Last updated
What is AssemblyAI used for in a Bolna voice agent?
AssemblyAI is a streaming speech-to-text provider that turns your caller's speech into text for the Bolna agent's LLM, using the universal-3-6-pro model trained on real voice-agent and phone audio. It is built for the turns that break voice agents: one-word answers, background voices and callers who switch language mid-sentence. Pick it when your calls are full of brand names or product codes, since it takes both a keyword list and a free-form description of the call.
How do I set up AssemblyAI with Bolna?
- Open your agent in the Bolna dashboard at platform.bolna.ai and go to the Languages tab.
- Under Transcription, set Provider to
assemblyand Model touniversal-3-6-pro. - Add your product and brand names in Keywords, and a sentence or two about the call in Context.
- If callers pause while reading out numbers or addresses, open Advanced settings and raise Longest pause in a turn. Mode and VAD threshold sit there too.
- Save the agent and place a test call.
For every setting, see the AssemblyAI setup guide in the Bolna docs.
Which AssemblyAI models and languages does Bolna support?
| Model | Languages | Status |
|---|---|---|
universal-3-6-pro | 32: Afrikaans, Arabic, Cantonese, Catalan, Danish, Dutch, English, Estonian, Finnish, French, Galician, German, Hebrew, Hindi, Italian, Japanese, Korean, Mandarin, Marathi, Norwegian, Norwegian Nynorsk, Persian, Portuguese, Romanian, Russian, Spanish, Swedish, Turkish, Urdu, Vietnamese, Xhosa, Zulu | Recommended default |
universal-3-5-pro | 19 of the above. It does not cover Afrikaans, Cantonese, Estonian, Galician, Korean, Marathi, Norwegian Nynorsk, Persian, Romanian, Russian, Urdu, Xhosa or Zulu | Previous generation, still supported |
universal | English only | Deprecated, not selectable on new agents |
Need an Indian language AssemblyAI does not cover? The docs point to Sarvam or Deepgram.
What can you build with AssemblyAI and Bolna?
Lead qualification and confirmations. Outbound calls that hinge on a short yes or no. universal-3-6-pro makes 52% fewer errors on short answers in heavy background noise than the previous model.
Multilingual support lines. One agent that code-switches across 32 languages, with 74% fewer sessions transcribed in the wrong language than universal-3-5-pro.
Calls taken in noisy places. Shops, households and open offices, where 22% less background speech leaks into the transcript. Set voice_focus to suppress other voices further.
Spelled names and card digits. Bolna passes the agent's last spoken reply to AssemblyAI as context, so when the agent asks for the last four digits, AssemblyAI knows a number is coming.