Wav2Letter Convolutional ASR Quiz

Answers and explanations appear as you take the quiz.

1. Which architecture defines the original wav2letter acoustic network?
   • A recurrent-only network
   • A decision-tree ensemble
   • A Gaussian-mixture-only acoustic network
   • A convolutional network without recurrent acoustic connections
2. Which research organization is associated with the original 2016 wav2letter paper?
   • Google Brain
   • Microsoft Research
   • Facebook AI Research
   • NVIDIA
3. Which output representation does the original wav2letter acoustic model predict from its studied speech inputs?
   • Only an intermediate human-labeled phoneme transcript
   • Letters or graphemes
   • Only a list of speaker names
   • Only a global tempo value
4. How does the original AutoSegCriterion (ASG) loss differ from the more common CTC loss?
   • It removes the blank symbol and models letter transitions directly
   • It requires per-frame manual alignment
   • It only works with RNNs
   • It cannot handle variable-length sequences
5. What computation difference does a convolutional acoustic stack offer compared with a conventional recurrent dependency across time?
   • Every complete speech decoder becomes fully parallel
   • No labeled transcription data is needed for its supervised objective
   • Convolution operations can be evaluated across positions without that recurrent chain
   • Every workload necessarily obtains zero transcription error
6. Which C++ machine-learning library documents ASR applications formerly associated with the wav2letter project?
   • scikit-learn only
   • An audio-compression format
   • A required spelling dictionary standard
   • Flashlight
7. Which tensor library does the later wav2letter++ paper identify for its C++ framework?
   • ArrayFire
   • A required Java virtual machine
   • Only the original Torch7 runtime
   • A fixed spreadsheet engine
8. For two length-three temporal convolutions at stride one and dilation one with no padding or intervening operation, how many original positions can one second-layer output depend on?
   • Three
   • Five
   • Six
   • Nine