Викторина, связанная с руководством · Средний Уровень
Смещение акцента в тесте на распознавание речи
Проверяет, как измеряется ошибка ASR, какие два первичных исследования обнаружили в ограниченных выборках и почему не следует исключать расу, диалект и акцент.
Вопрос 1 из 8
What does word error rate (WER) measure in speech recognition?
Предварительный просмотр каждого вопроса8 вопросы
Ответы и пояснения появляются по мере прохождения теста.
- What does word error rate (WER) measure in speech recognition?
- The speaker’s age
- The number of microphones used
- The share of audio with background music
- Substitutions, deletions and insertions relative to a reference transcript
- How many commercial ASR systems did Koenecke et al. test in their 2020 study?
- Два
- Twelve
- Сто
- Пять
- What average WER difference did the 2020 study report in its structured-interview sample?
- The same WER for both groups
- 0.35 for Black speakers and 0.19 for White speakers
- 0.19 for Black speakers and 0.35 for White speakers
- Zero errors for all speakers
- Which speaker sample did the 2020 study use?
- 42 White speakers and 73 Black speakers in structured interviews
- One speaker reading a single sentence
- Thousands of speakers from every country
- Only children with speech disabilities
- Why should the 2020 result not be called a universal “accent error rate”?
- WER is not a speech metric
- It measured a particular U.S. sample and race-associated comparison, not every accent or language
- It tested no speech systems
- Accent and language never affect recognition
- What did later accent-focused ASR research show?
- Every accent has identical WER
- Only microphones affect recognition
- Performance varies across tested accents, languages and model architectures
- Accent effects occur only in English
- Which terms should an auditor keep distinct?
- Accent, dialect, race, language proficiency and recording conditions
- Speaker identity and microphone brand only
- Accuracy and spelling only
- Gender and sentence length only
- Which safeguard is appropriate when an ASR transcript informs a consequential workflow?
- Reject every speaker with a strong accent
- Treat each transcript as ground truth
- Provide correction and human review for transcripts that may affect decisions
- Hide error rates from users