ASR Mistranscription Predictor

ASR-Mistranscription-Predictor is a planned project that aims to predict which words a speech recognition model will mistranscribe for a single speaker. Fine-tuning an ASR model on one person's voice is typically reactive: the speaker notices errors, records training sentences for them, and retrains. The project proposes instead to mine a confirmed fix list of correct words and their wrong transcriptions, explain those errors through generalisable features such as phoneme confusions, acronym structure, subword fragmentation, frequent English neighbours, and casing, and then score unseen candidate words by their likelihood of being mistranscribed. Top predictions would be checked against the real model with synthetic and then recorded probes before entering the fine-tuning data.

As of its October 2026 status, the repository contains the plan, documentation covering the training plan and confusion taxonomy, and a first LLM-based extrapolation run published as the Hugging Face dataset Hebrish/ASR-Error-Prediction. The pipeline is written in Python and uses scripts for building inputs, running predictions through an LLM gateway, and publishing to Hugging Face. The dataset is split by segment, including code-switching, tech, names, and general English, and the repository is designed so that no sentences or phrases from private dictation are published, only confirmed words and the incorrect forms the recogniser produced.

The work was planned for fine-tuning a Parakeet TDT model for English with Hebrew code-switching, but the method is described as general to any single-speaker ASR fine-tune that has a dictation history and a fix list. It may be of interest to developers working on speech recognition personalisation, ASR error analysis, or dataset construction for low-resource speaker adaptation.