Audio Embeddings and Representation Learning Quiz

Answers and explanations appear as you take the quiz.

1. What is an audio embedding?
   • A fixed-length vector of numbers representing a sound's meaningful features
   • A compressed MP3 version of the audio file
   • A written transcript of spoken words
   • A volume level measured in decibels
2. Why is self-supervised learning so important for audio embeddings?
   • It requires audio to be recorded in a studio
   • It learns useful representations from unlabeled audio, reducing the need for human labels
   • It only works on music, not speech
   • It guarantees perfect transcription accuracy
3. How does Wav2Vec 2.0 learn during pretraining?
   • By translating audio into French
   • By memorizing every training file exactly
   • By masking spans of audio and predicting the correct quantized unit among distractors
   • By measuring the file size of each clip
4. What does a model like CLAP align in a shared embedding space?
   • Two different video frames
   • Pixels and color histograms
   • Battery level and CPU usage
   • Audio and matching text descriptions
5. Before feeding audio to a neural network, what common transformation is often applied?
   • Converting it to a spectrogram or learned filter representation
   • Printing it as sheet music
   • Deleting all frequencies above human hearing permanently
   • Converting it to a JPEG image of the album cover
6. Which real-world task directly relies on comparing audio embeddings?
   • Calculating a song's purchase price
   • Speaker diarization, figuring out who spoke when
   • Choosing album cover artwork
   • Setting the screen brightness
7. During contrastive audio-text training, what happens to a matching audio clip and its text description?
   • They are converted into the same raw waveform
   • The text embedding is discarded
   • Their embeddings are pulled closer while mismatched pairs are pushed apart
   • The audio is assigned a speaker name as a class label
8. Why can one self-supervised audio embedding be reused across several downstream tasks?
   • It stores the original waveform without compression
   • It contains a manually written label for every possible sound
   • It works only for one speaker and one recording
   • It captures general acoustic structure that can support tasks with little extra labeled data