引導式測驗 · 中等 等級
SoundStorm 並行音訊產生測驗
探索 SoundStorm 的編解碼器令牌和基於置信度的並行解碼,以及語義調節、RVQ 順序、語音提示和範圍計時。
問題 1 的 8
哪項生成變化使 SoundStorm 與所引用的自回歸聲學基線區別開來?
預覽每個問題8 問題
當您參加測驗時,就會出現答案和解釋。
- 哪項生成變化使 SoundStorm 與所引用的自回歸聲學基線區別開來?
- 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
- SoundStorm 論文在其模型化 token 序列中使用了哪種注意力方向?
- 雙向注意力
- Only a mandatory causal left-to-right attention mask
- No token interactions of any kind
- Only a recurrent update without attention
- SoundStorm 產生什麼樣的表示形式?
- 直接原始波形採樣
- MIDI 音符
- 文字記錄
- 來自神經編解碼器的離散聲學標記
- SoundStorm 摘要中報告了哪項硬體範圍的時間結果?
- 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
- SoundStorm 的層內迭代解碼如何使用信心資訊?
- 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
- SoundStorm 通常以什麼為條件來控制內容?
- 僅隨機噪聲
- 固定的旋律
- 影像嵌入
- 來自 AudioLM 或 SPEAR-TTS 等模型的語意標記
- 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
- 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