Hebrish/ASR-Error-Prediction

ASR Error Prediction

Training data for a model that predicts which words a speech recognition model will get wrong for one speaker, by extrapolating from the words it has already been confirmed to get wrong. Built for the Hebrish fine-tunes of nvidia/parakeet-tdt-0.6b-v2 (English with romanized Hebrew code-switching). Plan, scripts and method: danielrosehill/ASR-Mistranscription-Predictor.

Segments

Every config is split by segment, so predictions can be made and used per training dataset:

Segment	What it covers	Source of candidates
codeswitching	romanized Hebrew and Yiddish terms	the Hebrish lexicon (places, food, bureaucracy, religious life, ...)
tech	tech, AI, Linux and smart-home terms and acronyms	the project's tech vocabulary packs, plus LLM proposals
names	people, companies, projects	none yet (fixes only)
general_english	ordinary English words	LLM proposals

Config	codeswitching	tech	names	general_english
fixes	7	59	9	12
candidates	3634	1112	0	0
predictions-haiku-r1	3634	1163	0	55

Configs

• fixes: the positives. Words confirmed wrong in the speaker's real dictation, from Hebrish/ASR-Error-List (pinned in error_list_rev): word, heard_as (the wrong forms the ASR wrote), kinds, occurrences, models, dates, and label (acoustic, or orthographic when the only error is casing or a hyphen).
• candidates: words the speaker is likely to say that are not on the fix list: term, domain, meaning, takes (recorded training takes so far; null = unknown) and pool (source file and commit).
• predictions-<run>: one model's scores for the candidates, plus terms it proposed outside the pools (origin: pool or proposed): p_fail (0-100), pattern (the confusion class it applied), likely_heard_as, reason, rank within the segment.

Predictions are not labels. They are rankings to be checked against the real model (a synthetic probe, then a few real recordings) before any word gets training sentences. Confirmed ones move to the error list with source: predicted.

Runs

Run	Date	Model	Notes
haiku-r1	2026-10-25	claude-haiku-5-5 (Anthropic Batch API, low effort; batch msgbatch_01SFKjYbEtvGqChr8eo3JXPk, US$0.13)	first pass; fix list at 87 words; scores every candidate in chunks of 60 with the whole fix list as evidence, plus 52 proposed tech terms and 55 proposed English words; 1 candidate missing from the replies

haiku-r1 caveats. Scores are uncalibrated and not comparable across segments. Code-switching scores run high (mean 58 for terms with no recorded takes, 37% of candidates at 60 or above) because the fix list has only 7 Hebrew words to learn from, so Haiku leans on "Hebrew = hard". Scores do fall with recorded takes (mean 58 at 0 takes, 49 at 1 to 4, 31 at 5 to 19, 22 at 20 or more), which is the expected direction. Use ranks within a segment, not thresholds. Wrong forms that equalled the term itself (1,175 of them) were removed after the run.

Not included

No audio, and nothing from the speaker's dictations beyond single confirmed words and the ASR's wrong forms. Negatives (rare words the model transcribed correctly) are planned but not published yet.