Main Content Main Menu

On-device AI

On-Device AI, GW2 Series

Empowering Intelligence in Everyday Devices

Gwanak Analog's embedded super low-power AI platform built upon a unique AI accelerator capable of delivering higher performance

Embedded AI Platform

Harnessing the full advantage of On-Device AI requires both powerful AI models and proven silicon. GW2 Series offers a full-stack solution that works together harmoniously.

Production-Ready AI Models

Gwanak Analog trained a collection of models that are optimized to deliver production-grade performance at unparalleled efficiency. You can immediately deploy them to unlock the power of artificial intelligence in your products with full readiness.

Software Development Kit (SDK)

Gwanak Analog built software development platform to help companies of all sizes deploy optimal AI models for future applications and form factors. SDK contains advanced AI model optimization tools, a custom compiler, and a fast performance simulation.

Text-to-Speech

Gwanak Analog\'s TTS model is delicately designed to have low memory footprint and fast inference speed, without sacrificing naturalness and sound quality of speech. As a result, our model perfectly fits in a tiny embedded chip.

Multiple Wake Up Words & Keyword Detection

Keyword detection captures various spoken keywords in real time. It can be used for triggering devices with wake up words or giving multiple commands to applications. The latest models utilize deep learning architecture such as CNN or RNN.

Keyword Spotting (KWS)

Gwanak Analog\'s KWS features open vocabulary with registered words and keywords. It is highly accurate thanks to powerful & lightweight neural network-based encoder and decoder, and configurable with false positive/negative control.

Audio Event Detecion & Acousitc Scene Classification

Gwanak Analog\'s AE recognizes at what temporal instances different sounds are active within an audio signal. ASC) is a classification task associating semantic labels with audio recordings. They use latest models of ResNet and CNN/RNN

Speaker Recognition

Gwanak Analog\'s speaker recognition is a task of determining the identity of a speaker from characteristics of voices utilizing CNN-based deep learning architecture. It is 95.83% accurate with equal error rate of 2.32% with 0.1 seconds response time.

Speech Enhancement

Gwanak Analog\'s SE aims to restore clean signal from noise-corrupted signal by reducing noise. By enhancing speech quality, this process enables to people to hear more clearly and helps speech recognition system to perform better.

Beamforming & Noise Suppression

Gwanak Analog\'s multi-microphone signals are processed to a single-channel by neural beamforming to improve the suppression of noise. The noise suppression model is applied to removing the noise part of the noisy speech to restore clean speech.