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

Bài kiểm tra chưng cất mô hình chưng cất-thì thầm và ASR

Kiểm tra việc gắn nhãn giả giữa giáo viên và học sinh, phạm vi điểm kiểm tra tiếng Anh, sự cân bằng giữa tốc độ và chất lượng, tài liệu tham khảo độc lập, bài kiểm tra im lặng và thẻ mẫu.

Đường dẫn hướng dẫn liên quanChưng cất thì thầm Asr chưng cất
Câu hỏi 1 của 8

In Distil-Whisper research, what supplies many of the student’s training targets?

Xem trước mọi câu hỏi8 câu hỏi

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

  1. In Distil-Whisper research, what supplies many of the student’s training targets?
    • Only manual labels from the final test set
    • Only the untrained student’s guesses
    • Transcriptions generated by a larger Whisper teacher
    • Image captions unrelated to the audio
  2. Why can student training inherit an error from the teacher?
    • Distillation validates every teacher word
    • The student never receives text targets
    • Teacher outputs are independent human reference transcripts
    • Pseudo-labels can contain the teacher’s wrong transcript
  3. Which scope claim is supported by the distil-large-v3 card?
    • It is verified for every language and task
    • It targets English ASR as a replacement within that stated use
    • It cannot produce errors on silence
    • It guarantees identical outputs to the teacher
  4. A team wants a faster model on its own device. What should it benchmark?
    • Latency, memory and transcript errors on that hardware and audio
    • Only a benchmark headline speedup
    • Only the number of model files
    • Only fluency of one demo clip
  5. What evidence is independent of teacher-student agreement?
    • Agreement on teacher-labeled training audio
    • Only the student confidence score
    • Human reference transcripts on held-out deployment-like audio
    • Only a model-family name
  6. Why include silence and non-speech in evaluation?
    • Silence guarantees correct output
    • Spoken-audio WER covers every quiet failure
    • Distillation rules out hallucination
    • Smaller ASR models may still emit unsupported text on quiet inputs
  7. Which property cannot be inferred from a checkpoint being called “distilled”?
    • That teacher-student training was used
    • Equivalent accuracy in every accent, language and task
    • That a smaller model may need fewer resources
    • That the exact checkpoint scope should be read
  8. Before using an English-targeted checkpoint for another language, what is the right step?
    • Check the exact model card and evaluate a suitable model for that language
    • Assume the teacher’s multilingual scope transfers
    • Translate the checkpoint name only
    • Treat an English benchmark as proof elsewhere