Pronunciation and speech control

Three layers decide how a voice AI Agent says names, numbers, and codes. The text of the reply comes first. The text-to-speech (TTS) settings then decide how the voice reads that text. The speech-to-text (STT) settings decide how well the AI Agent hears the same words from the caller.

Problem	Fix	Where
The voice reads a symbol, a date, or a URL badly	Write the text the way it is said	Voice guidelines and dialog responses
The voice reads numbers or abbreviations badly	Text normalization	Voice settings, text-to-speech row
The voice says a brand or a name wrong	Pronunciation dictionary	Voice settings, text-to-speech row
The voice is too fast or too slow	Speed	Voice settings, text-to-speech row
The transcript misspells a name or a term	Keyterms, keywords, or context terms	Voice settings, speech-to-text row

Write text the way it is said

For a reply that the AI Agent generates on a call, Moveo adds a voice formatting rule to the prompt. The rule removes markdown, emojis, and URLs. It also writes numbers, dates, amounts, and acronyms in spoken form. A text response that you write in a dialog does not always get this rule. Write those responses in spoken form yourself.

Before	After
`€1,250.50`	one thousand two hundred fifty euros and fifty cents
`15/03/2026`	the fifteenth of March, twenty twenty-six
`support@example.com`	support at example dot com
`GR16 0110 1250 0000 0001 2300 695`	the account ending in six nine five
`ATM`	A T M
`Visit https://example.com/pay`	You can pay on our website

Text normalization

Text normalization turns digits and abbreviations into words before the voice reads them. For example, it reads "1984" as "nineteen eighty-four" and "Dr." as "Doctor".

Provider	Setting	Values
ElevenLabs	Text normalization (`apply_text_normalization`)	`auto` (default) decides for each text. `on` always applies it. `off` reads the text as written.
MiniMax	English normalization (`english_normalization`)	On or off. It improves numbers and dates in English, with a small latency cost.

For eleven_flash_v2_5 and eleven_turbo_v2_5, ElevenLabs allows on only on an Enterprise ElevenLabs plan. On other plans, use auto and write numbers in spoken form, or use eleven_multilingual_v2.

Cartesia has no normalization setting in Moveo. For Cartesia, write numbers in spoken form.

Pronunciation dictionaries

A pronunciation dictionary holds rules that tell the voice how to say a word. For example, a rule can make the voice say SEPA as "see-pa". Moveo-hosted ElevenLabs and Cartesia rows have the toggle Use account pronunciation dictionary. A row with your own speech model does not show the toggle, because your key cannot read the provider account of Moveo.

Moveo keeps the dictionary in its own provider account. The name of the dictionary is your account ID, or starts with your account ID and _.
If no dictionary matches, the toggle is disabled and shows "No dictionary found for this account."
For ElevenLabs, the row saves the current version of the dictionary. After a change to the dictionary, open the voice settings and click Save again.

With the API, you can also set the dictionaries of your own provider account on a step with your speech model. An ElevenLabs step takes dictionaries in pronunciation_dictionary_locators. Each locator must have a pronunciation_dictionary_id and a version_id. Moveo ignores a locator without a version_id. ElevenLabs accepts at most three, so set no more than three. A Cartesia step takes one pronunciation_dictionary_id.

{
  "speech_model_id": "<speech_model_id>",
  "provider": "elevenlabs",
  "config": {
    "model_id": "eleven_flash_v2_5",
    "voice_id": "21m00Tcm4TlvDq8ikWAM",
    "pronunciation_dictionary_locators": [
      { "pronunciation_dictionary_id": "<dictionary_id>", "version_id": "<version_id>" }
    ]
  }
}

Speed and delivery

Provider	Settings
ElevenLabs	Speed 0.7 to 1.2. Stability 0 to 1, default 0.85. Similarity 0 to 1, default 0.75.
Cartesia	Speed 0.6 to 1.5, default 1.0. Volume 0.5 to 2.0, default 1.0. Emotion `neutral`, `calm`, or `content`.
MiniMax	Speed 0.5 to 2.0, default 1.0. Volume up to 10, default 1.0. Pitch -12 to 12, default 0.

Higher Stability keeps the voice consistent, but it can sound flat. Lower Stability adds expression. Click Test voice in the row to hear the result before you save.

Help the AI Agent hear terms

The speech-to-text row has settings that help the model hear names, products, and jargon.

Provider and model	Setting	Limit
Deepgram `nova-3` and Flux	Keyterms (`keyterm`)	30 terms, 128 characters each
Deepgram `nova-2`	Keywords (`keywords`)	30 words, 128 characters each
Deepgram `nova-3`, `nova-2`	Numerals (`numerals`) writes "twenty one" as "21"	On or off
Soniox	Terms (`context.terms`)	100 terms, 256 characters each
Soniox	Context (`context.text`)	10,000 characters

ElevenLabs scribe_v2_realtime has no vocabulary setting in Moveo. The next table shows an example of the effect of a term.

Caller says	Transcript without a term	Term to add	Transcript with the term
"Moveo"	movie oh	`Moveo`	Moveo
"Kifisias Avenue"	key fiss yes avenue	`Kifisias`	Kifisias Avenue

Next steps

Write the conversation rules in the voice guidelines.
Check names and numbers on a real call with Test a voice AI Agent.
See which providers support your language in Languages and providers.

Write text the way it is said​

Text normalization​

Pronunciation dictionaries​

Speed and delivery​

Help the AI Agent hear terms​

Next steps​