X-Vector Speaker Embeddings Quiz

Answers and explanations appear as you take the quiz.

1. What information is a speaker-recognition x-vector designed to emphasize?
   • Only the exact words spoken as text
   • Only the recording's file format
   • A guaranteed identity proof unaffected by recording conditions
   • Speaker-related characteristics rather than a transcript of the utterance
2. Which layer turns a variable-length utterance into a single fixed-length vector in the x-vector network?
   • A convolutional image filter
   • A statistics pooling layer that computes mean and standard deviation across time
   • A dropout layer
   • A final softmax used as the embedding itself
3. What type of neural network is traditionally used to extract x-vectors?
   • A generative adversarial network
   • A decision tree ensemble
   • A Time-Delay Neural Network (TDNN)
   • A plain linear regression
4. How are two x-vectors typically compared to decide if they come from the same speaker?
   • By cosine similarity or a PLDA backend
   • By counting matching words
   • By comparing file sizes
   • By checking the audio sample rate
5. Which established representation supplied comparison baselines in the 2018 x-vector speaker-recognition study?
   • A visual image-token embedding
   • A text-only word-count vector
   • i-vectors
   • An audio file-size histogram
6. Which newer architecture improves on the classic x-vector with channel attention and attentive statistics pooling?
   • ResNet-50 for images
   • GPT-2
   • YOLO
   • ECAPA-TDNN
7. Why is the classic x-vector extracted from a hidden segment-level layer rather than used as the final training-speaker softmax output?
   • The hidden representation can be compared as an embedding instead of being limited to probabilities for the training speaker labels
   • The hidden layer contains the recording's text transcript
   • Softmax probabilities are the only place where acoustic information exists
   • Hidden layers cannot process an unseen recording
8. What does a time-delay layer contribute before statistics pooling in an x-vector extractor?
   • It stores only a single final filename
   • It combines information from neighboring acoustic frames, building temporal context
   • It assigns a new random speaker label to every frame
   • It removes all time information before any acoustic processing