Voicebox Flow-Matching Speech Generation Quiz

Answers and explanations appear as you take the quiz.

1. What single training task is Voicebox optimized for?
   • Predicting masked audio given surrounding context and text
   • Classifying speaker emotion
   • Translating text between languages
   • Detecting deepfakes
2. Which generative technique does Voicebox use instead of diffusion?
   • Reinforcement learning
   • Conditional flow matching
   • Decision trees
   • K-means clustering
3. How does Voicebox condition speech generation on a target speaker without retraining?
   • It retrains the model for every speaker
   • It selects from a fixed list of speaker IDs only
   • It uses a short reference segment as unmasked audio context
   • It averages all training speakers into one voice
4. What efficiency result did the Voicebox paper report for its zero-shot TTS comparison?
   • It required 20× more generation time
   • It ran without any neural network
   • It always produced audio in a single millisecond
   • It generated speech up to 20× faster than the compared VALL-E system in the reported setup
5. Why did Meta initially withhold the full Voicebox model?
   • Concerns about voice-cloning misuse and deepfakes
   • It was too slow to run
   • It only worked in one language
   • It produced no usable audio
6. How does one model handle editing, cloning, and denoising in Voicebox?
   • By switching to a different network for each task
   • By framing every task as predicting masked audio from context
   • By using separate task-specific training runs
   • By manually scripting each output
7. How many languages did the Voicebox paper report for its multilingual training setup?
   • One language only
   • Two languages
   • Six languages
   • More than one hundred languages
8. How much speech data did the Voicebox paper report using?
   • About 50 hours
   • About 500 hours
   • More than 50 million hours
   • More than 50,000 hours