Question: A team is building a speech recognition system, but it has only a small set of labeled recordings. Explain how handcrafted audio components such as MFCC features and phoneme-based representations can influence what the model must learn from data. In your answer, describe both how these components affect the amount of training data needed and how they relate to the knowledge learned directly from examples.

Sample answer: Handcrafted audio features can make the learning task easier by removing some irrelevant variation, such as differences in speaker pitch or similar surface details. MFCCs are designed to capture useful characteristics of speech in a compact form, while phoneme representations help express the basic sound units of language. These components provide prior structure that works alongside what the model learns from training examples, so the system can reach useful performance with fewer labeled recordings.

Key points:

• Makes the learning task easier
• MFCCs reduce sensitivity to irrelevant variation such as pitch
• Phoneme representations capture basic speech units
• Adds prior structure to what the model learns from data
• Helps the system perform well with less labeled data

Rubric: A strong response will state that handcrafted components make the learning problem simpler and add useful prior structure, which reduces the amount of labeled data required. It should mention that MFCCs help filter out irrelevant variation such as pitch and that phoneme representations describe basic speech sounds.