OpenAI 耳语测验

当您参加测验时，就会出现答案和解释。

1. OpenAI 的 Whisper 的主要目的是什么？
   • 自动语音识别：转录和翻译音频
   • 从文本生成图像
   • 编写计算机代码
   • 创建 3D 动画
2. What was true of the language balance in Whisper’s training audio?
   • Every language contributed the same number of hours
   • The dataset was English-heavy, with much less audio for most languages
   • The dataset contained only English audio
   • Most data was studio-recorded speech in one language
3. What does Whisper use to condition the decoder on the requested task and language?
   • A separate camera feed
   • A hand-written transcript of the answer
   • Special tokens in the decoder prompt
   • A unique model checkpoint for every utterance
4. Whisper 本身不提供哪些功能？
   • 将其他语言的语音翻译成英语
   • 检测正在使用哪种语言
   • 向转录本添加时间戳
   • 根据提示生成逼真的视频
5. Which long-form transcription problems does OpenAI’s paper describe?
   • Repetition loops and missing words near segment edges
   • Perfect diarization of every speaker
   • Automatic correction of every proper noun
   • Guaranteed removal of background music
6. Whisper 有哪些已知的需要注意的弱点？
   • 它无法处理任何英语音频
   • 它偶尔会“产生幻觉”未说出的文字
   • 它只输出音频，不输出文本
   • 它拒绝添加标点符号
7. Approximately what share of Whisper’s training data was English audio paired with English transcripts?
   • About 18%
   • 约1%
   • About 65%
   • 约100%
8. What use does OpenAI caution against when transcription errors could have serious consequences?
   • Noncommercial experimentation with public samples
   • 向转录本添加时间戳
   • Using the model to identify a language
   • High-risk decision-making without appropriate verification