Skip to content
Nappai Documentation

OpenAI STT

OpenAI STT is one of the components that act as the hearing of a voice agent: it turns what the caller says into text so the agent can understand it. It uses OpenAI’s transcription models. It transcribes while the caller speaks, so the word-based interruption settings of the Voice Agent, such as Minimum Words To Interrupt, work with it.

It plugs into the STT connector of the Voice Agent node. It holds the speech recognition settings the agent uses on every call.

Before transcribing, the audio goes through OpenAI’s noise reduction for the near voice of a phone or a headset: with background noise it transcribes better. Nothing to configure.

When to use it

  • The agent’s model is already from OpenAI and you want to use the same credential for hearing.
  • You want cloud recognition that transcribes while the caller speaks.
  • Callers may switch language and you want the agent to follow them.

Configuration

You need an OpenAI API credential with your OpenAI account key. See Credentials.

FieldWhat it does
Credential (required)The OpenAI API credential used to connect to the service.
LanguageCode of the language the caller is expected to speak, for example es or en. Default: es.
Detect LanguageWhen on, the caller’s language is detected and the agent switches language with them. Off by default.
Modelgpt-4o-transcribe (default) is the most accurate; gpt-4o-mini-transcribe is cheaper and a little faster. gpt-realtime-whisper is also available.

Advanced options

FieldWhat it does
Supported LanguagesLanguages detection can switch between. Default: es and en. Only used when Detect Language is on.

Output

The STT output holds the speech recognition settings. It connects to the STT connector of the Voice Agent node, or to Fallback STT when you turn on Fallback Engines and want it as a backup. The editor will not let you plug it into a TTS connector by mistake.

Example

An estate agency uses OpenAI as its voice agent’s model.

  1. Add OpenAI STT and connect it to the STT connector of the Voice Agent.
  2. In Credential, choose the OpenAI API credential the model already uses.
  3. Leave Model as gpt-4o-transcribe and Language as es.
  4. Test the agent in the playground by asking about a flat.

Good practices

  • OpenAI bills the usage to your own OpenAI account. Check your usage there.
  • The differences between local and cloud engines are explained in Quality, noise and latency.