A new preprint describes a chip aimed at decoding auditory attention from EEG in real time. The design pairs an 8-bit quantized convolutional neural network (CNN) with a Pearson-correlation classifier—a method that scores how closely two signals track each other.
Algorithmic validation used real EEG datasets collected by cochlear-implant patients. The report also presents post-layout simulation results; it says the chip has been taped out, but measured power and real-time validation on patient-derived EEG remain future work.
Inside the streaming design
The system takes a 31 × 100 EEG window from 31 channels, reconstructs a stimulus, and produces two correlation coefficients plus an attention label. Its cross-layer streaming dataflow keeps intermediate results moving between operations; only the second convolutional-layer feature map is stored in SRAM, and model parameters can be updated through serial SRAM between listening sessions.
The classifier reformulates Pearson correlation so it can accumulate values sample by sample. It retains eight coefficients in a circular buffer and updates a moving average once per second, while division and square-root calculations run in multi-cycle sequential units.
Quantization kept the reported decoding pattern
Quantization converts model values into a smaller fixed representation. In the paper’s CNN-output comparison, the signal-to-noise ratio (SNR)—a measure of output relative to noise—was 19.79 dB for PyTorch Int8, 13.45 dB for forced Power-of-Two (PoT) quantization and 14.94 dB for the Quark PoT variant. The text describes the Quark result as about 1.5 dB better than forced PoT.
For the reported dataset, attended correlation coefficients across all evaluated models predominantly exceeded distractor coefficients. The authors describe the coefficients and final auditory-attention-decoding accuracy as robust to PoT quantization, but the report does not provide a numerical accuracy figure or uncertainty estimate.
The headline hardware numbers are still estimates
The abstract reports a fully implemented GF22FDX 22-nm ASIC with 2.09 mm² of total silicon area, a 1,264 µm × 1,654 µm footprint and 0.076 mm² for the CNN inference and streaming-classification engines.
The detailed physical-implementation section gives a different total-area figure: 76,048 µm² across 59,149 instances. It assigns 82.4% of area to the CNN, 14.4% to classification and 3.2% to peripherals. The preprint does not reconcile the two total-area figures.
Post-layout sign-off estimated average chip power at 0.4941 mW when operated at 0.55 V and 25°C; the CNN plus classifier accounted for 0.3265 mW. A separate worst-case leakage estimate was 0.1953 mW at 0.6 V and 55°C. These are simulation and sign-off estimates, not measurements from fabricated silicon.
Reported processing latency was 7.34 ms, within the 8 ms/sample requirement for a 125 Hz EEG stream. Using 0.49 mW and 7.34 ms, the paper calculates 3.63 µJ per inference. That is a derived energy figure, not measured fabricated-chip energy.
What the preprint does not establish
Hardware execution was bit-true identical to the software-based PyTorch PoT simulation. That checks consistency between the implementation and its software reference; it does not by itself establish patient-level decoding performance.
The report gives the source of the EEG data but not a participant count, dataset composition or numerical final auditory-attention-decoding accuracy. Its reported dataset did show attended coefficients generally exceeding distractor coefficients.
Measured power and real-time patient-derived EEG validation remain future work. As a result, the preprint does not establish improved hearing assistance or clinical benefit.
Paper data and sources
Original title: A Resource-Efficient CNN-Based EEG Auditory Attention Decoding ASIC
Authors: Qier Ma, Richard George, Stefan Scholze et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text