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.

• 157 words, 279 occurrences. Last review: 2026-10-28.
• 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).
• active: the word is still being recorded for retraining (takes under target); it switches off by itself when the target is reached. priority: high words are recorded before normal ones.
• source: manual means the word was added by hand, not from a reviewed error, so it has no occurrences.

Active words first, high priority first among them.

Word	Heard as	Occurrences	Takes / target	Active priority
arXiv	arcs of, Artives, arcsiv, argsiv, Clike	5	0 / 20	high
Konsole	console	4	5 / 20	high
abdominophrenic dyssynergia	abominophrenic dysinurgia	1	0 / 20	high
accent transfer	Axon Transfer	1	0 / 20	high
base	bass	1	0 / 20	high
code-switching	co-switching	1	0 / 20	high
DHL	HL	1	0 / 20	high
dyspepsia	dyspepsy	1	0 / 20	high
feed	fee	1	0 / 20	high
fluticasone	Fluticazone	1	0 / 20	high
hunches	punches	1	0 / 20	high
lifelike	lifly	1	0 / 20	high
OTC	ODC	1	0 / 20	high
pairs	pears	1	0 / 20	high
potentially	prudentially	1	0 / 20	high
re-architect	re-arcolate	1	0 / 20	high
Relvar	Ralvar	1	0 / 20	high
salbutamol	Samputumal	1	0 / 20	high
sans-serif	Sanserif	1	0 / 20	high
serif	Sara	1	0 / 20	high
Singulair	Singular	1	0 / 20	high
spacer	Spacerin	1	0 / 20	high
text	test	1	0 / 20	high
Ventolin	Ventilin	1	0 / 20	high
Vercel	inversel	1	0 / 20	high
My Weird Prompts	MyWord Prompts, My WordPrompts, My Word Promise, mywordprompt, My Word prompts	7	14 / 20	normal
Typst	TyPst, TIPS-type, tips toypt, TyPstip, tipst, Tips	6	9 / 20	normal
CLAUDE.md	userlevelclaw.md, 4 MD, Claude NMD, FlawdMD	4	9 / 20	normal
Easy Effects	EasyFX	4	5 / 20	normal
Infisical	Infiniscal, Infinisco, Inphysical, physical	4	8 / 20	normal
disfluency	disluency, Dispuency, fluencies	3	9 / 20	normal
mic	Mike	3	5 / 20	normal
Composio	Compi Posio, Compose	2	4 / 20	normal
OnePlus Nord 3 5G	OnePlus Nord 3.5G, OnePlus Nord 35G	2	4 / 20	normal
Parakeet	Power Key, PowerKey	2	7 / 20	normal
auto gain	auto game	1	4 / 20	normal
convention	invention	1	4 / 20	normal
cutlery	cuddlery	1	4 / 20	normal
Debian	Debin	1	7 / 20	normal
IBM Plex Sans	IBM Plexans	1	4 / 20	normal
MCP	MCP2	1	19 / 20	normal
MCP Hub	MCB Hub	1	3 / 20	normal
OTP	OTB	1	5 / 20	normal
STT	SST	1	7 / 20	normal
tech vocab	type vocab	1	4 / 20	normal
ThinkPad	ThinkPaddy	1	4 / 20	normal
Claude Code	clawed code, clawed, claud, clawed coat, claw code, Core Code, claudcode, claud Code, the claud, clause claud code	15	12 / 6
LoRA	Lores, Loas, Laura, lores, Loa, LoRa, lore, lures	15	8 / 6
ROCm	ROCM, Rockham, Rockin, RCM, Rock'em	15	12 / 11
Corn	Coren, Cornak's, Eminent Corn, Corner, Korn, Corin	6	5 / 5
Hugging Face	Hooking Face, hugging face, hooking face, Hogging Face, hogging face	6	11 / 11
Hebrish	HeBrish, Hebris, keybrush, Hebrew Shop, Hebrew shtal	5	13 / 11
