Case Study

Close the distribution gap in a wake-word detector

Case context: You are building a wake-word detector for a voice assistant. Your training set contains clean recordings made in a quiet studio using a desktop microphone. Your development set comes from household devices placed in kitchens, where recordings include echo, running water, and occasional clatter. The model performs well on training data but poorly on the development set.

Question: What data modification should you make to improve development-set performance, and why?

Sample answer: Add simulated kitchen echo and background noise to the clean training recordings. This makes the training distribution closer to the development distribution, so the model learns features that are more robust to the conditions seen on the development set.

Key points:

  • Add simulated echo and household noise to the clean training audio.
  • Make the training data resemble the development data.
  • Reduce the gap caused by the noisy kitchen recording environment.

Rubric: Learners should identify adding simulated echo and background noise to the training set and explain that it narrows the distribution gap between the clean training audio and the noisy development audio.

0

1

Updated 2026-08-12

Contributors are:

Who are from:

Tags

Machine Learning

Deep Learning

Supervised Learning

Dive into Deep Learning @ D2L

Data Science

Machine Learning Strategy

Machine Learning Yearning @ DeepLearning.AI