C
CustomWeb AI
Please add support for OpenAI’s newer speech-to-text models in Vapi’s realtime voice pipeline:
  • GPT-Live-Transcribe
  • GPT-Transcribe
These should be selectable directly as assistant transcriber models, with support for:
  • Low-latency streaming transcription
  • Language hints
  • Domain vocabulary and keyword hints
  • Partial and final transcript events
  • Configurable endpointing
  • BYOK OpenAI credentials
This matters for multilingual and domain-specific phone agents. We operate a Swiss German de-CH voice agent where the current transcription pipeline still produces occasional product-name and pronunciation errors. GPT-Transcribe supports contextual hints and is intended for higher-accuracy speech recognition, while GPT-Live-Transcribe is intended for low-latency realtime transcription.
Please clarify:
  1. Which model is supported for live phone calls through Vapi?
  2. Whether Swiss German or de-CH language hints are supported.
  3. Whether custom vocabulary or keyword boosting is available.
  4. Whether BYOK OpenAI credentials are supported.
  5. Expected latency and pricing compared with Deepgram Nova-3.