Customizable to detect any phrase - This reference design continuously searches for a key phrase utterance via a digital MEMS microphone. By updating the training dataset using deep learning frameworks such as Caffe, Tensorflow or Keras, designers can add wake word capability to a system
Optimized Power and Performance - The reference design enables a flexible system design trading off speed and power, the system can be put in low power listening for voice mode consumer as low as 0.6 mW.
Complete Reference Design - Training dataset, scripts for training using common Neural Network training tools, and NN model are provided to enable modification.
Features
- Accelerated, low-power key phrase detection at the network edge using neural network model
- iCE40 UltraPlus
- VGG8 like 8-bit CNN
- Sound samples converted to histogram and inputted to Neural Network as image
- Up to 40 evaluations per second
- 7 mW of power consumption
- Adjustable frame rate. Can be optimized between power and response time depending on system needs
This reference design is supported with 1.1 version of the Radiant Software Tool, this version of the tool can be downloaded on the Software Archive.
