subvocal

Automated Performance Benchmark Report

Generated on: Mon Jun 8 10:51:05 CDT 2026
Version: v0.3.0-alpha (Automated release run)


1. sEMG Classifier Inference Benchmarks

These metrics profile real-time execution speeds, computational footprint, and energy consumption estimates.

Model Type Mean Latency Median Latency Disk Footprint Params Size Est. FLOPs Est. Energy
1D CNN 1.66 ms 1.64 ms 409.7 KB 102,053 3,678,336 367.834 µJ
GRU 8.27 ms 8.25 ms 420.5 KB 105,669 30,190,912 3019.091 µJ
SVM Baseline 0.00 ms 0.00 ms 0.0 KB 0 0 0.000 µJ

2. Shorthand Intent Reconstruction Benchmark

These metrics profile shorthand-to-intent phonetic alignment and LLM resolution loops evaluated over 50 gold-standard phrases.

  • Intent Resolution Accuracy: N/A
  • Intent Resolution Latency: N/A

Full Intent Runner Output

================================================================================
SUBVOCAL SDK INTENT-RECONSTRUCTION BENCHMARK RUNNER
================================================================================
Live LLM Providers Detected: NONE (Running in Simulated mode)
Total Test Cases: 50
--------------------------------------------------------------------------------

================================================================================
BENCHMARK RESULTS SUMMARY
================================================================================
Decoder / Provider        | Accuracy   | Avg Latency    
--------------------------------------------------------------------------------
HEURISTIC                 |     74.0% |       0.70 ms
SIMULATED-LLM             |     96.0% |       0.70 ms

================================================================================
FAILURE PROFILE DETAILS
================================================================================
Provider HEURISTIC: 13 failures
  Noisy Input     : 'srch nrl ntwk'
  Expected Intent : 'SEARCH neural networks'
  Received Output : 'SEARCH Neural network'

  Noisy Input     : 'srch fgt calb'
  Expected Intent : 'SEARCH Figma calibration'
  Received Output : 'SEARCH logout confirm'

  Noisy Input     : 'srch emguka cp'
  Expected Intent : 'SEARCH EMG-UKA corpus'
  Received Output : 'SEARCH EMG-UKA copy'

  Noisy Input     : 'srch snt spch'
  Expected Intent : 'SEARCH silent speech'
  Received Output : 'SEARCH send Speech'

  Noisy Input     : 'srch flyt scdl'
  Expected Intent : 'SEARCH flight schedule'
  Received Output : 'SEARCH Flight seconds'

  ... and 8 more failures.
--------------------------------------------------------------------------------
Provider SIMULATED-LLM: 2 failures
  Noisy Input     : 'gt ggl.cm'
  Expected Intent : 'GOTO google.com'
  Received Output : 'GOTO error.com'

  Noisy Input     : 'srch bsnl'
  Expected Intent : 'SEARCH BioSignals'
  Received Output : 'GOTO error.com'

--------------------------------------------------------------------------------

©2026 Pranav Kalkunte MIT License English