Not just another "$600 microphone." This device has actual brains - Syntiant NDP120 handles wake detection and speaker ID at 150 microwatts. Your home server does the rest. No cloud. No subscription. Your data never leaves your network.
Commercial devices are fancy microphones. This is a smart AI device.
Syntiant NDP120 neural processor runs at 150 microwatts - always listening for human speech, not music, not TV, not silence. Wakes the main MCU only when needed.
The NDP120 knows WHO is talking - not just that someone is talking. Train it on your voice and your family. All on the device itself. No cloud processing needed.
Whisper transcription and LLM extraction run on YOUR hardware. Raspberry Pi 5 or a mini PC. Your data never touches the internet. Ever.
No cloud accounts. No telemetry. No data harvesting. Everything stays on your local network. Works even without internet.
One-time hardware cost. No monthly fees. No "$24/month forever" like commercial alternatives. Own your device, own your data.
Commitment tracking that knows the difference between what someone TOLD you to do vs what YOU said you'd do. Life logging for jokes, ideas, memories.
Commercial devices like Limitless ($300) and Humane AI Pin ($699) are literally just microphones.
All the AI runs in their cloud. You pay for the device, then pay $24/month forever to use it.
BrainGoblin puts real AI on the device (wake + speaker ID) and runs the heavy processing on hardware you control.
$115 total, $0/month ongoing.
Three components working together - device, server, and viewer
+-----------------------------------------------------------------------------+ | BRAINGOBLIN DEVICE | | | | +-------------------+ +-------------------+ +------------------+ | | | SYNTIANT NDP120 | | ESP32-S3 | | Storage | | | | (Always-on AI) | | (Main MCU) | | | | | | | | | | SD Card 32GB | | | | * Voice wake |--->| * Record audio |--->| Weeks of | | | | * Speaker ID | | * Compress | | recordings | | | | * 150 microwatt | | * WiFi/BLE | | Offline queue | | | +-------------------+ +-------------------+ +------------------+ | | | | ON-DEVICE: Wake detection, Speaker ID, Recording, Storage | +-----------------------------------------------------------------------------+ | | WiFi (home network) v +-----------------------------------------------------------------------------+ | HOME SERVER | | (Raspberry Pi 5 / Mini PC) | | | | +-------------------+ +-------------------+ +------------------+ | | | Whisper.cpp | | Local LLM | | Database | | | | | | (Llama/Mistral) | | | | | | * Transcription |--->| * Commitments |--->| * All your data | | | | * 100% local | | * Life logging | | * Searchable | | | | * No cloud | | * People | | * Encrypted | | | +-------------------+ +-------------------+ +------------------+ | | | | YOUR DATA NEVER LEAVES YOUR NETWORK | +-----------------------------------------------------------------------------+ | v +-----------------------------------------------------------------------------+ | PHONE APP | | | | * Just a VIEWER - no heavy processing | | * Shows transcripts, commitments, reminders | | * Configure device via BLE | | * Minimal battery usage | | | +-----------------------------------------------------------------------------+
| Task | Location | Power |
|---|---|---|
| Voice wake detection | NDP120 on device | 150uW |
| Speaker identification | NDP120 on device | 150uW |
| Audio recording | ESP32-S3 on device | 50mA |
| Local storage | SD card on device | - |
| Transcription | Home server | - |
| AI extraction | Home server | - |
| Viewing results | Phone app | Minimal |
Dual-board design with on-device AI and weeks of local storage
This is what makes BrainGoblin different from "dumb microphones." The NDP120 is a neural decision processor that runs tiny ML models at ultra-low power. It's the same technology Google uses for "Hey Google" always-on listening.
That's 0.00015 watts. A standard LED draws 20mW - 130x more power. This chip can listen 24/7 for MONTHS on a coin cell battery.
| Power Consumption | ~150 uW |
| Neural Network | 1M parameters (FW) or 4M (HW) |
| Audio Inputs | PDM, I2S |
| Sample Rates | 8kHz - 48kHz |
| Host Interface | SPI, I2C, GPIO |
| Training Platform | Edge Impulse |
| Package | 3.5mm x 3.5mm BGA |
# Train via Edge Impulse 1. Record 3-10 voice samples per person 2. Upload to Edge Impulse project 3. Train speaker classification model 4. Deploy to NDP120 # On-device identification NDP120 outputs speaker ID with each audio segment (0-N speaker index or "unknown")
The ESP32-S3 handles everything the NDP120 doesn't - recording audio, compressing it, storing on SD card, and uploading to your home server via WiFi. The XIAO Sense version includes an SD card slot built in.
The nRF52840 is great for BLE-only wearables, but we need WiFi to upload to the home server. The NDP120 handles the ultra-low-power always-on listening, so the ESP32-S3's higher power consumption only matters when actively recording/uploading.
