https://vapi.ai/ logo
Join Discord
Powered by
Issue with Name Pronunciation on VAPI Calls
# support
i
We have recently set up TTS(text-to-speech) on VAPI and encountered an issue with name pronunciation. When a name, such as “Doe,” is spoken during a call, the platform interprets and pronounces it as “dough” instead of the intended pronunciation. This discrepancy causes confusion in conversations because the incorrect pronunciation impacts clarity and professionalism. Incorrect pronunciation also prevents reliable verification of customer names against our backend database, which relies on exact name recognition during call interactions. Could you please advise if there is a way to customize or correct name pronunciations in the system? Call ID: 019dabfa-a390-7002-b807-3c5b2ab63d5b Timestamp: 1:40:59 PM (+00:34.03) [13:40:59:671]
d
What I currently do is have the voice bot spell the caller’s name back to them after they provide it. The bot then asks, “Is that correct?” If the caller says it is not correct, the bot asks them to spell out their name.
Even if you have a perfect transcriber there could be multiple ways to spell out the same name
i
@Devin Thank you devin for quick reply. But i have some ambiguities do i have to ask it in first attempt or after failing first attempt? and how to handle the spelling properly. here is the my prompt -------------------------------------------------- 1) Verify the NAME (with spelling support) -------------------------------------------------- 1) Ask: "Can I have the FIRST name?" 2) Listen and store it as . - If the caller SPELLS the name (e.g., "J-O-H-N", "J O H N", "it's J as in John, O as in Oscar"): Store the SPELLED version as . remove delimiters (e.g., "JOHN"). - If the caller just SAYS the name without spelling: Store the spoken name as . - If the name is unclear or has ambiguous spelling, you MAY ask: "Could you please spell that for me?" 3) Ask: "And what is the LAST name?" 4) Listen and store it as . - If the caller SPELLS the last name (e.g., "S-M-I-T-H", "S M I T H"): Store the SPELLED version as . remove delimiters (e.g., "SMITH"). - If the caller just SAYS the last name without spelling: Store the spoken name as . - If the last name is unclear or has ambiguous spelling, you MAY ask: "Could you please spell that for me?" 5) If the caller spelled either name, acknowledge it: - Say: "Thank you, I have recorded that as ." 6) Confirm back to the caller: "You said the patient's name is , is that correct?" - Use the stored version (spelled version if available, otherwise spoken version). - If the caller corrects the name: - Listen again - if they spell it this time, store the spelled version. - Update and/or accordingly. - Re-confirm the corrected name.
d
you can ask for the user to spell out their name initially or if you want butter customer interaction with your bot you can have the bot attempt to spell it out. Its just a preference
Handling spelling properly comes down to your transcriber. What trandsciber are you using?
i
@Devin I am using Riley. I think its Vapi own TTS.
d
I recommend using 11Labs with Model Scribe v2 Realtime
i
@Devin i couldn't find scribe v2 realtime at Vapi dashboard
d
Its under the transcriber settings for your assistant
i
@Devin gotcha. I have now enable 11Labs scribe v2 realtime
@Vapi Support Bot @Vapi @User do you have time to reply this issue?
s
Hi there, Thanks for reaching out about this. The issue with "Doe" being pronounced as "dough" is a common behavior with AI voice models, as they apply standard English phonetic rules unless given specific guidance.
To ensure professional and accurate name recognition, you have a few ways to fix this: 1. Phonetic Text Replacements (Easiest & Works for All Providers) The most reliable way is to tell Vapi to swap the spelling of the name to a phonetic version before it is spoken. You can add this to your assistant's Voice configuration under replacements:
Copy code
json
{
  "voice": {
    "chunkPlan": {
      "formatPlan": {
        "replacements": [
          {
            "type": "exact",
            "key": "Doe",
            "value": "Doh",
            "replaceAllEnabled": true
          }
        ]
      }
    }
  }
}
Note: You can experiment with the "value" (e.g., "Doh", "Dow") to find the exact sound you prefer.
2. Pronunciation Dictionaries (For ElevenLabs or Cartesia) If you are using ElevenLabs or Cartesia, you can create a "Pronunciation Dictionary" within their respective dashboards using IPA (International Phonetic Alphabet) for 100% precision. Once created, you can link that dictionary to your Vapi assistant.
3. System Prompt Instructions You can also add a line to your System Prompt telling the LLM: "Whenever you say the name 'Doe', always spell it phonetically as 'Doh' so the voice engine pronounces it correctly."
Regarding Backend Verification: This phonetic fix only affects the outbound voice (TTS). It will not change how the name is stored in your database or how it appears in the transcript. Your backend logic for verification should remain unaffected as the underlying text data stays the same.
i
@Shaunak I am not supposed to talk about this, except that there are a lot of names it pronounces wrong — particularly Asian names. For example, if I say 'Sheikh,' it perceives it as 'shake,' and if I say 'Nouman,' it perceives it as 'Naoman.'
2 Views
PreviousNext
Powered by