Guide-linked quiz · Medium Level

ASR Confidence Scores and Calibration Quiz

Test ASR word confidence, raw versus calibrated scores, reliability plots, threshold tradeoffs, high-confidence errors, and human review.

Related guide pathsAsr Confidence Scores
Question 1 of 8

A decoder gives a word score of 0.9. What is needed before calling it a 90% correctness probability?

Preview every question8 questions

Answers and explanations appear as you take the quiz.

  1. A decoder gives a word score of 0.9. What is needed before calling it a 90% correctness probability?
    • A larger vocabulary by itself
    • A polished display of the number
    • Held-out calibration evidence for that model and setting
    • The top decoder score without reference testing
  2. What does a reliability plot compare?
    • Predicted confidence with observed correctness across score groups
    • Word length versus audio duration
    • Speaker age versus vocabulary size
    • File bitrate versus punctuation count
  3. Why might high aggregate calibration be insufficient for a payment command?
    • Every critical number shares the average error rate
    • A high score constitutes authorization
    • Calibration prevents all rare-word errors
    • Rare amounts or names may have different error patterns
  4. A review threshold catches more transcription errors but flags many correct words. What tradeoff is this?
    • Higher detection with no review cost
    • Higher error detection with more unnecessary human review
    • A change in reference transcript length
    • Proof that all errors were fixed
  5. Why should editors still sample some high-confidence words?
    • High confidence guarantees correctness
    • Sampling changes the trained model weights
    • Confidently wrong recognitions can escape low-score filters
    • Only low-score words can be wrong
  6. How can a new microphone affect confidence interpretation?
    • The score-to-correctness relationship may drift
    • Microphones never change acoustic conditions
    • Calibration is fixed across every device
    • Only the human reference changes
  7. Which use fits ASR confidence best?
    • Authenticate the speaker’s legal identity
    • Replace all source-audio review
    • Certify no semantic errors
    • Prioritize uncertain spans and request confirmation where mistakes matter
  8. What should be reported alongside confidence calibration?
    • Only model file size
    • How well scores separate wrong from correct words and under what conditions
    • Only transcript display design
    • Only the number of output words

Keep testing yourself

More quizzes picked for your level.

View all quizzes