Audio Feature Extraction with Librosa Quiz

Answers and explanations appear as you take the quiz.

1. Which feature family summarizes spectral-envelope information using a mel filter bank and cosine transform?
   • Chroma vectors
   • Onset-strength envelopes
   • MFCCs
   • Raw STFT phase angles
2. What does a chroma representation commonly aggregate?
   • Energy in each distinct absolute-frequency bin without octave folding
   • Energy by pitch class
   • The time intervals between successive amplitude peaks
   • Cepstral coefficients describing the spectral envelope
3. Which setting controls the spacing between successive analysis frames?
   • Window length alone
   • Number of mel bands
   • Number of cepstral coefficients
   • Hop length
4. Why must a waveform be transformed when resampling?
   • Changing only the metadata does not create samples at the new rate
   • Relabeling samples alone always preserves playback pitch at any rate
   • A new sample-rate field automatically applies anti-alias filtering
   • Repeatedly copying samples is equivalent to every proper resampling method
5. Why should feature normalization statistics be fitted on training data only?
   • To guarantee that every normalized feature is Gaussian
   • To make preprocessing independent of the fitting data
   • To prevent validation or test information from influencing preprocessing
   • To permit a different learned normalization at inference without validation
6. What information is discarded when an STFT is reduced to magnitude?
   • Sampling rate
   • Phase
   • Time-frame count
   • Frequency-bin count
7. Which statement about onset strength is accurate?
   • It detects every onset at an exact boundary
   • It measures only frame loudness
   • It returns tempo-independent beat labels
   • It summarizes changes associated with likely attacks but is not a guaranteed event detector
8. Why group recordings by speaker or session before splitting overlapping windows?
   • Related windows can otherwise leak across training and validation
   • Window overlap guarantees independence once labels are shuffled
   • Grouping changes the waveform's sample rate
   • A source identifier alone removes every correlation between related windows