Machine learning bias has been in the news a lot lately. And automatic speech recognition can also suffer from bias, including bias related to gender, race, age, and non-native speaker bias. The recent 2022 State of Voice Technology report indicates that 92% of companies believe voice technology bias has an important impact on their customers, and companies like Amazon are focused on reducing this bias for consumers using their Alexa device. But what if we could use bias to our advantage and intentionally bias our AI ASR systems to better hear and understand different groups? Instead of focusing on “fixing” bias, we fix the underlying problems with ASR accuracy for all gender, race, and ages so every group can be heard and transcribed accuracy.
In this panel discussion, we have brought in Adam Sypniewski, CTO of Deepgram, Rachel Tatman, Linguist and Senior Developer Advocate at RASA, and Scott Sandland, CEO of Cyrano AI, to moderate.
In this webinar we will discuss...
- What are the definitions of real-world and machine learning bias?
- How could “fixing” bias lead to other biases of the ASR system?
- What are solutions to correct biases for certain groups?