Research note 01November 2026

We ran the decoder on36 people's ear EEG

The questions the earbuds depend on can be asked of real recordings before a single earbud exists. Electrodes around the ear, two stories at once, and the same code the earbuds will run.

77%
of one minute decisions name the story a person was following
The learned decoder, four electrodes per ear, two separate earbuds, no calibration. The linear one: 68%. A coin flip scores 50%.
1.6points
lost by cutting the wire between the ears
Each earbud referenced to itself, against one reference shared across the head.
0min
minutes of calibration needed
A decoder trained on 35 other people matches one trained on 20 minutes of a person's own data.
78s
for the EEG alone to follow a switch of attention
Median expected switch duration with the learned decoder; 120 s with the linear one. Enough to confirm, too slow to steer. So the head steers.

Live, in your browser

Watch it decide

A real recording: three minutes of EEG from electrodes around one person's ears while two stories played, one on each side. Your browser runs it through the earbud pipeline, filtering, resampling and decoding, with the linear decoder trained on the other 35 people that has never seen this one.

No sound: the data set keeps how loud each story was from moment to moment, not the stories themselves.

What the recordings say

Longer windows,clearer answers

How often a decision names the followed story, by how much EEG it may use, from one second to ten minutes. Four electrodes per ear on two separate earbuds, against all eighteen around both ears on a wire, and one earbud alone. Over five minutes, the two earbuds get it right 85% of the time.

4 per ear, wireless
All 18, wired
One earbud, 4

50%60%70%80%90%88.9%Chance for one person

125103060120300600

Decision window, seconds

Show the numbers

Decision window, seconds	4 per ear, wireless	All 18, wired	One earbud, 4	Chance for one person
1	53%	53%	52%	52%
2	54%	54%	54%	53%
5	56%	56%	55%	55%
10	58%	59%	57%	57%
20	61%	62%	59%	60%
30	63%	65%	61%	62%
60	68%	71%	66%	67%
120	74%	74%	71%
300	85%	85%	79%
600	89%	91%	88%

A few electrodesgo most of the way

One minute decisions on two separate earbuds, by electrodes per ear, each ear referenced to its own. Two per ear already reach 68%. Every electrode around both ears adds about 3 points, less than the spread between people.

50%60%70%

23468All

Electrodes per ear

Show the numbers

Electrodes per ear	Decisions right
2	68%
3	69%
4	68%
6	71%
8	71%
All	71%

Where the four sit

Left ear

Right ear

Positions on a cEEGrid, the arc of electrodes around each ear used in the recordings, seen from the side with the face toward the outer edge. Filled: the four per ear the decoder leaned on most.

No calibrationto start

One minute decisions by how much of a person's own data trained the decoder. A few minutes of it does worse than a decoder trained on other people, which needs none, and 20 minutes only edges past it.

Trained on 35 other people, none of theirs
Trained on their data only

50%60%70%68.1%70.5%

1251020

Minutes of the person's own data

Show the numbers

Minutes of the person's own data	Trained on their data only	Trained on 35 other people, none of theirs
1	54%	68%
2	59%	68%
5	61%	68%
10	66%	68%
20	70%	68%

Most people,not everyone

One dot per person, one minute decisions over 30 minutes, 4 per ear, wireless, no calibration. 25 of 36 clear the bar a single person needs to beat chance, and 4 sit at or below a coin flip.

Beats chance for one person

40%50%60%70%80%90%

Ringed: the three recordings in the replay

Show the numbers

#	Decisions right
1	70%
2	57%
3	80%
4	77%
5	70%
6	77%
7	53%
8	70%
9	67%
10	60%
11	80%
12	83%
13	63%
14	67%
15	77%
16	53%
17	60%
18	70%
19	77%
20	73%
21	73%
22	77%
23	87%
24	67%
25	47%
26	47%
27	77%
28	77%
29	70%
30	73%
31	43%
32	87%
33	80%
34	43%
35	67%
36	53%

A learned decoder,the same 36 people

The linear decoder rebuilds a story's loudness from the EEG through one fixed filter. The learned one trains two small networks together, one reading the electrodes and one reading each story, until the EEG's reading moves with the followed story's. Trained only on other people, scaled from the wearer's first two minutes, each window judged on its own samples. One minute decisions go from 68% to 77%.

