0:0019:30

Voice AI: The Future is By Everyone and For Everyone

Joan Bajorek 19:30 2020

Your voice assistant may grade you a D plus, and it isn't your fault

Read the same paragraph to a voice assistant and a white man might score an A minus while a mixed-race woman scores a D plus. Many users blame themselves when the system fails them, and companies lose buyers who give up on the product. Joan Bajorek, a speech technology PhD and CEO of Women in Voice, asks what it would take to build voice AI that truly works for everyone, everywhere.

See the evidence, the business case and the projects building voice AI for every speaker

Watch on YouTube

In short

Joan Bajorek, CEO of Women in Voice and Head of Conversational Research and Strategy at VERSA, shows how voice AI has spread into search, homes and multimodal devices worldwide. She argues that speech recognition still carries significant race, gender and dialect bias, and that building for every speaker is both a moral obligation and a large financial opportunity.

“And I think one of the things I struggle with terribly is that people believe it's their own fault, but the tech doesn't understand them. And that's utterly not the case.”
“I really don't believe that there are people at these companies being like, Oh, let's design something that's horribly biased against people.”

What you will learn

  • Explain what voice AI is and where multi-turn systems like Google Duplex still fall short.
  • Recognise how multimodal interfaces combine voice, vision, gesture and augmented reality to fit a use case.
  • Spot race, gender and dialect gaps in speech recognition instead of blaming users for being misunderstood.
  • Build a financial as well as a moral case for supporting more languages, dialects and speakers.
  • Find projects such as Mozilla Common Voice and communities such as Women in Voice that widen who voice tech serves.

Details

Level
Beginner
Format
Talk
Technical depth
Introductory technical
For
All career stages
Length
19:30
Published
July 2020

Questions this talk answers

What is voice AI?

It is a system where a computer interprets spoken input, processes it and may answer by voice. Google Duplex, which phones a salon to book an appointment in a multi-turn conversation, is one of the most sophisticated consumer examples.

How widespread is voice technology already?

As the speaker cited, Google reports that around 20% of its searches are made by voice, and smart devices see hundreds of millions of weekly interactions. Many people use voice tech without realising it.

Does speech recognition work equally well for everyone?

No. Research she cites shows accuracy gaps by gender, race and dialect; in her scorecard example a white man gets an A minus while a mixed-race woman gets a D plus reading the same paragraph.

Why should companies invest in voice tech that works for more speakers?

Beyond the moral obligation, there is money at stake: a coworker skipped buying a smart fridge because it would not understand him and his wife, and the speaker cited $1.7 trillion in Latinx purchasing power in the United States.

How can people help make voice AI more inclusive?

They can donate their voice to Mozilla's Common Voice, which makes voice data openly available with attention to demographics, and join communities such as Women in Voice.

About this talk

Voice AI is increasingly ubiquitous and democratized, voice AI is shifting from a glitchy IVR world to more and more human-like, sophisticated, multi-turn interactions. The ubiquity of these products cannot be underestimated, i.e. Google Assistant will replace voice search in the Chrome browser in 2020 (Voicebot.ai, 2020). Yet it is an open question whether voice AI supports users everywhere and how far we are from supporting a mission of everyone and everywhere. While significant financial opportunities alone may be enough for some companies to consider demographics during research and design phases (Bajorek, 2019), multilingual and multi-dialectal roadmap creation can be a tricky feat to prioritize numerous scripts, low-resource datasets, and models trained with English as the default framework that extrapolate poorly (Bender, 2019; van Esch, 2019). This talk provides concrete examples of robust fiscal and moral reasons to support a wide variety of languages and dialects internationally and the evolution of our builds to support voice technology of the future.

Topics

Related talks

Browse the library