WFST Decoding in Speech Recognition Quiz

Answers and explanations appear as you take the quiz.

1. In the conventional HCLG recipe, what does L represent?
   • The acoustic waveform
   • The context-dependency transducer
   • The pronunciation lexicon mapping phones and words
   • The grammar or language model
2. Which component encodes word-sequence grammar or language-model constraints in HCLG?
   • H
   • G
   • C
   • L
3. What do arcs in a weighted finite-state transducer commonly carry?
   • State transitions and weights only, with no label mapping
   • Input labels only, without output labels or weights
   • Word probabilities without graph states
   • Input-output labels and weights
4. What does composing compatible transducers achieve?
   • Combines sequential mappings into a composed search graph
   • Adds graph arcs without checking whether symbol labels match
   • Discards output labels and keeps only acoustic states
   • Chooses the final recognized words without using acoustic observations
5. What does the acoustic model contribute during hybrid recognition?
   • Only the final grammar
   • The text normalization dictionary
   • Scores associated with acoustic states or transitions
   • The lexicon's word spellings
6. Why can a narrow grammar be risky in a decoder?
   • It guarantees too many pronunciations
   • It may exclude valid expressions outside its allowed paths
   • It removes the acoustic model
   • It makes symbol tables unnecessary
7. What must align for WFST composition to work correctly?
   • Every component has an identical number of states
   • All components use numerically identical weight scales
   • Every graph arc has the same cost
   • Compatible symbol labels between component transducers
8. Which claim is appropriately scoped to HCLG?
   • It is a graph composition used in conventional hybrid HMM recipes
   • It describes every end-to-end neural decoder
   • It is the name of a language model alone
   • It guarantees perfect transcription