Moshi Full-Duplex Speech Quiz

Answers and explanations appear as you take the quiz.

1. What does 'full-duplex' mean in the context of Moshi?
   • It can listen and speak at the same time
   • It only works in two languages
   • It needs two microphones
   • It doubles the audio volume
2. Which organization released Moshi?
   • OpenAI
   • Kyutai
   • Google DeepMind
   • Meta
3. What is Mimi in the Moshi system?
   • A search engine
   • A graphics renderer
   • A neural audio codec that turns speech into discrete tokens
   • A text translation API
4. How does Moshi differ from a traditional voice-assistant pipeline?
   • It only outputs text
   • It requires no model at all
   • It works only offline with no AI
   • It processes audio directly instead of chaining STT, an LLM, and TTS
5. Roughly what conversational latency does Moshi target?
   • Around 160-200 milliseconds
   • About 2-3 seconds
   • About 10 seconds
   • Exactly 1 minute
6. What is the purpose of Moshi's 'Inner Monologue' method?
   • To hide the conversation from users
   • To predict text before audio, improving linguistic quality
   • To translate everything into French
   • To compress video
7. Mimi encodes audio at 12.5 discrete frames per second. About how far apart are consecutive frames?
   • About 8 milliseconds
   • About 800 milliseconds
   • About 80 milliseconds
   • About 1 second
8. Which two synchronized streams does Moshi model in parallel?
   • Two unrelated text transcripts
   • The user’s video and the system’s screen
   • Two copies of the same language-model weights
   • The user’s incoming audio and the system’s own outgoing audio