Bài kiểm tra chất lượng tổng hợp giọng nói

Câu trả lời và giải thích xuất hiện khi bạn làm bài kiểm tra.

1. Clip A sounds natural but says “fifty” instead of “fifteen.” Which dimension failed?
   • Semantic and word correctness
   • Naturalness of timbre only
   • The ability to play audio
   • The reference sentence length
2. Why can two reported mean opinion scores be hard to compare?
   • MOS always uses an identical test set worldwide
   • A higher number guarantees correct pronunciation
   • Listeners, samples, scales and playback conditions may differ
   • A listening score determines licensing
3. Which material should a TTS test include beyond easy demo sentences?
   • Only short common words
   • Names, abbreviations, numbers, code-switching and long passages
   • Only one synthesized vowel
   • Only text with no punctuation
4. A voice sounds expressive but misses a medication amount. How should quality be judged?
   • Average naturalness with correctness and ignore the amount
   • Treat fluency as proof of accurate words
   • Only check whether the voice is recognizable
   • According to the task’s requirement, treating the amount as a serious error
5. Why check the final exported audio after compression and mixing?
   • The source text changes automatically
   • Export always repairs pronunciation
   • Delivery processing can introduce clipping or reduce intelligibility
   • Platform processing cannot affect audio
6. What can a pronunciation review reveal that a naturalness rating may hide?
   • A fluent-sounding voice may say a name or number incorrectly
   • Pronunciation and naturalness are identical
   • The listening room is irrelevant
   • Correct words guarantee human-like rhythm
7. Why define the listening context before a quality test?
   • Every task uses the same speaker and device
   • Context determines no measurable outcomes
   • A model cannot generate long speech
   • A notification, lesson and phone call have different needs and noise conditions
8. Which accessibility aid should remain with generated speech when appropriate?
   • Only a hidden synthesis prompt
   • A readable transcript of the final spoken content
   • Only a model parameter list
   • Only one MOS number