently, we show that gender differences in writing style
may have parallels in reading preferences; papers and patents with more
informational features tend to be cited more by men, while those with more
involved features tend to be cited more by women, even after controlling for
the gender of the author, inventor, and patent attorney. Our findings suggest
that formal written text is not devoid of personal character, which could
contribute to bias in evaluation, thereby compromising the norm of
universalism.

本文研究了作者的性别是否会影响其在学术论文和专利申请中的写作风格，并发现女性在写作时更多地使用了涉及性的特征。研究表明，不同性别的作者所著作的文字在信息性和参与性特征上存在差异，这可能导致性别偏见并影响科学工作的公正性。