Skip to main content

​
Overview

Voice agents may occasionally mispronounce certain words — especially proper nouns, brand names, or Hebrew words with ambiguous pronunciation. The pronunciation dictionary lets you define word-to-respelling mappings per agent.

​
Format

Use English phonetic respelling with ALL-CAPS for stressed syllables:
Hebrew WordRespellingNotes
שקל (currency)SHE-kelStress on first syllable
שקל (weighed)sha-KALStress on second syllable
חשבוןkhesh-BONAccount / bill
ווייסיVOY-seeBrand name
מנכ”לmah-nah-KALCEO

​
How to Use

  1. Go to your agent’s settings in the dashboard
  2. Open the Pronunciation accordion section
  3. Click Add to create a new entry
  4. Enter the Hebrew word and its English phonetic respelling
  5. Save — the pronunciation takes effect on the next call

​
Tips

  • Only add mispronounced words — most Hebrew words are pronounced correctly by default
  • Keep the list short — long lists can dilute effectiveness. Focus on the words that matter most.
  • Test after adding — make a test call to verify the pronunciation improved
  • Iterate — if the first respelling doesn’t sound right, try a different one
  • Use CAPS for stress — Hebrew stress can fall on different syllables. SHE-kel vs sha-KAL produces very different results.
  • Include meaning for homographs — if a word has multiple pronunciations (like שקל), add separate entries with context in the word field (e.g., “שקל מטבע” and “שקל פועל”)

​
Generating Respellings

If you’re unsure how to respell a word, TaatikNet can help. It converts Hebrew text to romanization with stress marks — you can then convert the output to the ALL-CAPS format.

​
How It Works

The pronunciation dictionary works differently depending on the agent type:
  • Prompt agents (Gemini Live): A pronunciation guide is added to the agent’s system prompt. Since the model generates speech directly, it follows the pronunciation hints.
  • Flow agents (Chained pipeline): Hebrew words are replaced with their phonetic respelling before being sent to the text-to-speech engine.

​
Limitations

Pronunciation hints are best-effort. Expect approximately 90% consistency — the model may occasionally revert to its default pronunciation.
  • Results may vary slightly between calls
  • Works best for proper nouns, brand names, and uncommon words
  • Common Hebrew words are usually pronounced correctly without overrides
  • Maximum 50 entries per agent