Connectionist Temporal Classification Quiz

Answers and explanations appear as you take the quiz.

1. What core problem was CTC designed to solve for speech recognition?
   • Aligning unsegmented audio frames to output text without frame-level labels
   • Compressing audio files for storage
   • Translating speech from one language to another
   • Removing background noise from recordings
2. What is the special 'blank' token in CTC used for?
   • To mark the end of the entire sentence only
   • To represent 'no new output here' and to separate genuine repeated characters
   • To store the language model probability
   • To indicate silence in the original audio waveform
3. How does CTC compute the loss for a target transcription?
   • It picks the single most likely alignment and ignores the rest
   • It compares pixel values frame by frame
   • It sums the probabilities of all valid frame-paths that collapse to the target text
   • It averages the audio amplitude over time
4. What are the two collapsing rules CTC applies to turn a frame-path into final text?
   • Capitalize letters, then add punctuation
   • Reverse the sequence, then remove vowels
   • Add spaces, then translate to phonemes
   • Merge consecutive repeated characters, then delete blank tokens
5. What simplifying assumption does standard CTC make about its outputs?
   • Each frame's prediction is conditionally independent of the others given the audio
   • Outputs depend heavily on previously emitted characters
   • The audio must be exactly one second long
   • Every character takes exactly one frame
6. Why is beam search with a fused language model often preferred over greedy decoding in CTC?
   • It runs faster on every device
   • It compensates for CTC's lack of an internal language model, improving accuracy
   • It removes the need for the blank token
   • It eliminates the need to train the network
7. What length constraint does CTC place on an output sequence relative to input frames?
   • The target must always have exactly the same length as the input
   • The target must contain twice as many labels as input frames
   • The target cannot require more output labels than there are input time steps
   • CTC requires a fixed 256-token output
8. How can CTC compute the sum over many valid alignments without explicitly listing every path?
   • By selecting one random path and ignoring all others
   • By running a separate language model for every audio frame
   • By sorting the transcript alphabetically
   • Using a forward-backward dynamic-programming recurrence over an alignment lattice