Introduction¶

Human Perception of Sound¶

Dataset¶

Building¶

• Who are the users
• What do they need
• What task are they trying to solve
• How do they interact with the system
  • Distance
  • Environment
    • Background Noise
    • Reverb
• Quality Control
  • Only keep whatever a human can understand

Industry-Standard¶

• Google Speed Commands dataset
  • Recorded as individual words, not sentences
  • 1000-4000 examples of each word

Good Characteristics of Model¶

Volume Invariance

Pre-Processing¶

What aspects of the signal should you sent to the neural network

1. Align on start point
2. Normalization of amplitude
3. Denoise
4. Convert to frequencies, using Fast Fourier transform
   1. Extract features
   2. Sliding window
5. Cut on end point

Word	Volume	Waveform	Spectrogram	MFCC
Yes	Loud
	Quiet
No	Loud
	Quiet

Mel Filterbanks¶

Comments