v7 · hebrew

The Hebrew round

English is solved: Onyx narrates the site, free and local. Hebrew is a different country. The engines that read English beautifully either do not speak Hebrew at all or garble it, and every Hebrew mistake lands in front of a native speaker. Three takes below, one honest failure among them, and the options that remain.

The test passage · 221 characters, unpointed, no known homographs לפעמים הדרך הטובה ביותר לקרוא מאמר היא פשוט להקשיב לו. לוחצים על נגן, שמים את הטלפון בכיס, והקול ממשיך לקרוא בשבילך. הדפדפן ברקע, האוזניות מחוברות, ואפשר לדלג קדימה או אחורה בכל רגע. זה כל הרעיון של קורא הקול של האתר הזה.

01 · ListenThree takes, one honest failure

Your voice · ElevenLabs v3

The Amit Storyteller clone reading Hebrew — the path the story pipeline already proved. Transcribes near-perfect on this passage.

18.8spaid · he

Brian · speaking Hebrew

The site voice crossing languages: one narrator identity for the whole site, English and Hebrew. Also transcribes near-perfect — with an American accent you should judge by ear.

18.1spaid · he

Local · Chatterbox Multilingual

The free engine's official Hebrew, cloned from your voice. It fails — the transcription check came back garbled and the generator itself warned of token repetition. On the rack so the failure is heard, not just claimed.

~19slocal · fails

02 · EvidenceTranscribed back, word against word

Take	The check heard	Verdict
Your voice · v3	לפעמים הדרך הטובה ביותר לקרוא מאמר היא פשוט להקשיב לו…	near-perfect
Brian · Hebrew	לפעמים הדרך הטובה ביותר לקרוא מאמר היא פשוט להקשיב לו…	near-perfect
Local · Chatterbox MTL	לפעמים הדחת ובאוטר לקרום המהלי פשט, להקשיב לו…	garbled

The checker itself is a small multilingual model, so tiny deviations are its noise rather than the voice's — but the local take's failure is not subtle: הדרך הטובה became הדחת ובאוטר. This matches a month of ledger history: every local engine that touches Hebrew breaks somewhere — niqqud read as Arabic, homographs flipped, and now this.

03 · The optionsWhere Hebrew can go from here

Option	Cost class	The trade
ElevenLabs v3, your clone	credits	The proven path. Your own voice on your own Hebrew pages reads as authored, not narrated. Homographs need the pronunciation-dictionary workflow the story pipeline already built.
ElevenLabs v3, Brian	credits	One narrator identity across both languages. The accent is the question only your ear can answer.
VoxCPM, your clone (local)	$0	The story pipeline's measured local winner for Hebrew — but only WITHOUT niqqud, with known homograph misses, and slower than realtime. Worth a sample in the next round if the paid takes displease.
Azure / Google Hebrew neural	~$1–2 one-time	Competent, cheap, full SSML. Generic narrator character; no cloning of you. Untested here — no account keys on this machine yet.
Phonikud + Hebrew-specific TTS	$0, research	The proper fix for unpointed Hebrew (a dedicated grapheme-to-phoneme layer), and not a plug-and-play tool yet. Watch, do not build on.

The scale fact that changes the decision

The public site's Hebrew corpus is tiny — the real Hebrew lives in a few protected topics. At that scale the paid options cost a few hundred credits total, so unlike the English decision, price genuinely does not matter here. Pick purely by ear: your voice, or the site's voice speaking Hebrew.

04 · NextOne ear decision

Play the first two takes. If one of them reads Hebrew the way you want your pages read, the Hebrew story is closed with a few hundred credits a year. If neither does, the next round samples VoxCPM locally and prices an Azure key.