引導式測驗 · 中等 等級

裝置端語音辨識測驗

測試本地與伺服器 ASR、串流延遲、硬體和功率限制、部分雲端回退、資料流隱私和裝置測試。

相關引導路徑裝置端語音辨識
問題 1 的 8

What changes when the main ASR computation runs on device?

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。

  1. What changes when the main ASR computation runs on device?
    • No model weights are needed
    • Audio need not be sent to a remote recognizer for that local step
    • All app telemetry is automatically disabled
    • Every language becomes supported
  2. Why can local ASR still have privacy risks?
    • A local microphone never records sound
    • On-device models always upload raw audio
    • Offline use grants legal consent
    • Transcripts, sync or diagnostics may still leave the device
  3. Which metric should be measured on a target phone besides WER?
    • Latency, memory, battery and thermal behavior
    • Only a cloud server benchmark
    • Only the model parameter count
    • Only WER measured on a different phone
  4. Why may an app be partly on-device but not fully offline?
    • Local models cannot emit partial text
    • A phone has no microphone
    • Translation, sync or other features may still contact a server
    • Every function shares the ASR process
  5. How can RNN-T support a streaming recognizer?
    • It guarantees zero recognition errors
    • It can emit symbols incrementally without waiting for full audio
    • It needs a complete file before every token
    • It removes the need for acoustic encoding
  6. What may quantization trade when compressing a model?
    • Lower resource use for accuracy changes that require testing
    • Guaranteed improvement for every speaker
    • Automatic data deletion
    • A legal privacy certification
  7. A model performs well on a flagship phone. What remains to test?
    • Only flagship-device marketing specifications
    • Only a clean training-set score
    • Only the model checkpoint’s paper date
    • Other supported devices and real microphone conditions
  8. Which interface helps when local ASR mishears a rare name?
    • Treat the transcript as final and uneditable
    • Assume privacy makes it accurate
    • A correction or fallback path
    • Restart the device without showing the error