引導式測驗 · 中等 等級
串流語音辨識測驗
測試部分與最終成績單、串流上下文限制、RNN-T 式輸出、修訂穩定性、延遲指標和安全操作計時。
問題 1 的 8
What distinguishes streaming ASR from processing a completed recording?
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。
- What distinguishes streaming ASR from processing a completed recording?
- It waits for the full recording before output
- It guarantees partial words never change
- It operates without acoustic processing
- It emits provisional text as audio arrives
- Why might a partial transcript change after later words arrive?
- New acoustic and linguistic context changes the best hypothesis
- The previously captured audio necessarily changed
- WER is recalculated by the listener
- Later words are never relevant to earlier ambiguity
- Which metric captures instability that final WER misses?
- Only the final word count
- Only total audio file size
- How often previously shown partial words are revised
- Only network parameter count
- What does a larger future-audio lookahead usually trade?
- Less context with lower latency by definition
- More context for recognition but extra delay before output
- Automatic removal of all noise
- A guarantee that every partial token stays fixed
- How does endpointing relate to streaming recognition?
- It decides turn completion, separately from decoding the words
- It is identical to emitting each word
- It always happens before recording begins
- It measures only transcript vocabulary size
- A live assistant sees a volatile partial command. What should it avoid?
- Showing provisional text with a marker
- Measuring revision frequency
- Executing a consequential action before confirmation or finalization
- Waiting for further audio context
- Why report tail latency as well as a median?
- The median is always the maximum
- Some users may face long delays hidden by the median
- Tail delay measures only transcript content
- WER includes response delay in its formula
- What does on-device ASR not guarantee by itself?
- That audio can be processed locally
- That network dependence may be reduced
- That it changes the data-flow design
- Perfect accuracy or privacy without suitable implementation and controls