The Depths Must Add Up

Interval-Consistent Rescoring for Speech Recognition of Dictated Drill-Core Logs

A paper from the School of Underlying Conditions rescores a speech recogniser's n-best lists against the arithmetic of a drill-core run, first run by run and then jointly along the hole, and on 38.6 hours of dictation from a working core shed cuts the numeric field error rate from 10.8% to 3.7%. The same rescoring silently repaired 29 of the 37 runs the logger had spoken inconsistently; a margin test that flags such runs instead keeps all but six of those discrepancies in the record.

@misc{slop_5ib0r5,
  author       = {Ragnhild Uzoma and Vigdis Emenike},
  title        = {The Depths Must Add Up: Interval-Consistent Rescoring for Speech Recognition of Dictated Drill-Core Logs},
  year         = {2026},
  publisher    = {Slop University},
  doi          = {10.5555/slop.5ib0r5},
  url          = {https://slop.university/outputs/slop-paper-interval-consistent-decoding-5ib0r5/},
  version      = {1.0},
  note         = {Research paper},
}