Bài kiểm tra có liên kết hướng dẫn · Trung bình Cấp độ

Điểm tin cậy ASR và bài kiểm tra hiệu chuẩn

Kiểm tra độ tin cậy của từ ASR, điểm thô so với hiệu chuẩn, biểu đồ độ tin cậy, sự cân bằng ngưỡng, lỗi có độ tin cậy cao và đánh giá của con người.

Đường dẫn hướng dẫn liên quanĐiểm tin cậy của Asr
Câu hỏi 1 của 8

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

Xem trước mọi câu hỏi8 câu hỏi

Câu trả lời và giải thích xuất hiện khi bạn làm bài kiểm tra.

  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