Questions for ‘Talking through a tube can trick AI into mistaking one voice for another’

a photo of Yash Wani, a brown skinned man with short hair and glasse, positioning a microphone from the inside of a black tube

When people spoke through a simple tube with certain dimensions, a voice-recognition AI model wrongly tagged them as famous celebrities! Here, team member Yash Wani, an engineering PhD student at the University of Wisconsin–Madison, sets up a microphone to capture sound during a test of that system.

Todd Brown/University of Wisconsin–Madison

To accompany ‘Talking through a tube can trick AI into mistaking one voice for another


Before Reading:

  1. Describe what comes to mind when you think about artificial intelligence. How do you feel about advancements in artificial intelligence technology? For example, do you feel concerned? Do you feel optimistic? A little of both? How might a future world of advanced artificial intelligence look different from today?  
  2. Roll up a sheet of paper to make a tube shape. Working with a partner, compare the sound of your voice when speaking normally through the tube and then saying the same thing (in the same way) without the tube. Describe the differences you hear. Next, roll the paper tube tighter to narrow the opening. Then, compare the sound of your voice when speaking through a narrow tube versus the wider tube. Describe any differences you hear. 

During Reading:

  1. Why would technology that makes one person’s voice sound like another’s be a cybersecurity concern?
  2. How has deepfake software helped hackers target people’s bank accounts? What “added protections” helped address this problem?
  3. What did these added protections overlook that this new study attempts to address?
  4. Regarding sound waves, what is the relationship between pitch and frequency? 
  5. Besides tube width, what other physical aspects of a tube can affect how that tube alters someone’s voice? 
  6. How many volunteers participated in this study? How many volunteers could impersonate some of the celebrities in the data set? 
  7. Were artificial-intelligence models or human volunteers more accurate at identifying voices?
  8. Name one celebrity whose voice was impersonated by volunteers in this study.

After Reading:

  1. On average, what percentage of the time did the “tube voices” fool the AI? How did this share compare to that in human volunteers? How did researchers explain this difference? Based on this explanation, how might researchers update their AI models to improve their ability to spot “tube voice” fakes? 
  2. Besides what’s discussed in this story, what other changes might be made to a tube to alter someone’s voice? (Hint: What else could you change about the tube besides its shape?) Come up with two changes. For one of these, design a follow-up study to investigate the impact of this change on AI’s ability to identify people by their voice.