Permutation Invariant Training Quiz

Answers and explanations appear as you take the quiz.

1. What core problem does permutation invariant training (PIT) solve?
   • Deciding which output channel should correspond to which speaker during training
   • Compressing audio for storage
   • Reducing microphone hardware costs
   • Translating separated speech into text
2. How does PIT decide which error to use when updating the model?
   • It averages the error across all outputs equally
   • It tries every output-to-source pairing and keeps the one with the lowest error
   • It always assigns the loudest voice to output 1
   • It ignores the references entirely
3. Without a solution like PIT, what tended to happen to separation model outputs?
   • They became perfectly clean immediately
   • They ran much faster
   • They produced blurry, averaged outputs because of conflicting label assignments
   • They required no training data
4. For separating N speakers, how many permutations must PIT consider per step?
   • N (linear in speakers)
   • N squared
   • Always exactly two
   • N factorial
5. What advantage does utterance-level PIT (uPIT) add over frame-level assignment?
   • It keeps a speaker in a stable output channel across the whole utterance
   • It removes the need for any references
   • It only works on music, not speech
   • It doubles the number of speakers
6. Which approach can sidestep the permutation problem altogether?
   • Increasing the audio bitrate
   • Target-speaker extraction that conditions on a reference voice cue
   • Using only mono recordings
   • Disabling backpropagation
7. What extra inference step does utterance-level PIT avoid compared with frame-level PIT?
   • Running speech recognition before source separation
   • Estimating the number of microphones in the recording
   • Solving an additional output-speaker permutation problem during inference
   • Recomputing the training loss after every generated word
8. Which benchmark does the guide name as an example for evaluating speech-separation models?
   • LibriSpeech for single-speaker transcription
   • ImageNet for visual recognition
   • GLUE for language understanding
   • WSJ0-2mix for overlapping-speaker separation