🌿 transcript-fixer v1.0.0 · MIT

Deterministically fix proper names that speech-to-text mangles — before the transcript reaches any downstream consumer.

The problem

ASR output (e.g. gpt-4o-transcribe) renders recurring proper names badly — and prompt-level hints do not fully solve it, because the error is acoustic, not lexical. The fix is a deterministic rewriting pass: a curated, sourced table of mis-transcribed → correct forms applied with word boundaries and idempotence guarantees.

What it does

Sourced fix table — each entry backed by a real observed ASR error; edit it to match your own recurring mistakes.
Idempotent — fix(fix(x)) == fix(x); safe in multi-pass pipelines.
Protected forms — already-correct names are never rewritten (no false positives on close words).
Word boundaries, case-insensitive, longest multi-word expressions first.
CLI + Python API: --json, --list, --selftest, stdin/file input.
Zero dependencies, Python ≥ 3.9, single module.

Install (pip)

pip install transcript-fixer @ git+https://mandrilly.com/git/transcript-fixer.git

Usage

transcript-fixer "I flew to Nu York with Micheal"
# → I flew to New York with Michael   [2 fix(es)]

echo "Open Ai and Git Hub" | transcript-fixer -
transcript-fixer --selftest            # 7/7 OK
transcript-fixer --json report.txt     # machine-readable, with change details

Python API

from fix_names import fix_names, fix_report

fixed, changes = fix_names("Send the file to Micheal")
# ("Send the file to Michael", [{"from": "Micheal", "to": "Michael", "pos": 17}])

Tests

transcript-fixer --selftest   # 7/7 → OK (fixes, idempotence, protection)

Source

git clone https://mandrilly.com/git/transcript-fixer.git