Learned decoder
Linear decoder
Learned, stories out of step

50%60%70%80%90%94.4%Chance for one person

125103060120300600

Decision window, seconds

The check: the same stories, played out of step with the EEG. A network that had learned to tell the two voices apart and read something else in the EEG, such as where the eyes were turned, would keep scoring. This one falls to 49%, a coin flip, so what it follows is the speech.

Three different splits of the 36 people into training and testing, three networks each: one minute decisions from 76% to 77% per split.

Show the numbers

Decision window, seconds	Learned decoder	Linear decoder	Learned, stories out of step	Chance for one person
1	54%	53%	50%	52%
2	56%	54%	50%	53%
5	59%	56%	50%	55%
10	62%	58%	50%	57%
20	66%	61%	50%	60%
30	69%	63%	49%	62%
60	77%	68%	49%	67%
120	82%	74%	48%
300	89%	85%	47%
600	94%	89%	46%

What a decisionis worth

On the phone, a ten second decision moves the estimate of who is being listened to by how clearly it was won. How far was set by hand. Measured on people the decoders never saw, it was three times too far for the linear decoder: when the fusion was 91% sure, those decisions were right 67% of the time. Each decoder now ships with the weight its decisions earned, so a decision moves the estimate exactly as far as it deserves.

Learned, weight fitted (15.9)
Linear, weight fitted (3.0)
Linear, weight set by hand (10.0)

50%60%70%80%90%100%Sure as often as right

50%60%70%80%90%100%

How sure the fusion was

Show the numbers

How sure the fusion was	Decisions right	Decoder and weight
52%	52%	Learned, weight fitted (15.9)
56%	56%	Learned, weight fitted (15.9)
60%	62%	Learned, weight fitted (15.9)
66%	66%	Learned, weight fitted (15.9)
75%	74%	Learned, weight fitted (15.9)
51%	53%	Linear, weight fitted (3.0)
54%	53%	Linear, weight fitted (3.0)
57%	55%	Linear, weight fitted (3.0)
61%	62%	Linear, weight fitted (3.0)
67%	67%	Linear, weight fitted (3.0)
54%	53%	Linear, weight set by hand (10.0)
63%	53%	Linear, weight set by hand (10.0)
72%	55%	Linear, weight set by hand (10.0)
81%	62%	Linear, weight set by hand (10.0)
91%	67%	Linear, weight set by hand (10.0)

Weak alone,strong together

A ten second decision is right 62% of the time with the learned decoder: too weak to act on, so the phone acts on none of them alone. It keeps a belief about who is being listened to, adds every decision by what it is worth, and allows that attention may have moved since the last one. After the first minute the belief names the followed story 86% of the time, and after a switch it comes over in a median 60 seconds.

62%
of single ten second decisions right
86%
of the time the belief names the followed story
60s
to come over after a switch

Learned decoder
Linear decoder

50%60%70%80%90%100%

060120180240300

Seconds to come over after a switch

How readily the belief lets go is one number, the chance per decision that attention moved. Higher follows a switch sooner and wavers more. The phone uses 2%, ringed.

Show the numbers

Chance attention moved, per decision	Learned, names the story	Learned, to come over	Linear, names the story	Linear, to come over
0%	89%	260 s	85%	220 s
0.5%	88%	100 s	84%	140 s
1%	87%	80 s	83%	95 s
2%	86%	60 s	80%	70 s
5%	82%	40 s	76%	30 s
10%	77%	20 s	71%	20 s
20%	71%	20 s	65%	10 s
50%	62%	10 s	58%	10 s

Try a design

Pick an earbud,see what the data says

Every configuration here was run on the same 36 people and held out the same way, with the linear decoder. Choose how many electrodes each earbud carries, whether the two share a wire, and whether the decoder learns the wearer.

Electrodes per ear

Form

Decoder

Decisions right at 2 s: 54.2%
Decisions right at 10 s: 58%
Decisions right at 60 s: 68.1%
To follow a switch: 120 s
People who beat chance on their own: 25 of 36

What it decides about the earbuds

The head steers.The EEG confirms.

