Topic & Info
CAD & SoC Design Laboratory
Automatic Speech Recognition (ASR)

As the performance of GPU evolves, high-performance computing becomes possible and artificial neural network-based speech recognition is actively being studied. Currently, deep running-based speech recognition requires high-performance computation and is therefore primarily software-driven on high-performance servers. This has the disadvantages such as reducing the reliability of server connections, accessibility for various age groups, and security vulnerability.

To overcome these limitations, speech recognition hardware that works on-devices has been proposed. Convolutional Neural Network (CNN) can be used to reduce hardware cost while maintaining a high recognition accuracy of deep running. It shows specific performance for target application by learning only limited words such as IoT device operation.

In addition, the recognition rate can be greatly improved by combining speaker separation that extracts only the speaker's utterance by removing ambient noise and speaker adaptation that specializes in learning the speaker's accent.