Question: In machine learning, explain how collecting targeted training examples can reduce a data mismatch problem. Illustrate your answer with a speech recognition system.

Sample answer: A data mismatch problem occurs when the examples used for training are noticeably different from the examples the model must handle during development or deployment. One practical response is to collect additional training data that resembles the difficult cases. For example, suppose a speech recognizer is trained mostly on clean headset recordings, but the dev set includes many short utterances captured in a noisy subway station. The model may do well on the training distribution but poorly on those subway clips. If the team deliberately records 300 additional subway-like utterances and adds them to training, the model can learn the acoustic patterns of that environment and channel condition. This narrows the gap between training and dev data and improves performance on the hard cases.

Key points:

• Data mismatch means the training and dev distributions differ.
• A good remedy is to add training examples that resemble the hard dev examples.
• In the speech recognition example, clean headset audio and noisy subway audio represent the mismatch.
• Adding targeted speech data helps the model adapt to the difficult environment.

Rubric: The answer should define data mismatch, explain why targeted data helps, and apply the idea correctly to a speech recognition scenario with different training and dev conditions.