Skip to content
TriedDocs
Build agents

Pronunciation

Teach the agent how to say brand names, surnames and technical terms with a pronunciation dictionary in IPA, CMU, pinyin or jyutping.

Voices sometimes mispronounce names, brands and jargon. The pronunciation dictionary maps a word to how it should sound. Before each sentence reaches the voice, Tried finds dictionary words in the text and replaces them with a form the voice will say correctly.

The dictionary is the Pronunciation list at the bottom of Speech Settings. It is stored in the agent's pronunciation_dictionary.

Add an entry#

  1. Open Speech Settings and scroll to Pronunciation.
  2. Click Add.
  3. Type the Word exactly as the model writes it, for example Nguyen.
  4. Choose the alphabet: IPA or CMU.
  5. Type the Phoneme, for example wɪn in IPA or W IH1 N in CMU.

Remove an entry with its trash icon.

Alphabets#

alphabet Write the phoneme as Example for "Nguyen" Dashboard
ipa International Phonetic Alphabet wɪn Yes
cmu CMU / ARPAbet phones separated by spaces, stress digits allowed W IH1 N Yes
pinyin Mandarin pinyin with tone numbers ni3 hao3 API only
jyutping Cantonese jyutping with tone numbers nei5 hou2 API only

How entries are applied#

Matching is on whole words and is case-sensitive. What replaces the word depends on the alphabet and the voice model:

Case What the voice receives
ipa or cmu A plain respelling generated from the phonemes, for example kuh-mah.
pinyin The pinyin with tone marks, such as nǐ hǎo.
jyutping The jyutping without tone numbers.
Note

Premium and Standard voices both receive the respelling, not raw phonemes. Write phonemes that respell clearly, and check the result with Test Audio.

Fields#

pronunciation_dictionary is an array of:

Field Type Description
word string The word to replace, as the model writes it.
alphabet ipa, cmu, pinyin, jyutping How phoneme is written.
phoneme string The pronunciation. 
curl -X PATCH https://api.tried.com.au/update-agent/agent_3f1c9a7b2d4e6f8091a2b3c4 \
  -H "Authorization: Bearer $TRIED_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "pronunciation_dictionary": [
      { "word": "Nguyen", "alphabet": "ipa", "phoneme": "wɪn" },
      { "word": "Quinoa", "alphabet": "cmu", "phoneme": "K IY1 N W AA0" },
      { "word": "你好", "alphabet": "pinyin", "phoneme": "ni3 hao3" }
    ]
  }'

The array replaces the whole dictionary. Send every entry you want to keep.

Other ways to fix pronunciation#

  • Spell it in the prompt. Tell the model to write the word phonetically: "Write our brand name as Kwin-oh-ah."
  • Spell out codes. Ask the model to wrap codes in <spell> tags so they are read one character at a time. See Prompt engineering.
  • Help the transcriber too. Add the same words to Boosted Keywords so the agent also hears them correctly. See Speech recognition.
  • Try another voice model. Pronunciation of names varies between models and providers.