NocoDB	No CodeDB, Nocode, NokoDB, Noko DB	4	6 / 6
skill	scale, kills	4	5 / 5
8BitDo	8-bit do, four-bit do	3	9 / 9
Cloudflare	cloud fair, Cloud Claire, Cloud Clare	3	5 / 5
Colab	Collab, collab	3	6 / 6
evals	evalves, valves, haval	3	6 / 6
F-Droid	FDroid, Fdroid	3	10 / 10
Nemotron	Nematron, Nementron	3	13 / 11
Unsloth	Unsoft, On Sloth, Onsloth	3	8 / 7
ab initio	a benissio, abenisio	2	5 / 5
AliExpress	Alley, Alliance	2	7 / 7
DeepSeek	DeepSee, DeepC	2	12 / 9
DMARC	GMARC, Dimark	2	5 / 5
Hannah	Hanna, Anna	2	6 / 6
Jupyter	Jupiter	2	10 / 10
Karpathy	Karpathi, Karpathi's	2	10 / 10
MiniMax	Mini Max	2	8 / 8
Netanyahu	Etanyahu, Etnia	2	6 / 6
ONNX	ONIX, onyx	2	10 / 10
Qwen	Qwen1, quen	2	13 / 11
Sabich	Tzabich, Savich	2	22 / 11
series	Ceres, Siries	2	5 / 5
TTS	CTS, tth	2	9 / 7
Ubiquiti	ubiquity	2	6 / 6
WhisperX	Whisper X, Whisper Air	2	6 / 6
Z alias	ZALAS, Zed Elias	2	5 / 5
abliteration	Obliteration	1	5 / 5
agentic	agenda	1	5 / 5
Ali	Alley	1	8 / 8
AMD	MD	1	5 / 5
APK	ABK	1	6 / 6
ASR	ACR	1	6 / 6
backhaul	backall	1	6 / 6
Bezeq	Bezek	1	4 / 4
Bituach Leumi	betuachlomi	1	9 / 6
Boxty	E box d	1
Carrot Cake AI	Karrot Cake AI	1	10 / 10
Chatterbox	Shatterbox	1	10 / 10
Chevrat Hashmal	Chashmal Yisrael	1	4 / 4
Claude in Chrome	Claude Grouping Chrome	1	4 / 4
co.il	co.al	1	4 / 4
do you want me	Dramat me	1	4 / 4
DSR Holdings	Vidyester Holdings	1	8 / 8
eSpeak	Espeak	1	10 / 10
EXE	AXE	1	10 / 10
exports	experts	1	6 / 6
Ezra	Esther	1	5 / 5
fal	Fel	1	5 / 5
feedback loop	feedback leap	1	6 / 6
few-shot	fusil	1	7 / 5
fill	Phil	1	6 / 6
gut chodesh	Good church chodesh	1	6 / 6
Hagihon	Gichon	1	4 / 4
Hegseth	Hagsets	1	5 / 5
in bed	embed	1	6 / 6
inbound	for bound	1	4 / 4
inline	inland	1	5 / 5
Ivory	Ivor	1	8 / 8
Janglo	Danglo	1	5 / 5
Jupyter Notebook	Jupiter notebooks	1	10 / 10
Kaizen	Kaiser	1	5 / 5
Kdenlive	Caden Live	1	10 / 10
KSP	NKS	1	8 / 8
lav mic	Lav Mike	1	4 / 4
LLM	nelam	1	12 / 7
marking	marketing	1	4 / 4
mass sender	math sender	1	6 / 6
Meorav	Maorab	1	17 / 11
Modbus	Mod bus	1	6 / 6
Muse Spark	News Spark	1	5 / 5
NIC	NISH	1	6 / 6
nikud	ikut	1	6 / 6
Nvidia	NVideo	1	6 / 6
Ollama	ulama	1	13 / 8
OPNsense	OpenSense	1	5 / 5
Opus	Focus	1	7 / 5
populated	opulated	1	4 / 4
post-processing	PostgreSQL processing	1	5 / 5
prosody	porosity	1	10 / 10
pseudospam	Surasfam	1	3 / 3
Qdrant	Qdrint	1	5 / 5
rclone	Arclone	1	4 / 4
regex	RedJack	1	6 / 6
remote	ote	1	4 / 4
repo	Refo	1	4 / 4
SaaS	SAS	1	5 / 5
SBC	SPCAdmin	1	5 / 5
SFP	SFEN	1	6 / 6
Sonnet	Sommet	1	8 / 8
spray and pray	spray and prey	1	6 / 6
SSE	SSC	1	6 / 5
subscription	prescription	1	4 / 4
Superwhisper	Superwisper	1	5 / 5
system tray	system try	1	4 / 4
tab	top	1	4 / 4
tell her	Teller	1	4 / 4
TP-Link	TPM	1	6 / 6
vibe code	Vipecode	1	6 / 6
vocab	Rocab	1	4 / 4
Whisperian	Wisperion	1	5 / 5
Fritter		0	12 / 10
gezuntheit		0	6 / 6
vibe coded	Vipecode	0	6 / 6
vibe coding	Vipecode	0	6 / 6
zei gezunt		0	6 / 6

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.