引導式測驗 · 中等 等級
ASR 测验中的语境偏差和自定义词汇
測試罕見術語偏差、特定於會話的短語列表、聲學證據、錯誤插入、發音和私人詞彙處理。
問題 1 的 8
Why provide a speech recognizer with a meeting attendee list?
預覽每個問題8 問題
當您參加測驗時,就會出現答案和解釋。
- Why provide a speech recognizer with a meeting attendee list?
- It verifies who actually spoke
- It replaces the need for audio
- Relevant rare names may become easier to select when audio is ambiguous
- It requires every listed name to appear
- What risk grows when a custom term is boosted too strongly?
- The microphone records less acoustic data
- Rare names become impossible to predict
- Phrase lists automatically delete themselves
- It may appear in the transcript despite weak acoustic support
- Why prefer a short session-specific list over a huge global glossary?
- Large lists have no computation or privacy cost
- It reduces irrelevant competition and limits sensitive exposure
- A short list authenticates the speaker
- Global lists cannot contain personal names
- Which paired test reveals whether custom vocabulary causes false insertions?
- Audio with target terms and comparable audio without them
- Only recordings where every term is spoken
- Only text documents without audio
- Only the number of saved contacts
- Why might spelling alone fail to help with an unfamiliar name?
- Letters always determine an exact waveform
- Unfamiliar names cannot be pronounced
- Pronunciation may not be obvious from its letters
- Every name has the same pronunciation
- A clinic’s custom glossary improves term recall but worsens ordinary speech. What must be reported?
- Only recall on positive terms
- Only glossary length
- Only sentence fluency
- The tradeoff, including ordinary-word errors and false insertions
- How should a system handle a private contact list used for ASR context?
- Publish the list as a shared global vocabulary
- Limit access, retention and use to the authorized purpose
- Keep it forever without user choice
- Assume names are public because they aid ASR
- Which claim about a context-biased transcript is unsupported?
- A boosted name is authenticated as having been spoken
- The decoder had phrase suggestions available
- Candidates were compared using context
- Audio review is still needed