Case context: A team building a car-detection system wants more training data. They record 1 hour of background car engine noise and overlay it onto thousands of clean audio clips to create synthesized noisy training examples.

Question: Based on the concept of representative synthesized training examples, what problem should the team diagnose in this strategy, and what should they do differently?

Sample answer: The team should recognize that using only 1 hour of car noise as the sole noise source makes all their synthesized examples share the same underlying noise characteristics, which a learning algorithm could learn to detect as an artifact of synthesis rather than true car noise variation. This is precisely the failure mode described in the source. To fix it, they should gather noise recordings from many different cars, locations, and conditions so the synthesized data reflects the true diversity of real-world noisy audio.

Key points:

• The 1-hour noise clip is analogous to the source's stated example of unrepresentative synthesis.
• Reusing one narrow noise source lets the algorithm distinguish synthesized from real audio.
• This causes the model to key on synthesis artifacts rather than genuine car noise features.
• The fix is to diversify the source noise recordings across many conditions.

Rubric: Full credit: identifies the 1-hour noise source as too narrow, explains the distinguishability risk, and recommends broadening the noise source. Partial credit: identifies the problem without a concrete fix.