Hebrish/ASR-Error-List

Hebrish ASR Error List

The words the Hebrish speech-to-text fine-tunes still get wrong in real dictation, what they were heard as, and how far targeted retraining on each one has got.

The errors come from one speaker's own dictation with the Hebrish desktop app (English with Hebrew code-switching, Jerusalem). Each word here is fed back into training as a priority word: the recording app serves sentences with it first until there are enough takes of it. No transcript text, audio or other content of the dictations is published, only the word, the wrong forms and the model.

• 64 words, 141 occurrences. Last review: 2026-10-24.
• models / model_line: the Hebrish model that produced the error (Sabich, Shakshuka, …, with version); model_rev is the exact weights' revision, recorded from app 0.7.3. Errors from before the app recorded the model have none.
• source: reported means the speaker reported the error directly; it is not in the reviewed history.
• takes / target: recordings of the word made for retraining so far, against the number wanted. status: training (under target), trained (target reached; a later error means it still fails).

Word	Heard as	Occurrences	Takes / target
ROCm	ROCM, Rockham, Rockin, RCM, Rock'em	15	9 / 20
LoRA	Lores, Loas, Laura, lores, Loa, LoRa, lore, lures	14	2 / 20
Claude Code	clawed code, clawed, claud, clawed coat, claw code, Core Code, claudcode	7	2 / 20
Corn	Coren, Cornak's, Eminent Corn, Corner, Korn, Corin	6	1 / 20
Hugging Face	Hooking Face, hugging face, hooking face, Hogging Face, hogging face	6	9 / 20
My Weird Prompts	MyWord Prompts, My WordPrompts, My Word Promise, mywordprompt	6	9 / 20
Hebrish	HeBrish, Hebris, keybrush, Hebrew Shop, Hebrew shtal	5	9 / 20
NocoDB	No CodeDB, Nocode, NokoDB, Noko DB	4	1 / 20
Cloudflare	cloud fair, Cloud Claire, Cloud Clare	3	1 / 20
Colab	Collab, collab	3	2 / 20
F-Droid	FDroid, Fdroid	3	9 / 20
Nemotron	Nematron, Nementron	3	9 / 20
Unsloth	Unsoft, On Sloth, Onsloth	3	2 / 20
CLAUDE.md	userlevelclaw.md, 4 MD	2	0 / 20
disfluency	disluency, Dispuency	2	0 / 20
evals	evalves, valves	2	1 / 20
Infisical	Infiniscal, Infinisco	2	1 / 20
Jupyter	Jupiter	2	8 / 20
Karpathy	Karpathi, Karpathi's	2	8 / 20
Netanyahu	Etanyahu, Etnia	2	2 / 20
ONNX	ONIX, onyx	2	9 / 20
Qwen	Qwen1, quen	2	10 / 20
series	Ceres, Siries	2	0 / 20
Typst	TyPst, TIPS-type	2	0 / 20
Z alias	ZALAS, Zed Elias	2	0 / 20
ab initio	a benissio	1	0 / 20
abliteration	Obliteration	1	1 / 20
agentic	agenda	1	0 / 20
AMD	MD	1	0 / 20
APK	ABK	1	1 / 20
ASR	ACR	1	3 / 20
Bituach Leumi	betuachlomi	1	0 / 20
Boxty	E box d	1
Carrot Cake AI	Karrot Cake AI	1	8 / 20
Chatterbox	Shatterbox	1	9 / 20
DeepSeek	DeepSee	1	5 / 20
DMARC	GMARC	1	0 / 20
DSR Holdings	Vidyester Holdings	1	3 / 20
eSpeak	Espeak	1	8 / 20
EXE	AXE	1	9 / 20
Ezra	Esther	1	0 / 20
fal	Fel	1	0 / 20
few-shot	fusil	1	0 / 20
Hegseth	Hagsets	1	0 / 20
Janglo	Danglo	1	1 / 20
Jupyter Notebook	Jupiter notebooks	1	8 / 20
Kaizen	Kaiser	1	0 / 20
Kdenlive	Caden Live	1	9 / 20
LLM	nelam	1	5 / 20
Meorav	Maorab	1	9 / 20
Muse Spark	News Spark	1	0 / 20
Nvidia	NVideo	1	1 / 20
Ollama	ulama	1	1 / 20
post-processing	PostgreSQL processing	1	0 / 20
prosody	porosity	1	9 / 20
Qdrant	Qdrint	1	0 / 20
regex	RedJack	1	1 / 20
SaaS	SAS	1	1 / 20
Sabich	Tzabich	1	9 / 20
Sonnet	Sommet	1	2 / 20
SSE	SSC	1	0 / 20
Superwhisper	Superwisper	1	0 / 20
TTS	CTS	1	4 / 20
Whisperian	Wisperion	1	1 / 20

Release tests

release_tests: each Hebrish model release scored on held-out read recordings of these words (test split of the speaker's read-aloud corpus, never trained on), with the shipped int8 ONNX files and the Hebrish desktop engine on CPU. exact: clips whose transcript has the word exactly as the reference writes it; ci: ignoring case; heard_as_hits: a known wrong form still appears. Small counts while the test recordings accumulate (1 in 5 of each priority word's sentences is recorded for test).

Maintained by Daniel Rosehill (Hebrish). Built by scripts/publish_error_list.py in the Voice-Training-Recorder repo.