RNN-Transducer Models Quiz

Answers and explanations appear as you take the quiz.

1. Which limitation of CTC does the RNN-Transducer specifically address?
   • CTC's conditional-independence assumption, which prevents modeling output-token dependencies
   • CTC cannot process audio at all
   • CTC requires GPUs to run
   • CTC only works for images
2. What are the three main components of an RNN-Transducer?
   • Encoder, decoder, and discriminator
   • Encoder (transcription network), prediction network, and joint network
   • Tokenizer, embedder, and softmax only
   • Two attention heads and a router
3. What role does the prediction network play in an RNN-T?
   • It compresses the audio waveform
   • It measures microphone volume
   • It acts like a language model, conditioning on previously emitted tokens
   • It selects the sampling rate
4. Why does RNN-T support low-latency streaming so naturally?
   • It processes the entire utterance at once before emitting anything
   • It ignores time order entirely
   • It only works on pre-recorded files
   • Its alignment is monotonic and left-to-right, advancing through audio with blanks
5. In RNN-T decoding, what does emitting a blank token do?
   • Advances to the next audio time step
   • Ends the entire transcription
   • Deletes the previous word
   • Switches to a different language
6. Which encoder type increasingly replaces LSTMs in modern production RNN-T systems?
   • Decision trees
   • Conformer encoders
   • Plain bag-of-words
   • Hidden Markov Models
7. What two histories condition the RNN-T joint network’s output distribution?
   • Only the future transcript and a speaker photo
   • Only the last acoustic frame, without output context
   • The encoder representation of acoustic input and the prediction-network representation of prior output labels
   • The optimizer state and learning rate
8. What alignment behavior does an RNN-T allow before advancing to the next audio frame?
   • It must emit exactly one label for every frame
   • It advances time only after a word-final token
   • It generates all output before reading any audio
   • It can emit zero or more nonblank labels at the current frame, then a blank advances time