ESP32-S3 stays in deep sleep (~14uA) until NDP120 sends a wake interrupt. Then it records, compresses, stores, and goes back to sleep. WiFi only activates for batch uploads to save power.
| Processor | ESP32-S3 Dual-core 240MHz |
| RAM | 8MB PSRAM |
| Flash | 8MB |
| WiFi | 802.11 b/g/n |
| Bluetooth | BLE 5.0 |
| SD Card | MicroSD slot (built-in) |
| Battery | LiPo charging circuit |
| Dimensions | 21mm x 17.5mm |
| Deep Sleep | ~14uA |
# ESP32-S3 to Nicla Voice ESP32 GPIO4 -> Nicla WAKE_INT (interrupt) ESP32 GPIO5 -> Nicla SPI_CS ESP32 GPIO6 -> Nicla SPI_CLK ESP32 GPIO7 -> Nicla SPI_MOSI ESP32 GPIO8 -> Nicla SPI_MISO 3.3V -> Nicla 3.3V GND -> Nicla GND # I2S Microphones ESP32 GPIO1 -> I2S BCLK ESP32 GPIO2 -> I2S LRCK ESP32 GPIO3 -> I2S DOUT
The latest generation I2S MEMS microphone from TDK InvenSense. Unlike the older (now discontinued) INMP441, the ICS-43434 has native I2S output with standard timing - no hacks needed.
The ICS-43434 is a bottom-port microphone. The sound hole faces DOWN toward the PCB. Your enclosure needs an acoustic port aligned with this hole!
| Mic | Interface | SNR | Status |
|---|---|---|---|
| ICS-43434 | Native I2S | 65dB | Chosen |
| INMP441 | PDM->I2S | 61dB | EOL |
| SPH0645 | Weird I2S | 65dB | Problematic |
| Nicla PDM | PDM | ~60dB | Backup |
# Both mics share same I2S bus # L/R pin determines which clock edge Mic 1 (Left): L/R -> GND Mic 2 (Right): L/R -> 3.3V # Same BCLK, LRCK, DOUT lines # MCU receives stereo interleaved data
The magic of BrainGoblin is that the NDP120 does the always-on listening at microwatts, while the ESP32-S3 only wakes when there's actual speech to record.
| State | Current | Battery Life (1000mAh) |
|---|---|---|
| Idle (NDP120 listening) | ~150uA | Months |
| Active recording | ~50mA | ~20 hours |
| WiFi sync | ~100mA | ~10 hours |
Assuming 4-6 hours of actual speech per day (realistic for most people), a 1000mAh battery should last 2-3 days between charges. With WiFi duty cycling (sync every 15-30 minutes), even longer.
ALWAYS ON (NDP120 listening): - NDP120: ~150 uA - Total: ~150 uA RECORDING (ESP32 active): - NDP120: ~150 uA - ESP32-S3: ~40 mA (with I2S) - ICS-43434: ~1 mA (2 mics) - SD Card: ~5 mA (write) - Total: ~46 mA WIFI UPLOAD: - NDP120: ~150 uA - ESP32-S3: ~100 mA (WiFi TX) - Total: ~100 mA DEEP SLEEP: - NDP120: ~150 uA - ESP32-S3: ~14 uA - Total: ~164 uA
The XIAO ESP32-S3 Sense has a built-in LiPo charging circuit. Just plug in USB-C and it charges the battery automatically. No extra components needed.
Your data never leaves your network - run everything locally
Any computer that can run Docker. Two main options:
Everything runs in Docker containers for easy setup and isolation.
# docker-compose.yml (simplified)
services:
whisper:
image: ghcr.io/ggerganov/whisper.cpp
volumes:
- ./audio:/data
ollama:
image: ollama/ollama
volumes:
- ./models:/models
api:
image: braingoblin/server
ports:
- "8080:8080"
depends_on:
- whisper
- ollama
Everything you need to build a BrainGoblin device (~$115 total)
Compare to Limitless ($300 + $24/mo) or Humane AI Pin ($699 + $24/mo). And this has ACTUAL on-device AI, not just a microphone streaming to the cloud!
Optional: Home server adds ~$80-150 depending on whether you use a Raspberry Pi 5 or a Mini PC. But you may already have suitable hardware lying around.
How BrainGoblin stacks up against commercial alternatives
| Feature | Limitless | Humane AI Pin | BrainGoblin |
|---|---|---|---|
| Wake Detection | Cloud | Cloud | On-Device (NDP120) |
| Speaker ID | Cloud | Cloud | On-Device (NDP120) |
| Transcription | Their cloud | Their cloud | Your server |
| AI Processing | Their cloud | Their cloud | Your server |
| Monthly Cost | $24/mo | $24/mo | $0 |
| Data Location | Their servers | Their servers | Your network |
| Works Offline | No | No | Yes (stores locally) |
| Open Source | No | No | Yes |
| ADHD-Focused | No | No | Yes |