Hearing (speech to text)
The four hearing options, what each is best at, which languages they support, and how to choose one.
Hearing is how the caller's speech becomes text for the model. Choose it on the Hearing card of the agent's page (the panel is titled Hearing: "How the agent turns the caller's speech into text").
The options
| Option | Best for | Notes |
|---|---|---|
| Deepgram Flux | English, and the default | New English agents start on Flux when Deepgram is set up on the platform. English only. It streams the transcript and decides itself when the caller has finished speaking, so replies start sooner. |
| GPT-4o Mini Transcribe | Every language | OpenAI's hosted transcription. Works for every supported language. New agents use it when Deepgram is not set up. It transcribes once the caller has stopped speaking, so replies start later than with Flux. |
| Deepgram Nova 3 | Low delay in one language | About 150 to 300 ms. Listens in the agent's language, including Arabic, Urdu and Chinese. |
| Deepgram Nova 3 Multilingual | Callers who switch language | Follows a caller who changes between English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian and Dutch. Not for Arabic, Urdu or Chinese. Choosing it also makes the agent reply in the caller's language (see below). |
Each option shows its price per minute in the list. The default only applies to new agents: an agent you have already set up keeps the hearing you chose.
Hearing and the reply language
- Every option except Deepgram Nova 3 Multilingual listens for the agent's Language only, and the agent always replies in that language, even if a caller speaks another one or a knowledge document is written in another one.
- Deepgram Nova 3 Multilingual is how you choose a multilingual agent: it hears the caller's language, and the agent replies in the language the caller is speaking (starting in the agent's Language).
Step-by-step setup
- Open the agent, Assistant tab, and click the Hearing card.
- Open Hearing and choose an option. Options that cannot listen in the agent's language are greyed out with the reason.
- Press Save and test with Talk.
Configuration details
- Hindi callers who mix Hindi and English do best on Deepgram Nova 3 Multilingual.
- For Arabic, Urdu and Chinese use Deepgram Nova 3 or GPT-4o Mini Transcribe. Flux is English only, so changing an agent's language moves it off Flux (the page tells you when it does).
- If Deepgram is not set up on the platform, Deepgram options are greyed out and a call on one fails with a message saying the key is missing; it never falls back silently.
- Urdu agents wait for a short pause to decide the caller has finished, because the smarter end-of-turn detection was not trained on Urdu.
Testing
Speak the way your callers do (accent, pace, background noise). Check the transcript in the test panel or in History for mis-heard words.
Expected result
The transcript matches what you said, and the agent replies without long gaps.
Troubleshooting
See Speech recognition problems.
Related guides
Something here does not match what you see in the dashboard? Let us know.