Whisper Speech Recognition Quiz

Answers and explanations appear as you take the quiz.

1. Roughly how many hours of audio was Whisper trained on?
   • 680,000 hours
   • 6,800 hours
   • 68,000 hours
   • 6.8 million hours
2. What intermediate representation does Whisper compute from raw audio before the encoder?
   • A raw waveform tensor
   • A log-Mel spectrogram
   • MFCC delta features only
   • A one-hot phoneme vector
3. How does a single Whisper model perform multiple tasks like transcription and translation?
   • It loads a different model per task
   • It runs separate fine-tuned heads in parallel
   • Special prompt tokens specify the task during decoding
   • Users edit the spectrogram manually
4. What overall neural architecture does Whisper use?
   • A convolutional-only network
   • A recurrent LSTM stack
   • A diffusion model
   • A Transformer encoder-decoder
5. Why is Whisper considered especially robust compared to earlier ASR systems?
   • It was trained on diverse, noisy real-world web audio
   • It only works on studio-clean audio
   • It uses no machine learning
   • It requires a custom microphone
6. Under what license were Whisper's model weights released?
   • Proprietary, API-only
   • MIT (open source)
   • GPL with no commercial use
   • Creative Commons non-derivative
7. Approximately how long is each audio segment Whisper converts to a log-Mel spectrogram?
   • 3 seconds
   • 3 minutes
   • 30 seconds
   • 30 minutes
8. What does Whisper’s English-translation task produce from non-English speech?
   • A translated recording in the speaker’s voice
   • Only a transcript in the original language
   • A list of detected speakers
   • An English text translation