On-Device Voice AIKWS · Wake Word · STT
28 nmSeM-V1 silicon planned 2027.03.31
3–5 TOPS/WCurrent internal range · validation ongoing
7Korean patent applications pending

CLOUD-STREAMED VOICE PIPELINE

REMOTE

Continuous audio uplink
Network-dependent inference
Cloud round-trip latency

Cloud-Streamed vs.
On-Device Inference

ON-DEVICE VOICE PIPELINE

LOCAL

KWS · Wake Word · STT
Local inference execution
No continuous audio uplink

ARCHITECTURE · ON-DEVICE VOICE INFERENCE

On-Device Voice Inference Pipeline

SeM-V1 is designed to execute KWS, wake-word detection, and STT workloads on device, so supported functions do not require continuous raw-audio streaming to the cloud.

  • Always-On Voice MonitoringExecution path designed for continuous keyword and wake-word detection
  • Energy-Efficient InferenceClustered compute and local SRAM reduce data movement
  • On-Device ExecutionKWS, wake word, and STT run on the local device path
Appliance concept showing voice input processed locally by SeM-V1 before the device responds

PRODUCT DETAIL · VUI EDGE AI

SeM-V1

Always-On Voice Agent SoC

A low-power on-device Voice AI SoC development system for KWS, wake-word detection, and STT workloads, including compiler, runtime, profiling, and evaluation hardware.

  • Keyword Spotting (KWS)
  • Wake-Word Detection
  • Speech-to-Text (STT)
  • Compiler · Runtime · Profiling
SeM-V1 package concept with ZedBoard and Trenz evaluation platforms

Model → Software → NPU → Silicon

  1. Voice Model

    Profile operators, tensor shapes, data reuse, and latency

  2. Compiler & Runtime

    Partition graphs; map dataflow and memory

  3. Clustered NPU

    Couple compute clusters with local SRAM

  4. 28 nm Silicon

    Samsung Foundry 28 nm physical implementation

Always-On Voice Monitoring

Execution path designed for continuous keyword and wake-word detection

Energy-Efficient Inference

Clustered compute and local SRAM reduce data movement

On-Device Execution

KWS, wake word, and STT run on the local device path

PRODUCT LINES

Choose a product line.

01 / 05

SEM PRODUCT · IN DEVELOPMENT

SeM-V1 Development Status

  • SeM-V1Physical silicon planned 2027.03.31
  • 3–5 TOPS/WCurrent internal range
  • 3 Voice WorkloadsKWS · Wake Word · STT
  • Compiler · Runtime · ProfilingDeployment software · silicon · evaluation hardware
  • 7 Patent ApplicationsKorean applications · pending

The image shows the SeM-V1 package concept and evaluation platforms. It is not a photograph of finished SeM-V1 silicon.

SeM-V1 package concept with ZedBoard and Trenz evaluation platforms
SeM-V1 package concept · AI accelerator evaluation hardware

Voice Model Evaluation

Assess SeM-V1 mapping feasibility from operator mix, tensor shapes, memory footprint, and latency targets.

Request Model Evaluation

SoC System Integration

Define the NPU IP, silicon, and deployment-software scope from target-product interfaces and power and area budgets.

Discuss System Integration

CONCEPT FILM · 23 SEC

SeM-V1 Always-On Voice Agent SoC

The supplied CES concept film was optimized for the web. It is a technology visualization, not live product footage.