We are facing a peculiar issue. VAPI applies a phonetic transformation to our custom-LLM response in the same call. We always send back to VAPI the response with our phonetic transformation already applied, but VAPI applies other transformations after. The problem is that vapi only applies the transformation the first time; after that moment, it never applies again. In this call, we sent both times the number plate of a vehicle in the same format (you can see in the picture, before our transformation and after our transformation). The format is always equal, but the first time the voice mispronounced the number plate, and the second time was correct callId: 70e71a54-7887-4c5b-861b-a3c766be629f url call: https://dashboard.vapi.ai/calls/70e71a54-7887-4c5b-861b-a3c766be629f https://cdn.discordapp.com/attachments/1364864305211641889/1364864305627136061/image.png?ex=680b3870&is=6809e6f0&hm=fe41a1b0d4587ab6cb40df608728dc0937039f0b9b15dfd6bd60a308e836c1be& https://cdn.discordapp.com/attachments/1364864305211641889/1364864305962418267/image.png?ex=680b3870&is=6809e6f0&hm=2c28d3d1bef39f918aa6574abcf0e8ce7f37d51e2353d6652e8db7f99df9f4dd& https://cdn.discordapp.com/attachments/1364864305211641889/1364864306373595190/image.png?ex=680b3870&is=6809e6f0&hm=626c22e07402dc57e37fd5ae25be59320be8ba8a7c626b588f5340787e027bfb&

To disable Vapi's phonetic transformations that are being applied after your custom transformations, you can toggle off the "Voice Input Formatted" feature. This can be done through the API for your assistant's voice formatting plan:

javascript
// In your assistant configuration
{
  "voiceFormattingPlan": {
    "enabled": false  // Disable Vapi's automatic phonetic transformations
  }
}

This will prevent Vapi from applying additional transformations to your already-formatted text. Sources: - [Voice Input Formatting](https://docs.vapi.ai/assistants/voice-formatting-plan) - [Special Characters in LLM Outputs](https://discord.com/channels/1211482211119796234/1342824967833653268)

@Vapi Support Bot but why is it only applied the first time?

Based on the information I found, the most likely explanation is a caching issue. The voice provider (not Vapi) may be caching the audio file or the chunk itself after the first transformation:

// Potential solution: Try clearing cache or adding a unique identifier
{
  "voiceFormattingPlan": {
    "replacements": [
      // Add timestamp or random string to force new processing
      { "type": "exact", "key": "uniqueMarker", "value": "" }
    ]
  }
}

This might force the provider to process each response as new content rather than using cached transformations. Source: - [Discord Discussion on Caching Issues](https://discord.com/channels/1211482211119796234/1342338583721414677)

@Vapi Support Bot tell me more about cachingEnabled for assistant with 11labs voices and about chunkPlan enabled true and false (pros and cons)

I found information about

chunkPlan

but couldn't find details about

cachingEnabled

for 11labs voices. For `chunkPlan`: - It controls how text is processed before being sent to voice providers - When enabled: Improves quality, allows audio control tokens like

<flush />

, and helps with non-English languages - When disabled: Relies on the voice provider's processing, may reduce latency, and bypasses Vapi's text formatting Could you clarify what specific functionality of

cachingEnabled

you're trying to understand for 11labs voices? This would help me provide more targeted information. Sources: - [Discord Discussion](https://discord.com/channels/1211482211119796234/1343543475383046185) - [Discord Discussion](https://discord.com/channels/1211482211119796234/1342138198230302870)

@Vapi Support Bot cachingEnabled is an option that you can modify at vapi assistants with 11labs voices. The default value is true but i would to understand why could be happen if we set the cachingEnabled false

Hey David! It looks like I need to get a human agent to help you out with this. I've escalated your query, and someone from our team will reach out to you soon. Thanks for your patience! 😊

cachingEnabled reuses previously generated audio from first cache, which is cost saving but less dynamic. if it is turned to false, it regenerates each time its requested which is more dynamic