Noise-Conditioned State Space Model

Voice AI on a
$5 Chip

NC-SSM-20K achieves 96.4% keyword spotting accuracy in 10.4ms on a Cortex-M7 MCU — matching DS-CNN-S accuracy with 5x lower latency and 10x fewer MACs.

20K
Parameters
10.4ms
Latency
2.44M
MACs
96.4%
Accuracy
$5
Chip Cost
⚡

Ultra-Low Latency

7.1ms inference on Cortex-M7 @ 480MHz. Stateful SSM hidden state enables progressive classification without rebuffering.

🔋

Tiny Footprint

7,443 parameters, 720 bytes hidden state. INT8 quantized model fits in 7.3KB. Runs on $1.50 FPGA or $3.50 MCU.

🎤

Noise-Conditioned

Learned noise embedding (sigma) adapts the model to real-world SNR conditions. No manual noise calibration needed.

🚀

Streaming Native

SSM processes audio frame-by-frame with stateful hidden state. Short keywords detected in ~350ms vs 1053ms for CNN.

🔒

Patented (US/KR)

Noise-conditioned SSM architecture protected under US and KR patent applications. Royalty licensing available.

📦

SDK Ready

pip install nano-ssm. PyTorch training + C SDK for Cortex-M deployment. ONNX/TFLite export in one line.

Live Keyword Detection

Real-time voice command recognition running NC-SSM inference

Tap to Start
Browser microphone required
Yes
No
Up
Down
Left
Right
On
Off
Stop
Go
[system] NC-SSM demo ready. Click microphone to begin.

Performance Benchmark

NC-SSM vs industry-standard CNN models on Cortex-M7 @ 480MHz

OURS
NC-SSM
SSM Family
Parameters7,443
MACs (Model)0.86M
Latency7.1ms
Accuracy95.3%
BC-ResNet-1
CNN Family
Parameters7,464
MACs (Model)4.99M
Latency12.8ms
Accuracy95.0%
DS-CNN-S
CNN Family
Parameters23,756
MACs (Model)24.3M
Latency53.5ms
Accuracy96.4%
ModelParams (K)Full MACs (M)Model MACs (M)Latency (ms)Clean Acc (%)Efficiency
BC-ResNet-17.56.154.9912.895.015.4
DS-CNN-S23.825.6724.3253.596.43.8
NC-SSM7.43.420.867.195.327.8
NC-SSM-Large10.23.801.247.995.625.2
NC-SSM-15K15.84.481.929.396.221.5
NC-SSM-20K20.05.002.4410.496.419.3

Accuracy vs Parameters

Latency vs MACs

Streaming Advantage

Why SSM architecture fundamentally outperforms CNN in real-time voice AI

SSM

NC-SSM Streaming

Stateful Sequential Processing
Mic → VAD → Onset Detection → Progressive Classify
  +300ms → 1st attempt (short words)
  +500ms → 2nd attempt (medium)
  +750ms → 3rd attempt (confirm)
confidence > threshold → DETECTED!
⚡
Stateful hidden state — ht accumulates temporal context naturally.
🎯
Progressive classification — Short words at 300ms, long at 500ms.
🔋
Zero feature map memory — Only 720 bytes hidden state.
💡
Energy-aware — 0 MACs during silence. 10x battery life.

CNN Streaming (DS-CNN-S)

Stateless Window Processing
Mic → Buffer 1.0s → Full Inference → Result

Every 1s: re-buffer → re-compute entire window
No state carried between windows
Short words may be split across windows
🔄
Stateless — Must recompute 24.3M MACs every window.
⏳
Fixed 1s buffer — Minimum latency = 1000ms.
💾
196 KB feature maps — Cannot fit on low-cost FPGA.
🔋
Always computing — Burns MACs on silence.

Detection Timeline: "Go" (200ms word)

NC-SSM
onset → classify → DETECTED
~350ms
DS-CNN-S
buffer 1000ms ... full inference 53ms ... DETECTED
~1053ms
NC-SSM detects 3x faster for short keywords
~350ms
Short Word Detection
720 Bytes
Hidden State Memory
0 MACs
During Silence
10x
Battery Life

Business Model

Multiple revenue streams from a single core technology

🔒

IP Licensing

US + KR patents on noise-conditioned SSM architecture. Per-chip royalty model for semiconductor companies and OEMs.

$0.01-0.05 / chip
📦

Nano AI SDK

pip install nano-ssm. PyTorch training, C SDK for edge deployment, ONNX export. Community free, Pro $500/mo.

SaaS / B2B
🎤

Custom Wake Word

Train custom keywords for enterprise. "Hey Samsung", "OK LG" on NC-SSM with full edge deployment package.

$10K-50K per project
⚡

Edge AI Module

All-in-one KWS module: STM32H7 + MEMS mic + NC-SSM firmware. BOM $3-5, sell $15-30. 60-70% margin.

Hardware Product

Hardware Module

All-in-one edge AI voice module for mass production

NC-SSM
Edge Voice Module v1.0
STM32H743 + MEMS Mic
25mm x 25mm x 5mm
⚙

STM32H743 (Cortex-M7)

480MHz, 1MB RAM, 2MB Flash, FPU

🎤

ICS-43434 MEMS Microphone

I2S digital output, 65dB SNR

⚡

Ultra-Low Power

Active: 100mW | Listen: 15mW | Sleep: 0.1mW

🔌

Interface

UART/SPI/I2C, GPIO wake, 3.3V supply

🚀

NC-SSM Inference

7.1ms, 7,443 params, INT8 (7.3KB)

Bill of Materials (Unit Cost @ 10K qty)

$3.50
STM32H743
$0.80
MEMS Mic
$0.50
Passives + PCB
$4.80
Total BOM

Why NC-SSM Wins

7.5x
Faster than CNN
5.6x
Fewer Model MACs
$5
Target Chip Cost
0.1W
Power Consumption