Conv-TasNet Time-Domain Separation Quiz

Answers and explanations appear as you take the quiz.

1. What is the defining feature of Conv-TasNet compared to earlier separation systems?
   • It operates directly on the raw waveform instead of a fixed spectrogram
   • It requires no training data
   • It only works on music, not speech
   • It converts audio to MIDI first
2. What kind of network does Conv-TasNet use as its separation module?
   • A single fully connected layer
   • A Temporal Convolutional Network (stacked dilated 1D convolutions)
   • A decision tree ensemble
   • A diffusion model
3. What replaces the traditional Short-Time Fourier Transform in Conv-TasNet?
   • A wavelet transform
   • A Hann window
   • A learned 1D-convolution encoder
   • A Mel filterbank
4. Which loss function is central to training Conv-TasNet?
   • Cross-entropy on word labels
   • Mean absolute pixel error
   • Contrastive image loss
   • Scale-invariant SI-SNR with permutation-invariant training
5. What did Conv-TasNet’s paper report about ideal time-frequency magnitude masks?
   • It surpassed several ideal magnitude masks in two-speaker separation evaluations
   • It eliminated the need for microphones
   • It only worked on one identical speaker
   • It used ideal masks as its encoder
6. What does Conv-TasNet’s learned encoder produce for the separator?
   • A transcript of each speaker
   • A waveform representation optimized for speaker separation
   • A fixed mel image
   • A list of microphone locations
7. What type of speech-separation task did the Conv-TasNet paper address?
   • Text-to-speech from a written prompt
   • Speaker identification from video frames
   • A single-channel mixture containing multiple speakers
   • Multichannel music transcription with MIDI output
8. What allows Conv-TasNet’s TCN to model long-term speech dependencies?
   • A database lookup for every audio sample
   • A text decoder
   • A fixed Fourier transform alone
   • Stacked one-dimensional dilated convolutional blocks