So there is a provider that expects body in the below format:

{
  "audioDuration": 0,
  "channelType": "MONO",
  "encodeAsBase64": true,
  "format": "WAV",
  "modelVersion": "GEN2",
  "multiNativeLocale": "string",
  "pitch": 0,
  "pronunciationDictionary": {},
  "rate": 0,
  "sampleRate": 24000,
  "style": "string",
  "text": "string",
  "variation": 1,
  "voiceId": "string"
}

and responds in this format:

{
  "audioFile": "string",
  "audioLengthInSeconds": 0,
  "consumedCharacterCount": 0,
  "encodedAudio": "string",
  "remainingCharacterCount": 0,
  "warning": "string",
  "wordDurations": [
    {
      "endMs": 0,
      "pitchScaleMaximum": 0,
      "pitchScaleMinimum": 0,
      "sourceWordIndex": 0,
      "startMs": 0,
      "word": "string"
    }
  ]
}

And vapi uses this particular structure for request and response https://cdn.discordapp.com/attachments/1211484481848873020/1323183441943986176/image.png?ex=6773961c&is=6772449c&hm=fca197ea2ffc7b75f8d712e4307fcf925768ded035d309655b7dc366cd7db2ba&