Transcribe-HTR: An Auditable, LLM-Based Historical Document Intelligence Platform

Abstract

Transcription remains a major obstacle to large-scale research with handwritten historical documents. Recent multimodal large language models (LLMs) can read many difficult documents without collection-specific model training, but a single generated transcription provides little information about its own reliability. We present Transcribe-HTR, a locally deployable platform for batch transcription using models from multiple commercial providers. The system normalizes provider output, supports document-specific transcription instructions, combines independent attempts into a consensus reading, and identifies passages on which attempts disagree. It also provides deterministic consistency statistics, two-pass grounded named-entity recognition, summarization, and archival exports. Transcribe-HTR treats AI transcription as a repeatable measurement rather than an unquestioned result. This design helps historians, archivists, and digital-humanities researchers process collections at scale while directing human review toward uncertain documents and passages.