Pronunciation dictionary

Voice models can’t always pronounce brand names, domain names and foreign product names correctly. The pronunciation dictionary tells the assistant how a written form (e.g. magaza.com.tr) should be spoken (e.g. “mağaza nokta kom nokta te re”). It lives in the assistant configuration under runtime.tts.pronunciations: {"written": "spoken"}.

{
  "runtime": {
    "tts": {
      "pronunciations": {
        "magaza.com.tr": "mağaza nokta kom nokta te re",
        "AirPods": "er pods",
        "X2": "iks iki"
      }
    }
  }
}

How it is applied

Rules (identical to the runtime speech_text module):

• The dictionary is applied after the Turkish transform. Each key is looked up both as written and in its transformed form, so magaza.com.tr is spoken the same way even inside an email address.
• Matching is on word boundaries and case-insensitive: airpods and AirPods hit the same entry, AirPodsPro does not.
• Longer keys first: if you have both iPhone and iPhone 16, “iPhone 16” matches first.
• Rows with an empty written or spoken form are ignored.
• The transcript and call record always keep the original text; the dictionary only affects audio.
• The dictionary is also used for ready phrases; when it changes, affected phrases are regenerated.

Limits: up to 200 entries; written form up to 200 characters, spoken form up to 400.

Pronunciation studio

The Assistants → assistant → Pronunciation tab lets you edit the dictionary and pick the right pronunciation by ear. The tab appears when Pronunciation studio is on in your plan.

1. Press New word and enter the written form that is mispronounced (up to 100 characters). Optionally add an example sentence (e.g. “Sitemiz magaza.com.tr adresinde.”); candidates are voiced inside this sentence.

2. Optionally record the right pronunciation in your own voice (webm, ogg, mp4, mp3 or wav; up to 2 MB and 15 seconds). The recording is transcribed and added to the candidates.

3. Press Generate candidates. The studio produces rule-based candidate spellings (up to 12) and voices each with up to 3 voice models. Candidate kinds:

   Kind	Example (ab.com.tr)
   Transform	how the runtime reads it without a dictionary
   Syllables	“a be nok ta kom nok ta te re”
   Commas	“a be, nokta, kom, nokta, te re” (with pauses)
   Joined	“abe nokta kom nokta tere”
   Hyphenated	“a-be nokta kom nokta te-re”
   Circumflex	with a lengthened vowel (kâ, lâ)
   English letter names	for English-leaning models

4. Pick the candidate × model cell that sounds right. The preview voices a sentence with the assistant’s own dictionary and model.

5. Publish. The dictionary (and the voice model, if you picked one) is published as a new version of the assistant. If nothing changed, the version stays the same.

Caution

Only admins and owners can publish the dictionary; members can generate and listen to candidates. If the assistant has other unpublished changes, the studio asks you to publish them first.

Audio files made by the studio are kept for 30 days and then deleted. Requests that produce audio (candidates, previews, recordings) are limited to 30 per minute per company. Your own provider key (BYOK) is used if set, otherwise the platform key.

Tips

• First listen to how the transform reads the word; most numbers, prices and emails are read correctly without a dictionary entry.
• For abbreviations (e.g. “KDV”), comma or hyphenated candidates often sound most natural.
• If a word is spoken wrong, use the pronunciation dictionary; if it is heard wrong (the customer says it and it isn’t understood), add it to keyterms.