Voice-by-Induction

Induction earpieces have existed for decades but stayed niche. Real-time AI with always-on voice mode is new. Combine them and you get a voice only you can hear that knows everything — at least in theory. The terrain to cross was Ampere's Law at room scale, with whatever amplifier I could find online.

• LM386
• TPA3118
• Induction coil
• 30 AWG magnet wire
• Neodymium magnets
• 9V battery
• Gemini Live

$34 gift card balance.

1. Mar 2026 — opening

   Induction earpieces have been around for decades but stayed niche. Always-on AI voice mode is new. Combine them and you get a voice only you can hear that knows everything. That's the bet.

2. The theory

   Current through a loop draws a magnetic field. The field alternates at audio frequencies. A magnet in the ear canal vibrates with it. You hear sound. Ampere's Law at room scale.

   [image: Ampere's Law schematic — profile of a human head and neck wearing an induction loop, with labeled arrows running from an audio amplifier to the coil and a small magnet drawn inside the ear canal]

3. The build — $34

   Spent the gift card balance in one order: LM386 amplifier module ($7.49), 50g spool of 30 AWG magnet wire ($7.99), 120-pack of 5×2mm neodymium discs ($5.99), 9V battery and a 3.5mm aux cable for the rest.

   [image: Top-down product grid of the four parts ordered — blue LM386 PCB, red copper magnet-wire spool, stack of small silver disc magnets, Duracell 9V battery beside a black 3.5mm aux cable]

4. Thursday night

   Wound the loop in a friend's quiet room. Dropped a magnet by my ear. Started Gemini Live on the laptop. There was a voice only I could hear.

   [image: First listen — wire loop draped across a dim bedroom desk, laptop open with Gemini Live mid-sentence, single neodymium disc pinched between fingers and held to the ear]

5. School the next day

   Same rig, hallway between classes. The loop went silent unless I cupped both ears and stopped breathing.

6. More turns

   Assumed the field was weak and wound extra turns. No change at all. Was shown the math — N×I = V / R_per_turn. Double the turns, resistance doubles, current halves, field identical. Turns cancel completely. Loudness lives in amplifier voltage and wire gauge.

   [image: Whiteboard derivation of N×I = V / R_per_turn with the two N's crossed out in red — annotated to show that only amplifier voltage and per-turn resistance survive]

7. 18V

   Tried to brute-force more current at 18V. Puff of smoke. Dead board. LM386 maxes at 15V. Wrong chip entirely.

   [image: Stylized 3D render of the LM386 module mid-failure — wisps of blue-white smoke curling off the capacitor and potentiometer after the 18V push]

8. Three boards

   First board was twisted wires and globbed solder; I probed live pins with metal tweezers between them. Second had actual joints, still ugly. Third was solid enough that music came through the loop clearly for a moment.

   [image: Macro lineup of the three LM386 modules left-to-right — bird's-nest first attempt with copper twisting around the pins, slightly cleaner second with exposed copper, third with solid joints and the magnet-wire leads neatly anchored]

9. Where it goes next

   The chip is the ceiling, not the coil. A TPA3118 outputs 30W and handles 24V — roughly 40× the power. Pair it with 26 AWG (0.13 Ω/m vs 30 AWG's 0.34) and there's actual current to push. Next round, if there is one.

   [image: Side-by-side schematic comparison — small blue LM386 module labeled 0.7W / 15V / 30 AWG at 0.34 Ω/m, next to a larger TPA3118 amplifier board with toroidal inductors labeled 30W / 24V / 26 AWG at 0.13 Ω/m]

Audible in a quiet room; silent in a classroom without cupping both ears.

Geometry can't paper over an undersized amplifier — adding turns multiplied resistance and divided current at the same rate, so the field never grew. The ceiling was the chip, not the coil. At room scale, induction lives on amp output voltage and wire gauge, and a 0.7W module can't out-shout a hallway.