This resource is no longer available
Preventing Bias in Speech Technologies
The accent gap is an issue that plagues many speech recognitionmodels, caused by a lack of diverse training data. The result is systematic algorithmic bias within speech technology that causes many users to be misunderstood. At worst, the accent gap perpetuates the cycle of real-world racial bias, and alienates potential usersof various backgrounds.
The solution to this problem lies in the use of diverse training data, representative of a variety of accents and demographics.For example, non-native English speakers make up 350 million people in the US, 60% of which are native Spanish speakers. For companies to tap into this market with vast potential, they must train their models with Spanish-accented English speech data: data that represents the variety of ways in which people really speak.
Companies must prioritize this diverse training for speech models to remain competitive in the speech recognition marketand continue to move towards more inclusive speech technology for all.
- Nov 8, 2021
- Sep 20, 2021
- White Paper