Industrial Research into ASR Bias'

ASR 偏差的工业研究

• 批准号: 10093501
• 金额: $ 145.15万
• 依托单位: WYSER LTD
• 依托单位国家: 英国
• 项目类别: Collaborative R&D
• 财政年份: 2024
• 资助国家: 英国
• 起止时间: 2024 至 无数据
• 项目状态: 未结题
• 来源: https://gtr.ukri.org/projects?ref=10093501
• 关键词: Industrial; Research; into; ASR; Bias

This research will focus on reducing biases in foundational Automatic Speech Recognition (ASR) services (e.g. OpenAI's Whisper). The aim is to reduce ASR biases for minority linguistic groups in UK advice giving settings.This research follows on from our identification of bias in foundational ASR models for UK accents. Our aim is to further develop this research by now focusing on understanding what is the most efficient method to reduce this bias now and in the future. Here we focus on advice giving settings, but this work will transfer to other uses of ASRs such as personal assistants or other conversational settings e.g. customer service.By mitigating the biases in these foundation ASR services this research paves the way for more equitable AI, AI that is more socially responsible, trustworthy, and therefore widely adopted.

这项研究将侧重于减少基础自动语音识别（ASR）服务（例如OpenAI的Whisper）中的偏见。其目的是减少英国建议提供环境中少数语言群体的ASR偏见。这项研究是在我们对英国口音的基本ASR模型的偏见识别之后进行的。我们的目标是进一步发展这项研究，重点是了解现在和将来减少这种偏见的最有效方法。这里我们关注的是建议提供设置，但这项工作将转移到asr的其他用途，如个人助理或其他会话设置，如客户服务。通过减轻这些基础ASR服务中的偏见，本研究为更公平的人工智能铺平了道路，人工智能更具有社会责任感，更值得信赖，因此被广泛采用。