Mimi Streaming Audio Codec Quiz

Answers and explanations appear as you take the quiz.

1. What lab released Mimi and the Moshi voice model?
   • Kyutai
   • OpenAI
   • DeepMind
   • Meta AI
2. Roughly how many tokens per second does Mimi emit for speech?
   • About 200
   • About 12.5
   • About 1,000
   • About 16,000
3. What information does Mimi’s first semantic codebook aim to represent?
   • Only the waveform’s stereo channel count
   • A transcript copied verbatim from a separate recognizer
   • Semantic speech information distilled from WavLM representations
   • The exact speaker identity with no acoustic tokens
4. Why is Mimi described as 'streaming' or 'causal'?
   • It only works on streamed music files
   • It requires a cloud stream to function
   • It streams text subtitles
   • It emits tokens as audio arrives instead of waiting for the full clip
5. Which quantization technique does Mimi use to encode audio?
   • Residual vector quantization (RVQ)
   • JPEG block compression
   • Run-length encoding
   • Principal component analysis
6. What does Mimi’s discrete audio-token representation let Moshi do?
   • Treat audio tokens as ordinary words with identical meanings
   • Model audio continuations as token sequences and decode them back into speech
   • Avoid using an audio decoder
   • Recover the original recording session stems
7. What is a reported frame size for Mimi’s streaming codec in the Moshi paper?
   • 8 seconds
   • 1 millisecond
   • 80 milliseconds
   • 10 minutes
8. Which encoder/decoder design does the Moshi report describe for Mimi?
   • A text-only BERT encoder with no audio decoder
   • A fixed lookup table with no learned encoder
   • A bidirectional non-causal decoder that waits for the entire conversation
   • Neural encoder and decoder Transformers with causal masking for streaming