Voice + wearables
Murmur
An open intelligence layer for Omi and other AI wearables. Own the recordings, choose the models, preserve memory, and securely direct a computer or server from anywhere.
View repositoryWega Labs · Open interface infrastructure
New ways to tell computers what we mean.
October is the spatial canvas. Murmur opens voice wearables. Fovea turns gaze and gesture into control. Noema translates deliberate biosignals into text. Together, they move computing beyond keyboard and mouse.
The product + the open layerOctober · Murmur · Fovea · Noema
The architecture / 01
One canvas. An open interface layer.
October is the shared canvas containing the agents, computers, machines, and outputs a person needs to supervise.
Murmur opens voice wearables and remote action. Fovea combines gaze with hand gestures. Noema develops deliberate biosignals into text and commands. Each project exposes a reusable interface instead of locking the modality to one product.
Each project works alone. Together, they give October new ways to understand intent.
Open source programme
Three projects. One human control layer.
Useful on their own, interoperable by design, and open under Apache 2.0.
Gaze + gesture
Fovea
An open visual-input engine that turns ordinary cameras into control. Look to target, select with intent, and use a hand gesture to move, drag, scroll, or manipulate.
View repositoryBiosignals + text
Noema
Open infrastructure for translating deliberate EEG, EMG, EOG, and motion signals into confidence-aware text and commands—without pretending to read unrestricted thought.
View repositoryThe protocols and reference foundations stay open. Wega builds the most capable products, models, and managed experiences on top.
Authoring and selection / 02
Two different information costs.
Authoring needs bandwidth
Writing a sentence means producing every bit. A sentence carries about fifty bits.
Selection needs less
Choosing among three candidates costs 1.6 bits. The model provides the candidates.
Supervision is selection
Reviewing, approving, and redirecting agents are selection tasks.
Where this matters
When keyboard + mouse are the wrong interface.
Robotics and hardware labs
Direct tests and agents while standing beside a robot, vehicle, or electronics rig.
Factories and automotive teams
Supervise several machines while moving through the environment or working with both hands.
Accessibility and hands-free work
Choose the control channel that fits the person and the situation—not the desk.
Plate II / 03
Bits per decision.
Yes / no
1.0bits
One of three replies
1.6bits
One of five predicted words
2.3bits
The top three rows are selection. The bottom row is authoring.
Human budget1–2 bits / second
Stage 01 · Live / 04
One instruction in fourteen bits.
Four candidates per step. Seven steps. One instruction out of 16,384.
Stage 01 · Live
Compose an instruction with fourteen bits.
Four choices, seven steps, two bits each. This stage tests the interaction before the sensor.
Apollo · agent
Variants A and B are ready. What should happen next?
Select a candidate to begin.
Selections0/7
Bits emitted0
Bits in output—
Supplied by model—
Elapsed0.0s
Effective wpm—
Method · log₂(4) = 2 bits per choice. Output uses Shannon’s ≈1.1 bits per character. Model-supplied bits equal output bits minus emitted bits. Time includes reading.
The baseline / 05
Compare it with typing.
Type the sample to measure the authoring cost.
The authoring baseline
Typing cost
Type the line at your normal speed. The text stays on this page.
Elapsed0.0s
Words / min—
Bit / s · naive—
Bit / s · entropy—
Typing uses ten fingers. A single deliberate signal carries one or two bits per second.
Method · bit/s = characters per second × bits per character. Naive: log₂(27) ≈ 4.75. Entropy: Shannon’s ≈1.1. Rungs are typical estimates, not measurements.
Our thesis / 06
Attention, not gesture.
Code-modulated visual evoked potentials, or c-VEP, assign a visual code to each target. Attention to a target produces a matching response in the visual cortex.
Compare the recorded signal with each code to identify the selected target.
Attention is the input.
Information rate / 07
Compare non-invasive inputs.
Motor imagery
~15 bits/min
Imagined movement, extensive per-user training
Blink / gesture selection
~20–40 bits/min
Visible voluntary movement, fatigues quickly
c-VEPour direction
100–260 bits/min
Occipital electrodes, precise stimulus timing
Published c-VEP results report 0.77–0.84 accuracy across subjects and 0.98 for the best performer. These are literature results, not ours.
What we build
It all builds October.
October is the software we build and ship — a workspace for supervising agents, computers, and machines. Murmur, Fovea, Noema, model-ranked selection, and attention research expand how you control it.
If an engineer is beside a robot, car, electronics rig, or factory machine, walking back to a laptop to type “Agent 4, rerun this test” is absurd. Speak from across the room, look at the agent to focus it, or use a gesture to control the canvas.
Wega Labs is the lab. October is the flagship product. Murmur, Fovea, and Noema are open infrastructure that any product can use.
For companies, October becomes the interface layer that lets people supervise complex agent and machine operations without sitting in front of one keyboard.
Product path
Now, next, later.
October desktop
A spatial canvas for agents, screens, terminals, and previews—with Otto voice control.
Three interface foundations
Murmur, Fovea, and Noema are public Apache-2.0 projects for voice wearables, gaze + gesture, and biosignal-to-text.
Attention itself
Attend to a canvas object to focus it, then choose an action through gaze, gesture, biosignals, or model-ranked selection.
Artifact control / 08
Muscle can dominate EEG results.
The artifactual component of the EEG signal is significantly more informative than brain activity with respect to classification accuracy — consistent across different feature extraction methods and classification pipelines.Artifacts in EEG-Based BCI Therapies: Friend or Foe? · Sensors 22(1):96
Jaw, eye, and neck activity can be more predictive than brain signal. Many BCI studies do not report artifact controls.
Artifacts can help a communication interface. They cannot support a claim about brain state.
We test every neural result against a muscular explanation.
Programme / 09
Stages and gates.
Each stage has a public pass condition. Stages 01 and 02 test interaction and display timing before neural hardware.
Selection interaction, measured end to end
A laptop camera detects deliberate blinks. Natural blinks last 100–150 ms; deliberate blinks exceed 400 ms. Each deliberate blink commits a model-ranked candidate.
The interface reports bits, selections, and effective words per minute.
Gate
95% detection with under one false positive per minute, across three faces in two lighting conditions.
Stimulus timing
A photodiode measures when each flash reaches the screen. A 60 Hz display updates every 16.7 ms and can add frame jitter.
This separates display error from biological variance.
Gate
Recover a known modulation sequence from the photodiode alone, with no participant involved, plus a published jitter distribution for our display.
Attention-driven selection on a live canvas
Occipital electrodes record responses to coded canvas objects. Each object uses a circular shift of one base sequence.
Gate
80% accuracy across four targets within two seconds. The programme begins with four targets; published forty-target results remain literature references, not the Stage 03 baseline.
More targets, less calibration, better ergonomics
Increase target count, reduce calibration, and improve the visual design.
Scheduling starts after Stage 03 passes.
Gate
Defined after Stage 03 clears its gate.
Research standard / 10
How we report results.
We describe selection as selection.
We name the signal, input method, and model contribution.
Every accuracy figure includes information rate.
We report target count and time per selection beside accuracy.
Artifact controls ship with every neural result.
We test every neural result against a muscular explanation. Mixed signals remain labelled as mixed.
Published results stay attributed.
Prior results keep their citation, apparatus, participants, and conditions.
Every stage stays public.
We publish each gate, result, and dataset.
Join early access.
Test the selection interface before public release.