引導式測驗 · 中等 等級

Meta MMS 與 Massively Multilingual Speech 測驗

測試 MMS 任務特定的語言覆蓋範圍、ASR 與識別和合成、朗讀-語音域轉換以及本地評估。

問題 1 的 8

Why do MMS papers report different language counts for ASR and language identification?

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. Why do MMS papers report different language counts for ASR and language identification?
    • Language ID is the same task as speech synthesis
    • All MMS checkpoints have identical language lists
    • They are different tasks with different model coverage
    • The counts refer to countries rather than languages
  2. What does an MMS language-support listing not guarantee?
    • Equal accuracy for every dialect and real-world task
    • That a specific language is listed in a model scope
    • That a task-specific checkpoint may exist
    • That the paper reports multilingual training
  3. Why test ordinary conversations if training used public readings?
    • Conversation cannot be transcribed by any model
    • Read speech contains no audio variability
    • Language ID cannot operate on readings
    • Read style and vocabulary may differ from daily speech
  4. Which evaluation partner is especially useful for an underrepresented language?
    • Only a vendor outside the speech community
    • Native speakers who can assess local varieties and reference transcripts
    • Only translated English benchmark text
    • Only a single global average score
  5. A team needs written text from audio. Which MMS task should it inspect?
    • 文字轉語音合成
    • Language identification without transcription
    • 自動語音識別
    • A language embedding without a decoder
  6. A language-ID model predicts a language for a clip. What should it not be used to infer?
    • The speaker’s nationality or ethnicity
    • A candidate audio language label
    • A possible route to ASR
    • A reason to check confusion with related languages
  7. Why verify a specific MMS checkpoint and license before deployment?
    • Every MMS artifact is interchangeable
    • A paper count replaces license review
    • All checkpoints use identical scripts
    • Project family names do not define one model’s task, rights or scope
  8. What would make a broad coverage claim more meaningful to users?
    • Only the total language count
    • Per-language tests on representative speech with a correction path
    • Only one clean read-speech clip
    • Only an unverified vendor headline