Pronunciation & Keyword Boosting

Overview

Two separate mechanisms let you control how your assistant handles specific words: pronunciation dictionaries shape how the text-to-speech (TTS) voice pronounces a word out loud, and keyword/keyterm boosting shapes how the speech-to-text (STT) transcriber recognizes a word the caller says. Which one you need — or both — depends on the provider behind your assistant's voice and transcriber.

ElevenLabs voices support two pronunciation rule types:

Phoneme rules

Phoneme rules give the most precise control, using one of two phonetic alphabets:

• IPA (International Phonetic Alphabet) – e.g. /tə'meɪtoʊ/
• CMU Arpabet – ASCII-based, e.g. T AH M EY T OW

Phoneme rules only work with the eleven_turbo_v2 and eleven_flash_v2 models.

{
  "rules": [
    {
      "stringToReplace": "tomato",
      "type": "phoneme",
      "phoneme": "/tə'meɪtoʊ/",
      "alphabet": "ipa"
    },
    {
      "stringToReplace": "Sulus",
      "type": "phoneme",
      "phoneme": "S OO L UH S",
      "alphabet": "cmu-arpabet"
    }
  ]
}

Alias rules

Alias rules replace a word or acronym with an alternative spelling or phrase, and work with all ElevenLabs models. Use them to expand acronyms or standardize a phonetic spelling:

{
  "stringToReplace": "UN",
  "type": "alias",
  "alias": "United Nations"
}

Dictionary lookups are case-sensitive, so add separate entries for different capitalizations if needed. Rule order matters — the first matching rule is the one applied. Adding a pronunciation dictionary to an ElevenLabs voice also automatically enables SSML parsing.

Cartesia uses a text + alias format, supporting two alias styles:

Cartesia pronunciation dictionaries are only available on the sonic-3 model or newer.

{
  "items": [
    { "text": "Sulus", "alias": "SOO-lus" },
    { "text": "Nginx", "alias": "Engine-X" },
    { "text": "GIF", "alias": "<<ˈ|dʒ|ɪ|f>>" }
  ]
}

On the transcription side, Deepgram-backed transcribers support two complementary boosting features to improve recognition of specialized terminology and uncommon proper nouns. Both are configured in the transcriber section of your assistant's settings.

keywords accepts single-word tokens (letters and digits only) with an optional integer intensifier after a colon — e.g. apple, apple:3, apple:-2. Multi-word phrases ("order number") or non-letter characters aren't accepted in keywords — use keyterm instead. Intensifiers are exponential factors: the default is 1, higher values increase recognition likelihood, negative values suppress it, and 0 is equivalent to omitting the keyword.

{
  "transcriber": {
    "provider": "deepgram",
    "model": "nova-2",
    "language": "en",
    "keywords": ["snuffleupagus:5", "systrom", "krieger"],
    "keyterm": ["order number", "account ID", "PCI compliance"]
  }
}

• Start without any boosting or dictionary rules, and add them only where you observe real misrecognition or mispronunciation.
• Focus on uncommon words: proper nouns, brand names, and domain-specific terms the model tends to miss.
• Use keywords for single words and keyterm for phrases — don't put spaces in keywords.
• Avoid duplicate entries; sending each keyword once is sufficient.
• Use moderate intensifiers on keyword boosting to avoid introducing false positives.
• Provide the exact spelling and capitalization you want reflected in transcripts and dictionary matches.
• For very large vocabularies, a custom-trained transcription model may outperform keyword boosting.