twistedninja02/DialectLoop

A Multi-Agent LLM Workflow for Iterative Quality Control (QC) in Low-Resource Dialectal Speech Corpus Curation. # DialectLoop 📂🔄 > **A Multi-Agent LLM Workflow for Iterative Quality Control in Low-Resource Dialectal Speech Corpus Curation** > *Author: Anuj Sarker (anujsarker02@gmail.com)* > *Affiliation: Ahsanullah University of Science and Technology, Dhaka, Bangladesh* > *Prepared / Submitted for the upcoming International Conference on Machine Learning (ICML) / Association for Computational Linguistics (ACL)* --- DialectLoop treats corpus quality control (QC) as an iterative research task rather than a one-shot classification problem. A Transcription Auditor and a Dialect Verifier independently inspect transcript batches; a Critic reconciles their outputs and escalates uncertain cases; and a Summariser presents batch-level findings to a researcher before the next iteration. The paper evaluates this workflow on a 74-hour Bengali speech corpus containing 8,400 segments from 12 districts of Bangladesh. On a stratified, expert-adjudicated subset of 1,200 segments, DialectLoop reports a 91% error detection rate, 89% dialect-label accuracy, and Cohen's $\kappa=0.86$. ### 🌐 Live Application URLs * **Production/Shared Preview Build:** ais-pre-ruuqklpsttzfrbjes7x… * **Active Development Build:** ais-dev-ruuqklpsttzfrbjes7x… --- ## 🔍 Scope DialectLoop is designed for QC of an existing corpus with transcripts and district-level metadata. It does **not** perform data collection, initial transcription, ASR training, or dialect classification at deployment time. It is a research accelerator with explicit human oversight, not an autonomous annotation system. --- ## 📂 Repository Directory Structure The following tree separates executable code, reproducibility records, and publication assets. Entries marked `[planned]` are allocation placeholders and are not yet present in this review snapshot. ```text DialectLoop/ ├── README.md # Publication & setup d …