Conformer Architecture Quiz

Answers and explanations appear as you take the quiz.

1. What two mechanisms does the Conformer architecture combine in a single block?
   • Convolution and self-attention
   • Recurrence and pooling
   • Decision trees and clustering
   • Beam search and tokenization
2. Why is combining convolution and attention especially useful for speech?
   • Attention handles local phoneme detail while convolution handles global context
   • Convolution captures local patterns while self-attention captures long-range context
   • Both only matter for images, not audio
   • Neither helps; it is purely for speed
3. What is the 'Macaron' or sandwich structure of a Conformer block?
   • Two attention layers around one convolution
   • A single feed-forward layer only
   • Attention and convolution wrapped between two half-step feed-forward modules
   • Convolution stacked on top of convolution
4. Which organization introduced the Conformer, and in what year?
   • OpenAI in 2018
   • Meta in 2015
   • Microsoft in 2012
   • Google in 2020
5. What does the convolution module inside a Conformer typically include?
   • A pointwise conv, a GLU, a depthwise conv, normalization, and a Swish activation
   • Only a single fully connected layer
   • A recurrent LSTM cell
   • A random forest classifier
6. On LibriSpeech test-clean/test-other, what WER did the original Conformer paper report with an external language model?
   • 2.1% / 4.3%
   • 1.9% / 3.9%
   • 9.1% / 9.3%
   • 0.1% / 0.3%
7. What positional information does the Conformer paper add to attention?
   • A geographic GPS coordinate for every speech frame
   • A fixed class label for each utterance
   • Relative positional encoding
   • A separate video frame index
8. What word error rates did the Conformer paper report without an external language model on LibriSpeech test/test-other?
   • 20.1% / 40.3%
   • 0.1% / 0.3%
   • 12.1% / 14.3%
   • 2.1% / 4.3%