Skip to main content
Deepgram is Bolna’s default transcriber for English. Its streaming Nova-3 model delivers sub-150ms transcription latency, making it the fastest option for real-time voice AI.

​
Quick config

Set this in your agent’s tools_config.transcriber block. See Create Agent API for the full schema. To use your own Deepgram account, connect it at platform.bolna.ai/auth/deepgram.

​
Supported models

Recommendation: Use nova-3 for English agents. Use nova-2 with language set explicitly for non-English.

​
Key settings

​
Endpointing guidance

  • 250ms — default; works well for fluent English speakers
  • 350–500ms — better for non-native speakers, elderly callers, or callers who pause mid-sentence
  • 100–150ms — fast-paced scripts; reduces dead air but may interrupt on pauses

​
Keyword biasing

Deepgram takes a term list, which Bolna sends as your agent’s keywords.
Which parameter Bolna uses depends on the model, and it picks the right one for you:
  • nova-3 and Flux use keyterm prompting, which takes plain terms and no weights. Bolna strips any term:number suffix before sending, and the dashboard warns when it is going to.
  • The nova-2 family uses the older keywords parameter, which does accept a per-term weight: Bolna:5, Plivo:3.
Keep the list focused. Deepgram budgets roughly 500 tokens across all terms, and 20 to 50 well-chosen terms outperform an exhaustive list. Deepgram has no equivalent of the context field, so that box is shown disabled in the dashboard. Use AssemblyAI, Soniox or Sarvam if you need context biasing. See Keywords and context for how to choose terms.

​
Multilingual config

Set language explicitly. Auto-detect adds 100–200ms latency and occasionally misclassifies.
For Indian languages, Sarvam may give better accuracy. See Multilingual Voice Agents.

​
FAQ

Use nova-3 for English — it’s faster and more accurate. Use nova-2 when you need a non-English language or a domain-specific variant (nova-2-phonecall, nova-2-medical, etc.).
Your endpointing value is too low. Increase it to 350–500ms. If callers have a strong accent or speak with filler words (“um”, “uh”), a higher endpointing value gives them more time to complete their thought.
Match the telephony provider’s output: Plivo and Exotel use linear16 at 16000 Hz. Twilio uses mulaw at 8000 Hz. If you use the wrong values, transcription accuracy degrades.
Yes — the same transcriber config applies to both outbound and inbound calls. There’s no separate configuration.