Pronunciation Guide

​

Overview

Voice agents may occasionally mispronounce certain words — especially proper nouns, brand names, or Hebrew words with ambiguous pronunciation (Hebrew is written without vowels, so one spelling can be read several ways). The pronunciation dictionary lets you define a per-word override for each agent.

​

Format

The override format is provider-aware — it depends on the voice your agent uses. The dashboard always asks for the Hebrew word, then a second field whose format matches the voice provider: Each entry stores both formats independently, so switching an agent’s voice between Soniox and DeepDub never reinterprets a niqqud string as IPA (or vice-versa) — the non-matching field is simply ignored.

​

How to Use

1. Go to your agent’s settings in the dashboard
2. Open the Pronunciation accordion section (the title shows the active format — “Pronunciation (Niqqud)” or “Pronunciation (IPA)”)
3. Click Add to create a new entry
4. Enter the Hebrew word, then its override in the format shown for your voice (niqqud for Soniox, IPA for DeepDub)
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 — focus on the words that matter most
• Test after adding — make a test call to verify the pronunciation improved
• Iterate — if the first override doesn’t sound right, try a different one
• Include meaning for homographs — if a word has multiple readings (like שקל), add separate entries with context in the word field (e.g., “שקל מטבע” and “שקל פועל”)

​

Generating Overrides

If you’re unsure how to write the override:

• Soniox (niqqud) — nakdan.dicta.org.il adds vowel points to plain Hebrew text.
• DeepDub (IPA) — phonikud-tts converts Hebrew text to IPA.

The dashboard links to the matching tool directly under the override field.

​

How It Works

For a flow agent, the word is rewritten before the text reaches the Text-to-Speech engine:

• Soniox voices — the word is replaced with its niqqud-pointed form, which Soniox renders faithfully.
• DeepDub voices — the word is wrapped in an inline SSML <phoneme> tag carrying the IPA, which DeepDub honors.

Because this is a deterministic text substitution (not a hint to a model), an override applies on every utterance where the word appears.

​

Limitations

• Works best for proper nouns, brand names, and uncommon words
• Common Hebrew words are usually pronounced correctly without overrides
• An override only applies when the word appears exactly as written in the word field
• Maximum 50 entries per agent