SpecAugment for Speech Recognition Quiz

Answers and explanations appear as you take the quiz.

1. What does SpecAugment operate on?
   • The log-mel spectrogram
   • The raw audio waveform
   • The text transcript
   • The model's weights
2. Which of these is one of SpecAugment's three operations?
   • Pitch reversal
   • Frequency masking
   • Word substitution
   • Echo injection
3. What role did the SpecAugment paper assign to its feature transformations?
   • They guarantee that every recognizer avoids overfitting
   • They replace the need for transcribed training data
   • They augment ASR training inputs and were evaluated for recognition performance
   • They improve every speech model by the same amount
4. What does 'time masking' do?
   • Deletes the transcript timestamps
   • Speeds up the audio playback
   • Removes silence at the start
   • Blanks out a span of consecutive time steps
5. How are time- and frequency-mask locations chosen during SpecAugment training?
   • They are randomly sampled, so the same input may receive different masked regions on later passes
   • They are fixed permanently before the training set is collected
   • They are chosen by a language model after each recognition result
   • They are identical for every utterance and every training step
6. On which benchmark did SpecAugment notably help achieve state-of-the-art results?
   • ImageNet
   • LibriSpeech
   • SQuAD
   • GLUE
7. What does time warping do in the original SpecAugment policy?
   • It replaces speech frames with translated text
   • It masks every frequency channel permanently
   • It deforms feature positions along the time axis
   • It changes the speaker identity in the waveform
8. Which recognizer architecture did the original SpecAugment paper use in its experiments?
   • A region-based image detector
   • A music source-separation U-Net only
   • A text-only masked language model
   • Listen, Attend and Spell (LAS)