Misophonia Labs is developing wearable in-ear hardware powered by on-device neural filtering. Our pipeline surgically isolates chewing, breathing, and oral triggers within an 8-millisecond budget on embedded microcontrollers—eliminating distress while maintaining natural conversation.
Well beneath the human auditory threshold for echo perception (10-15ms).
Peak suppression depth for masticatory and salivary acoustic bursts.
High perceptual speech quality score during active trigger elimination.
Inference runs entirely offline on embedded micro-controllers.
Misophonia is a recognized neurobehavioral condition where specific pattern-based biological sounds (chewing, swallowing, lip smacking, heavy breathing) trigger severe sympathetic nervous system arousal—causing involuntary panic, intense anger, and social withdrawal.
Commercial ANC headphones (like AirPods Pro or Sony WH-1000XM5) are designed strictly for stationary, predictable noise (airplane engines, air conditioning hums). They fail entirely with sudden, non-stationary oral transients.
Conversely, passive foam earplugs completely isolate individuals, creating severe conversational barriers during meals, lectures, and collaborative work.
Inverts phase for stationary frequencies below 1 kHz. Cannot track or react to unpredictable, rapid biological transients such as mouth clicks or crunching.
Dampens speech frequencies uniformly across the spectrum. Makes interpersonal communication impossible, forcing individuals into social isolation.
Extracts real-time STFT spectral features, estimates a complex ratio mask, and surgically attenuates the trigger signature while maintaining speech harmonics and directional cues.
Operating in the time-frequency domain to jointly reconstruct magnitude and phase with deterministic latency.
Signals are framed at 16 kHz with a 512-point window (32ms) advanced by an 8ms hop size (128 samples). Both real and imaginary spectral coefficients are fed into the tensor core to preserve phase clarity.
Trained on authentic trigger corpora blended with clean speech databases using an L1 Complex + Spectral Convergence composite loss function for rapid convergence.
The neural mask generator is quantized to 8-bit integer weights and tuned for sub-10ms execution budgets on high-speed embedded micro-controllers with hardware floating point units.
The hardware package pairs high-SNR front/rear MEMS microphones with an ergonomic custom in-ear acoustic canal. Dedicated low-power circuitry ensures rapid analog-to-digital conversion, neural computation, and balanced-armature output without perceptible phase delay.
Captures directional sound vectors to preserve spatial localization so you always know where conversation partners are speaking from.
Direct streaming interface bypassing operating system audio buffers, locking roundtrip audio processing latency below 8 ms.
Allows instant parameter updating from the companion mobile agent without routing live audio over wireless channels.
Every person with misophonia has a unique trigger signature. Instead of tedious manual equalization curves, our companion mobile app uses an intelligent conversational agent communicating via Model Context Protocol (MCP) to map user discomfort into runtime DSP filter parameters.
set_acoustic_profile
• High-frequency transient filter boosted (-30 dB on 2.5–6 kHz crunch harmonics).Users communicate in natural language with an intelligent conversational assistant. Through an MCP server interface, the agent directly invokes hardware tools to update the earphone's DSP registers and spectral coefficients over BLE.
Assists audiology professionals with gradual sound exposure therapy. The agentic assistant gently modulates attenuation levels over weeks, tracking emotional resilience without overwhelming the user.
Trigger detection statistics are stored locally on device and only metadata summaries are processed, maintaining strict privacy and ethical data standards.
Misophonia Labs began as an applied engineering research initiative driven by a critical technological gap: while audio AI has transformed speech recognition and generation, real-time edge processing for sensory and neurological sound sensitivities has remained largely neglected by major consumer hardware manufacturers.
The project focuses on digital signal processing, embedded firmware (C/C++), and deep learning acoustic models, developing bench prototypes to bring clinical-grade relief to everyday social and professional environments.
Understanding the difference between traditional noise cancelling and selective neural gating.
Standard ANC operates via analog phase inversion targeted at low-frequency, stationary background hums (engines, HVAC). Misophonia Labs uses digital neural masking to classify and eliminate sudden, non-stationary oral transients (chewing, smacking, clicking) while keeping vocal frequencies unmuted.
If audio passed to your eardrum experiences more than 10-15 ms of delay, you perceive an unnatural comb-filtering echo effect that makes normal conversation disorienting. Our quantized pipeline computes masks in under 3.6 ms, keeping total acoustic latency imperceptible.
No. All live audio processing and neural mask inferencing occurs strictly on-device on the embedded micro-controller. No raw audio ever leaves the earphone or streams over the internet.
The mobile companion app hosts an intelligent conversational agent communicating via Model Context Protocol (MCP). The agent translates user natural language feedback into structured hardware parameters sent over Bluetooth LE to adjust filter floors and frequency cutoffs.
We are accepting inquiries from clinical audiologists, researchers, and individuals seeking access to upcoming developer prototypes.