Insiders Hub

small-footprint keyword spotting on the edge with Intel NCS 2

Shreekantha Nadig

Shreekantha Nadig

Bengaluru, Karnataka

1 1
  • 0 Collaborators

This project is a demonstration of on-device KWS using Intel NCS 2. We would like to use an end-to-end model trained for ASR to do small-footprint keyword spotting which is compact enough to be run on-device with Intel NCS 2. ...learn more

Project status: Concept

Artificial Intelligence

Groups
Student Developers for AI

Intel Technologies
AI DevCloud / Xeon, Intel Opt ML/DL Framework, Intel Python, OpenVINO, Movidius NCS

Code Samples [1]Links [2]

Overview / Usage

  • We would like to use an end-to-end trained ASR model to do small-footprint keyword spotting on-device using Intel NCS 2
  • The model would be built with TensorFlow and trained on Intel AI Dev Cloud
  • This is a demonstration of ASR capabilities of Intel NCS 2

Methodology / Approach

  • Feature Extraction - Either python_speech_features, kaldi wrappers or TensorFlow implementation of kaldi-like feature extraction procedure for feature extraction. log mel fbank with 23-80 coefficients
  • An unidirectional LSTM Encoder network with 1-3 layers and 128-512 hidden units
  • Joint CTC/Attention loss for training
  • Small-footprint model with deep-KWS for keyword spotting

Technologies Used

  • Intel AI Dev Cloud
  • Intel NCS 2
  • Intel TensorFlow
  • For on-device demonstration - RaspberryPi 3

Repository

https://github.com/sknadig/intel_asr

Comments (1)