Cross-day speech decoding without retraining the network: 65,792 input parameters lift all six T12 sessions

Cross-day adaptation for speech decoding can happen only at the neural input, without retraining the high-capacity part that models temporal sequence and phoneme output. Earlier approaches either fine-tuned the model without restriction, adding many target-session-specific degrees of freedom, or relied on aligning low-dimensional neural representations to stabilize decoding. On six chronologically held-out T12 sessions, macro phoneme error rate fell from 0.4211 to 0.3732 with 40 calibration trials and from 0.4219 to 0.3565 with 80; all six sessions improved under all three calibration seeds, with nominal paired signed-rank p = 0.03125 for both budgets.

The calibration rule selected on T12 was transferred unchanged to T15, with no retuning against T15 test performance. T15 used a separately trained source decoder, and the learning rate, adaptation scope, optimization steps, calibration budgets and seed set all carried over from T12 development. On T15, macro phoneme error rate fell from 0.6324 to 0.6152 at 40 trials, with four of eight sessions improving under every seed, and from 0.6123 to 0.5508 at 80 trials, with four of six eligible sessions improving under every seed; nominal p values were 0.546875 and 0.15625.

Error decomposition shows the calibration mainly corrects decoded sequences that were too short, rather than trading one edit-error type for another. On T12 at 40 trials, substitution rate fell from 0.2754 to 0.2445 and deletion rate from 0.1057 to 0.0857, while insertion rate moved only from 0.0469 to 0.0491 and the decoded-to-reference phoneme-length ratio rose from 0.9412 to 0.9634. On T15 at 40 trials, deletion rate fell from 0.171822 to 0.107795 and the length ratio rose from 0.850713 to 0.950242, but substitution and insertion rates rose; only at 80 trials did substitution and deletion fall together.

Target-session calibration costs seconds because more than 99.95% of the pretrained parameters stay frozen. Only the session-specific 256×256 affine matrix and its 256-dimensional bias are optimized, giving 65,792 trainable parameters, 0.049398% of the 133,187,113-parameter T12 source decoder and 0.048939% of the 134,436,649-parameter T15 decoder. On the evaluated GPU, T12 calibration averaged 7.52 s at 40 trials and 7.56 s at 80, with peak allocated memory near 1.28 GB; T15 took 20.12 s and 23.46 s, with peak memory near 1.80 GB and 1.91 GB.

The work evaluates two public intracortical attempted-speech datasets, each from a single BrainGate2 participant, and all analyses are offline. Calibration is supervised and requires 40 or 80 labeled trials from the target session, and only restricted affine and residual-adapter scopes were compared. T12 and T15 differ in native feature dimensionality, recording chronology and source-model construction, so the T15 experiment should be read as independent cross-dataset evaluation rather than a controlled single-factor replication.

Neither T15 calibration budget fell below the nominal 0.05 reference threshold, and session-level responses were less uniform than on T12, so a single low-capacity transformation cannot be claimed to remove cross-session mismatch universally. The correctable fraction depends on the target recording session and the amount of calibration information available. Validation phoneme error rates for affine-only and affine-plus-adapter calibration were nearly identical, 0.295796 and 0.295989, so affine adaptation is not shown to be statistically superior to more expressive parameter-efficient methods. Generalization to additional participants, implant configurations, longer recording periods and closed-loop communication remains unknown.