Hebrish/Hebrew-Term-Pronunciations-IPA

Hebrew Term Pronunciations (IPA)

How Hebrew words are pronounced inside English speech, written in IPA, measured from real recordings of one Anglo-Israeli speaker (Daniel Rosehill). Pilot of 20 terms, 2026-10-28.

Variation is intended. The speaker deliberately records more than one pronunciation of a term: Ashkenazi and Israeli (Sephardi-based) forms, common ways Anglo-Israelis say the word inside English, and sometimes common mispronunciations, because an ASR model has to transcribe all of them as the same word. A term's pronunciation also shifts with its place in the sentence (initial, medial, final) and with the stress and speed around it. So variants is the data, not noise around one right answer, and ipa is only the most typical form among these takes, not a correct or standard pronunciation.

• terms: one row per romanized term: the most typical form (ipa, ipa_phones), the form fed to Kokoro TTS (kokoro, stress on the last syllable by assumption), the number of takes, and every distinct phone string heard with its count (variants). group is chet (contains [χ]), resh (contains [ʁ]) or plain.
• takes: one row per source take: the take id in danielrosehill/Read-Aloud-Voice-Sentences, the sentence, the term's time span in seconds, the recognized phones, the term's position in the sentence, and the take's hebrew_pronunciation and variant_spelling tags from the recording app (ashkenazi, israeli, mixed, or null when the take was not tagged). Most pilot takes are untagged (4 of 150 tagged), so most variation here is not yet labelled by style.

Method. Each take is force-aligned to its known text (MMS aligner); the term's span is phone-recognized by facebook/wav2vec2-lv-60-espeak-cv-ft; ipa is the medoid take (smallest summed phone edit distance to the term's other takes). The medoid is chosen without regard to the pronunciation tags.

Caveats. Not yet checked by ear, so recognizer errors and intended variants are not yet told apart. The phone recognizer is trained on espeak's phone set, and it often hears [χ] as [x] and drops or replaces uvular [ʁ]: cheshbon's consensus starts with [ɾ] although most of its takes start with [x], and ishur has no final r. Use variants and the per-take rows rather than ipa alone.

Part of the Hebrish code-switching ASR project; used to seed synthetic voices in Hebrish/Synthetic-Codeswitching-Voices-Pilot.