01
True wireless works
Two earbuds share no electrical reference, so each has to measure against itself. That costs about a point and a half here. No neckband, no wire.
02
Four electrodes per ear
Two per ear reach most of what every electrode does, so four leave room for contact lost to movement and sweat. The ranking says where they should touch.
03
Wear it and go
A decoder trained on other people works from the start, and the earbud scales it to the wearer from two minutes of wear without asking anything.
04
EEG is the slow channel
Two second decisions are barely better than a coin flip; with the learned decoder one minute decisions are right about three times in four, five minute ones 85% of the time. Turning toward someone steers in a fraction of a second, a clench is the explicit command, and the EEG's running belief tells, over a minute or two, who a person keeps coming back to.

What is already built

Everything butthe hardware

The earbud's job is small on purpose: sample the electrodes, the motion sensor and the microphones, stamp the time, send frames. Everything after that is software, and it is written and tested.

Decoding pipeline
Filtering, resampling, the linear and the learned attention decoders, the running belief, head and EEG fusion, clench detection. TypeScript, matched against the research code sample for sample.
lib/sense
Wire protocol
One frame per Bluetooth notification. A C encoder for the firmware writes the exact bytes the app reads, checked in both directions.
hardware/PROTOCOL.md
Digital twin
A pair of virtual earbuds streaming real recordings, head motion and jaw clenches as frames, so the whole app runs end to end today.
lib/sense/twin.ts
Dev kit bridge
An OpenBCI board over USB, or Signalive earbuds over Bluetooth, behind the same interface as the twin.
lib/sense/transport.ts
Hardware spec
What the earbud must provide, each requirement with its reason and its test: electrodes, front end, referencing, radio, power.
hardware/SPEC.md
Research code
Everything on this page, rebuilt from the public data with three commands.
research/aad

Method and limits

The data set The study

Data

Attended speaker paradigm (cEEGrid data), OpenNeuro ds004015, released under CC0 by Holtze, Rosenkranz, Jaeger, Debener and Mirkovic (Frontiers in Neuroscience, 2022). 36 people, electrode arcs around both ears, two audiobooks at once from left and right, each person told to follow one side. Three 10 minute blocks per person in which both stories were equally loud.

Pipeline

A causal band pass from 1 to 9 Hz, resampling to 64 Hz, and a linear decoder over half a second of EEG that rebuilds the followed story's loudness over time. A window's verdict is the story whose loudness matches the rebuilt one best.

Testing

Personal decoders were tested on a 10 minute block they never saw; generic ones on a person they never saw, scaled with that person's first two minutes and no labels. The electrode ranking was chosen on the whole group, which can flatter the per ear counts slightly.

Switch time

Minimal expected switch duration (Geirnaert, Francart and Bertrand): the attended voice held in the top 35% of the gain range 80% of the time, with at least five gain steps.

Learned decoder

Two networks of one dimensional convolutions, about 16,000 weights each: one over the eight electrode channels, one over each story's loudness, trained on 5 second crops so that the followed story's embedding moves with the EEG's. Six folds by person, five more people held back to stop training, three networks per fold from different starting points, three different splits of the 36 people. The EEG is scaled from the person's first two minutes and each story by its own running spread; nothing reaches more than 0.44 s ahead. The control rotates both stories 97 s out of step with the EEG.

Trust and belief

Each decoder's ten second decisions on people it never saw fit the weight a decision's margin earns, by maximum likelihood. The belief is a hidden Markov model over the two stories, filtered forward one decision at a time. Its steadiness counts from the second minute of each block; its switch times come from switches simulated halfway through each block, the later decisions counting for the other story.

Limits

Electrode arcs around the ear are not earbuds. Attention here was instructed and held for whole blocks, never switched. A quiet room and people sitting still. These are published research recordings, not Signalive product measurements.

Reproduce it

python3 research/aad/preprocess.pypython3 research/aad/analyze.pypython3 research/aad/export.pypython3 research/aad/calibrate.pyfor s in 7 11 23; do AAD_SEED=$s AAD_ENSEMBLE=3 python3 research/aad/deep.py; donepython3 research/aad/belief.pypython3 research/aad/deep_export.py

Decision window, seconds	4 per ear, wireless	All 18, wired	One earbud, 4	Chance for one person
1	53%	53%	52%	52%
2	54%	54%	54%	53%
5	56%	56%	55%	55%
10	58%	59%	57%	57%
20	61%	62%	59%	60%
30	63%	65%	61%	62%
60	68%	71%	66%	67%
120	74%	74%	71%
300	85%	85%	79%
600	89%	91%	88%

