VOICE MODELS
Voice AI
Local STT, TTS, VAD, and echo control that let devices hear and speak without the cloud.
EDGE SYSTEMS / §01 — VOICE
ON-DEVICE VOICE AI
Build
voice agents
for physical AI
An on-device AI platform for building voice agents that make devices speak, listen, and reason locally.
LOCAL RESPONSE700ms
CLOUD CALLS0
DATA EGRESS0
§02 — Runs on the hardware you ship
NVIDIAAMDIntelQualcommRaspberry PiWebAssembly
§03 / Fig. 03 — Signal path
Speech, reasoning, and response run on-device.
SPEECH → REASONING → RESPONSE
700 ms round trip · 0 packets egress · 1B–35B models
700 ms round trip · 0 packets egress · 1B–35B models
FIG. 03 — Full-duplex voice loop. Mic → STT → SLM (+ memory, guardrails, tools) → TTS → speaker, entirely local. No request ever leaves the device.
§04 — WHY ON-DEVICE
Physical AI cannot wait on a cloud round trip.
THE CLOUD WAY
Mic → network → datacenter → back.
LATENCY1–2s
DATASOMEONE ELSE'S SERVERS
OFFLINENEEDS CONNECTIVITY
COSTSCALES WITH USAGE
Every utterance leaves the building. One dropped packet, one silent failure.
THE EDGEAI WAY
Mic → speaker, all on the same chip.
LATENCY700ms
DATA0 PACKETS EGRESS
OFFLINEWORKS OFFLINE
COSTFLAT · UNLIMITED
Nothing leaves the chip. Speech, reasoning, memory, tools, and response stay on-device.
FIG. 04 — Cloud vs. on-device. Measured on reference NVIDIA Jetson Orin hardware, idle system, 8B int4 model.
§05 — PRODUCTS
Everything an on-device voice agent needs.
AGENT ORCHESTRATION
Agent Harness
C++ runtime to create autonomous agents that reason, plan, and execute multi-step workflows.
EDGE MODELS
Model Catalog
Edge-ready SLMs selected, quantized, and compiled for local reasoning on embedded hardware.
1B – 35BOPTIMIZED PARAMETERS
FIG. 05 — Product overview. Voice AI, agent runtime, and edge models: one stack for devices that need to speak, listen, and reason locally.
§06 — WHY EDGEAI
Three things no competitor has all of.
01
One-stop-shop
The full agentic stack, not a single tool — orchestration, voice, and models in one runtime.
02
Ultra-portable
Hardware-agnostic by design. Build once, deploy across NVIDIA, Qualcomm, Apple, and AMD.
03
Purpose-built
Not a wrapper. Built from the ground up for real-time, on-device execution.
§07 — BY THE NUMBERS
The constraints physical AI buyers care about.
700ms
FAST · END-TO-END LATENCY
0
SECURE · DATA EGRESS
$0/min
COST-EFFECTIVE · UNLIMITED USAGE
Runs on
FIG. 06 — Telemetry baseline. Measured on reference NVIDIA Jetson Orin hardware, idle system, 8B int4 model. Latency is the full mic-to-speaker response loop.
§08 — Friday · The agent inside the machine
Your bot stopped. Ask it why.
Friday runs on your hardware — voice accessible, reads the machine’s own logs, faults, and config. Walk up, scan the code on the enclosure, and it tells you what happened: In words. Before the vendor picks up.
Friday — Unit 4417 · Cell 3 Controller
Friday
Unit 4417 · online
+ New session
This unit
Fault 217 — servo overtemp
E-stop at 04:12
Torque calibration
Overnight fault log
Shift handoff notes
On-unit · 4417
Voice · Lana — On-unit
Runs where the machine is
One optimized binary. Runs on Jetson Orin, Raspberry Pi, or whatever is already in the cabinet.
Authentication is proximity
Stand in front of the unit, scan the pairing code, and you’re in. Nothing to provision, nothing to revoke when someone leaves.
The network is optional
Local models, local tools, local memory. Works on a floor with no uplink and no cell service — and keeps working when connectivity doesn’t.
If you can see the unit, you can talk to it
Early access
§09 — IN THE FIELD
One platform. Every physical frontier.
FIG. 01
Robotics
Talk to your robots. Spoken commands, understood on-device.
FIG. 02
Automotive
An in-cabin co-pilot that works with no signal.
FIG. 03
Aerospace
A mission voice agent where the radio can't reach.
FIG. 04
Defense
Private voice AI in comms-denied environments.
§10 — From the lab
Notes from the team.
Jul 7, 2026 · voice-ai
Barge-in and interruption handling for on-device voice agents
Barge-in is the hardest part of a voice agent to get right: hearing yourself over your own TTS, deciding an interruption is real, and stopping in a couple of audio frames. Why the DIY Whisper + llama.cpp + Piper stack breaks it, and how an on-device C++ loop fixes it.
Read→Jul 6, 2026 · voice-ai
Time-to-first-token: why voice agents live or die on latency
A latency budget for real-time voice agents: the 200ms human turn-taking rule, where cloud pipelines spend 600ms–1.7s, and how on-device buys it back.
Read→Jun 22, 2026 · jetson
How to build an offline voice assistant on NVIDIA Jetson Orin
A builder's guide to running a fully offline voice agent on Jetson Orin — local STT, an SLM, and TTS, with real latency numbers and where the DIY Whisper + llama.cpp + Piper stack breaks down.
Read→FIG. 08 — Engineering notes, benchmarks, and field reports.
LIVE · ON-DEVICE
Want to see a robot hold a conversation without WiFi?
Book a 30-minute technical demo. See it run on your hardware.