Jasper and QuartzNet ASR Quiz

Answers and explanations appear as you take the quiz.

1. Which factorization distinguishes QuartzNet’s time-channel separable convolution from a dense temporal convolution?
   • Replacing every neural layer with a GMM
   • Removing all channels from the input
   • Adding a mandatory recurrent connection at every step
   • Separate depthwise temporal filtering and pointwise channel mixing
2. In the QuartzNet paper’s stated comparison, which approximate count is listed for QuartzNet-15x5 versus JasperDR-10x5 at 333 million parameters?
   • 250 million
   • 500 million
   • 19 million
   • One billion
3. Which organization is associated with the cited Jasper and QuartzNet research papers?
   • Google
   • NVIDIA
   • Amazon
   • OpenAI
4. What loss function do Jasper and QuartzNet use to train without explicit alignments?
   • Connectionist Temporal Classification (CTC)
   • Mean squared error
   • Cross-entropy on fixed frames
   • Triplet loss
5. Which structural feature is described for the Jasper convolutional stack?
   • Only alphabetical sorting of frames
   • A guarantee that all gradients have equal magnitude
   • Residual connections
   • Random deletion of every deep layer
6. Which augmentation combination does the QuartzNet paper report for its best evaluated training setup?
   • Only deleting all training transcriptions
   • A required pronunciation dictionary replacement at each frame
   • A guarantee that no data augmentation is used
   • Speed perturbation with a cutout-style spectrogram method
7. Under the stated CTC path mapping, what labels result from A, A, blank, A?
   • AA
   • A
   • AAA
   • An empty sequence
8. For a length-three depthwise temporal filter on four input channels plus a pointwise map to eight output channels, how many filter weights are counted, excluding biases?
   • 96
   • 44
   • 12
   • 32