HuBERT Self-Supervised Speech Quiz

Answers and explanations appear as you take the quiz.

1. How does HuBERT generate the targets it tries to predict during training?
   • Offline k-means clustering of audio features into hidden units
   • Human transcripts
   • Random noise labels
   • Contrasting against distractor samples
2. What well-known text model inspired HuBERT's masked-prediction training scheme?
   • GPT-3
   • BERT
   • ResNet
   • Word2Vec
3. How does HuBERT improve its target labels over successive training iterations?
   • It hires more annotators
   • It adds more raw noise
   • It re-clusters using the model's own improved internal representations
   • It freezes the first clustering forever
4. What features are typically clustered in HuBERT's very first iteration?
   • Final transcript words
   • Image pixels
   • Speaker ID embeddings
   • Simple MFCC acoustic features
5. Why can HuBERT learn useful structure even when its clustering is imperfect?
   • It only needs to predict cluster IDs at masked positions, which stays a consistent task
   • It ignores the clusters entirely
   • It uses a perfect oracle clustering
   • It memorizes the test set
6. Which pretraining challenge does HuBERT address by using offline cluster labels?
   • Speech has no temporal structure
   • There is no lexicon of sound units available during pretraining
   • Every speaker uses the same pitch
   • Audio contains no repeated patterns
7. In the HuBERT paper, what relative WER reductions did the 1B-parameter model report on dev-other and test-other?
   • Exactly 50% on both sets
   • No measurable change on either set
   • Up to 19% on dev-other and 13% on test-other
   • Up to 13% on dev-other and 19% on test-other
8. Which benchmark sets did the original HuBERT paper use for speech-recognition results?
   • ImageNet and COCO
   • GLUE and SuperGLUE
   • SQuAD and TriviaQA
   • LibriSpeech and Libri-light