Викторина, связанная с руководством · Средний Уровень
Объясненная викторина о частоте ошибок в словах
Тестовые замены, удаления и вставки, знаменатель опорного слова, нормализация, срезы подгрупп и пределы WER.
Вопрос 1 из 8
A reference says “red car” and ASR says “blue car.” Which word error occurred?
Предварительный просмотр каждого вопроса8 вопросы
Ответы и пояснения появляются по мере прохождения теста.
- A reference says “red car” and ASR says “blue car.” Which word error occurred?
- A substitution of red with blue
- A deletion of red with no replacement
- An insertion of blue while red remains
- No error because both words name colors
- Which count forms the WER denominator?
- The number of words in the ASR output
- The number of audio frames
- The number of utterances
- The number of words in the reference transcript
- Why can WER exceed 100 percent for a short reference?
- Substitutions always count twice
- Many inserted words can outnumber the reference words
- The denominator is the output length
- The formula clips all rates at one
- The reference is “send five boxes”; ASR says “send nine boxes now.” What errors appear?
- Two deletions with no insertion
- Only an insertion of now
- One substitution and one insertion
- No error if punctuation is ignored
- Why agree on number normalization before comparing models?
- “21” versus “twenty one” can score differently under different token rules
- Numbers are not spoken words
- Normalization re-records the audio
- One universal normalizer resolves every ambiguity
- What does low overall WER not guarantee?
- That clinically important terms were all right
- That a reference transcript was used
- That relatively few word edits occurred overall
- That medication names or negations are transcribed correctly
- Which report can reveal a hidden accent or microphone gap?
- Only the model size
- WER broken down by relevant speaker and recording groups
- Only one aggregate WER
- Only training loss on the same speakers
- For a live captioning product, what should accompany final WER?
- Only the audio compression ratio
- Only the vocabulary size
- Partial/final output latency and stability
- Only punctuation counts