Викторина SoundStorm по параллельной генерации звука

Ответы и пояснения появляются по мере прохождения теста.

1. What generation change distinguishes SoundStorm from the cited autoregressive acoustic baseline?
   • Removal of all learned neural components
   • A guarantee of a single unconditional step for every token level
   • Parallel prediction of acoustic-token positions with iterative refinement
   • A requirement to produce only one waveform sample
2. Which attention direction does the SoundStorm paper use within its modeled token sequence?
   • Двунаправленное внимание
   • Only a mandatory causal left-to-right attention mask
   • No token interactions of any kind
   • Only a recurrent update without attention
3. Какое представление генерирует SoundStorm?
   • Непосредственные образцы необработанных сигналов
   • MIDI-ноты
   • Текстовые расшифровки
   • Дискретные акустические токены от нейронного кодека
4. What hardware-scoped timing result is reported in the SoundStorm abstract?
   • Every device generates all audio in exactly 0.5 seconds
   • 30 seconds of audio generated in 0.5 seconds on a TPU-v4
   • 30 seconds of audio always requires two hours
   • No hardware or duration was identified
5. How does SoundStorm’s within-level iterative decoding use confidence information?
   • It retains selected confident token candidates according to its schedule
   • It guarantees confidence equals factual correctness
   • It keeps only tokens whose names are alphabetically first
   • It never predicts masked positions
6. Что обычно используется SoundStorm для управления контентом?
   • Только случайный шум
   • Фиксированная мелодия
   • Встраивание изображений
   • Семантические токены таких моделей, как AudioLM или SPEAR-TTS.
7. In the stated SoundStorm decoding hierarchy, when are finer RVQ levels sampled relative to coarser ones?
   • Before any coarser-level information is available
   • After the required coarser-level tokens have been sampled
   • Only after all waveform samples have been printed as text
   • Without any ordering or conditional dependencies
8. What input supplies speaker-voice information in the reported prompted dialogue setup, distinct from semantic content tokens?
   • Only the output filename
   • Only a numerical browser-tab count
   • A short acoustic prompt containing the speakers’ voices
   • A required image of a music score