Automatic speech recognition (ASR) these days in general has two main components: 1) Feature extraction and 2) classification. Check https://theaisummer.com/speech-recognition/ for an introduction. In our case, the application is somewhat simpler than the general speech recognition in as far as we only want to recognize a single specific keyword or phrase such as 'Hello HAL' or 'Wake up!'. Particularly attractive for HW implementation with HW components that we are also investigating, would be (for 1) ) a particular variant of extracting the short term frequency power spectrum of the incoming sound with a 'silicon cochlea' in combination with (for 2) ) a recurrent neural network (RNN) structure. The latter would employ machine learning to learn to classify a particular keyword.
This project would investigate such an architecture and evaluate its performance for key word recognition. It is part of a bigger project for neuromorphic (computing inspired by the nervous system) speech recognition where others concentrate more on the hardware development.
Keyword Recognition Algorithm for Analog HW Implementation
Uncharacteristically for the NANO group, this project shall not investigate hardware directly, but investigate a learning algorithm with the aim of making/choosing said algorithm to be suitable for implementation as analog integrated circuit.
Published Oct. 12, 2022 9:40 AM
- Last modified Oct. 12, 2022 11:14 AM