VITS End-to-End Speech Synthesis Quiz

Answers and explanations appear as you take the quiz.

1. What does the acronym VITS stand for?
   • Variational Inference with adversarial learning for end-to-end Text-to-Speech
   • Voice Inference Transformer System
   • Vocoder-Integrated Time-Series synthesizer
   • Visual Input Text Synthesis
2. What is the main architectural difference between VITS and traditional TTS pipelines?
   • VITS uses two separately trained stages
   • VITS trains the acoustic model and vocoder jointly, going from text directly to waveform
   • VITS requires hand-labeled phoneme durations
   • VITS only outputs mel-spectrograms
3. How does VITS learn the alignment between text and audio frames?
   • A pre-trained external forced aligner
   • Manual annotation by linguists
   • Monotonic Alignment Search (MAS) during training
   • Random assignment refined by reinforcement learning
4. Why does VITS include a stochastic duration predictor?
   • To make every utterance identical for consistency
   • To compress the audio file size
   • To translate text into other languages
   • To produce varied, natural speaking rhythms for the same text
5. Which component pushes VITS output toward realistic-sounding audio?
   • A GAN-style discriminator using adversarial training
   • A reinforcement learning reward model
   • A rule-based signal-processing filter
   • A spell-checker module
6. Which cited system explicitly builds on VITS for zero-shot multispeaker and multilingual synthesis?
   • FastSpeech 2
   • YourTTS
   • WaveNet
   • A deterministic vocoder with no speaker conditioning
7. What does the normalizing-flow component contribute to VITS?
   • It replaces text-audio alignment with a speaker identity check
   • It guarantees identical duration for every spoken phrase
   • It increases the expressiveness of the latent prior distribution
   • It removes the waveform decoder from the model
8. Why does VITS use both variational inference and adversarial learning?
   • They make the model independent of text input
   • They guarantee every generated voice is a real speaker
   • They remove all duration variation
   • They support latent-variable modeling and realistic waveform generation