Викторина, связанная с руководством · Средний Уровень
Шепчущие галлюцинации в викторине о тишине
Тестирование галлюцинаций «Шепот тишины», проверки источника звука, эвристика отсутствия речи, пороговые компромиссы, отрицательные тестовые клипы и контрольные журналы.
Вопрос 1 из 8
A transcript contains a fluent sentence during a verified silent interval. What is the main concern?
Предварительный просмотр каждого вопроса8 вопросы
Ответы и пояснения появляются по мере прохождения теста.
- A transcript contains a fluent sentence during a verified silent interval. What is the main concern?
- A hidden speaker is guaranteed
- The words may be a hallucination unsupported by audio
- Silence proves the sentence is quoted audio
- The source file necessarily changed
- Why should a no-speech threshold be tested on faint real speech?
- All faint utterances are noise
- Threshold changes cannot create omissions
- An aggressive setting may suppress words that were actually spoken
- The decoder always knows the speaker
- Which source establishes Whisper’s documented possibility of audio-unrelated transcripts?
- The original Whisper research paper
- Only a product screenshot
- Only an unrelated vision benchmark
- Only a user’s social profile
- Что требует открытый контроль возможных галлюцинаций в тишине в своей документированной реализации?
- A camera depth stream
- A human transcript for every incoming clip
- A fixed speaker identity label
- Word timestamps enabled for that skip logic
- Why is a low WER on ordinary spoken clips insufficient for this failure mode?
- WER already measures silent false text in every corpus
- It may not test silent or non-speech segments where false text can appear
- Silence always adds many reference words
- All quiet sounds appear in spoken-speech tests
- A reviewer cannot hear a disputed phrase clearly. What should the transcript show?
- Publish the most fluent completion as fact
- Assign a speaker without evidence
- Increase display confidence without listening
- Uncertainty, not an unmarked best guess
- Which evaluation pair captures the threshold tradeoff?
- False words on silence and missed words in faint speech
- Only model size and throughput
- Only transcript fluency
- Only total audio file count
- What should a high-stakes workflow keep for disputed text?
- Only the polished transcript
- Only a screenshot of the output
- Original audio, timestamps, model/settings and human review record
- Only the highest decoder score