DeepSpeech Architecture Quiz

Answers and explanations appear as you take the quiz.

1. What loss function allows DeepSpeech to train without frame-level alignment between audio and text?
   • Mean squared error
   • Cross-entropy over fixed-length windows
   • Connectionist Temporal Classification (CTC)
   • Triplet loss
2. What was the main architectural philosophy that DeepSpeech popularized?
   • End-to-end training of a single neural network from audio features to text
   • Splitting recognition into many hand-engineered, separate components
   • Using only rule-based phonetic dictionaries
   • Recognizing speech purely with template matching
3. Why does DeepSpeech use a bidirectional recurrent layer?
   • To compress the audio file size
   • To give each prediction context from both past and future frames
   • To translate between languages
   • To remove the need for any training data
4. What kind of input features does DeepSpeech typically operate on?
   • Raw pixel data
   • Pre-written text transcripts
   • Compressed MP3 metadata tags
   • Spectrogram or MFCC features computed over short audio frames
5. What is often added at decode time to improve DeepSpeech's word choice and spelling?
   • An external language model (e.g., an n-gram model)
   • A second microphone
   • A larger image classifier
   • A hardware encryption chip
6. What does the original DeepSpeech output layer predict at each acoustic time step?
   • A fixed speaker-identity vector only
   • A single complete translated document with no time steps
   • A probability distribution over character labels and the CTC blank
   • The recording's filename and compression format
7. How does CTC relate an acoustic-frame label path to the final DeepSpeech transcript?
   • It sorts all character labels alphabetically
   • It requires exactly one audio frame for each word
   • It keeps every repeated frame label as a separate written character
   • It collapses repetitions and removes blank symbols according to the CTC rule
8. Why does original DeepSpeech's bidirectional acoustic context matter when designing a live recognizer?
   • Bidirectional recurrence needs no future observations
   • Future-frame context requires lookahead or waiting, so it is not automatically zero-lookahead streaming
   • The network can know later audio before it is recorded
   • A character language model removes all buffering requirements