#	Decisions right
1	70%
2	57%
3	80%
4	77%
5	70%
6	77%
7	53%
8	70%
9	67%
10	60%
11	80%
12	83%
13	63%
14	67%
15	77%
16	53%
17	60%
18	70%
19	77%
20	73%
21	73%
22	77%
23	87%
24	67%
25	47%
26	47%
27	77%
28	77%
29	70%
30	73%
31	43%
32	87%
33	80%
34	43%
35	67%
36	53%

Decision window, seconds	Learned decoder	Linear decoder	Learned, stories out of step	Chance for one person
1	54%	53%	50%	52%
2	56%	54%	50%	53%
5	59%	56%	50%	55%
10	62%	58%	50%	57%
20	66%	61%	50%	60%
30	69%	63%	49%	62%
60	77%	68%	49%	67%
120	82%	74%	48%
300	89%	85%	47%
600	94%	89%	46%

Decision window, seconds	4 per ear, wireless	All 18, wired	One earbud, 4	Chance for one person
1	53%	53%	52%	52%
2	54%	54%	54%	53%
5	56%	56%	55%	55%
10	58%	59%	57%	57%
20	61%	62%	59%	60%
30	63%	65%	61%	62%
60	68%	71%	66%	67%
120	74%	74%	71%
300	85%	85%	79%
600	89%	91%	88%

#	Decisions right
1	70%
2	57%
3	80%
4	77%
5	70%
6	77%
7	53%
8	70%
9	67%
10	60%
11	80%
12	83%
13	63%
14	67%
15	77%
16	53%
17	60%
18	70%
19	77%
20	73%
21	73%
22	77%
23	87%
24	67%
25	47%
26	47%
27	77%
28	77%
29	70%
30	73%
31	43%
32	87%
33	80%
34	43%
35	67%
36	53%

Decision window, seconds	Learned decoder	Linear decoder	Learned, stories out of step	Chance for one person
1	54%	53%	50%	52%
2	56%	54%	50%	53%
5	59%	56%	50%	55%
10	62%	58%	50%	57%
20	66%	61%	50%	60%
30	69%	63%	49%	62%
60	77%	68%	49%	67%
120	82%	74%	48%
300	89%	85%	47%
600	94%	89%	46%

We ran the decoder on36 people's ear EEG

Watch it decide

What the recordings say

Longer windows,clearer answers

A few electrodesgo most of the way

No calibrationto start

Most people,not everyone

A learned decoder,the same 36 people

What a decisionis worth

Weak alone,strong together

Pick an earbud,see what the data says

The head steers.The EEG confirms.

True wireless works

Four electrodes per ear

Wear it and go

EEG is the slow channel

Everything butthe hardware

Decoding pipeline

Wire protocol

Digital twin

Dev kit bridge

Hardware spec

Research code

Method and limits

Decision window, seconds	4 per ear, wireless	All 18, wired	One earbud, 4	Chance for one person
1	53%	53%	52%	52%
2	54%	54%	54%	53%
5	56%	56%	55%	55%
10	58%	59%	57%	57%
20	61%	62%	59%	60%
30	63%	65%	61%	62%
60	68%	71%	66%	67%
120	74%	74%	71%
300	85%	85%	79%
600	89%	91%	88%

#	Decisions right
1	70%
2	57%
3	80%
4	77%
5	70%
6	77%
7	53%
8	70%
9	67%
10	60%
11	80%
12	83%
13	63%
14	67%
15	77%
16	53%
17	60%
18	70%
19	77%
20	73%
21	73%
22	77%
23	87%
24	67%
25	47%
26	47%
27	77%
28	77%
29	70%
30	73%
31	43%
32	87%
33	80%
34	43%
35	67%
36	53%

Decision window, seconds	Learned decoder	Linear decoder	Learned, stories out of step	Chance for one person
1	54%	53%	50%	52%
2	56%	54%	50%	53%
5	59%	56%	50%	55%
10	62%	58%	50%	57%
20	66%	61%	50%	60%
30	69%	63%	49%	62%
60	77%	68%	49%	67%
120	82%	74%	48%
300	89%	85%	47%
600	94%	89%	46%