Universal-3.5 Pro Realtime is now available in Vapi

Why it matters

Universal-3.5 Pro Realtime posted a 1.22% pooled word error rate on Pipecat's open speech-to-text benchmark of real agent conversations — ahead of Deepgram Flux (1.62%) and ElevenLabs Scribe v2 (3.12%). For a receptionist bot or an outbound sales agent, that gap shows up as fewer "sorry, can you repeat that?" moments and fewer misrouted calls because the model misheard a name or an account number.

It also comes with the controls that actually matter for production voice agents:

• 18 languages with native mid-sentence code-switching, including Hinglish and English↔Spanish
• Keyterm prompting for up to 100 custom terms — names, product SKUs, industry jargon — so the model leans toward the words your callers actually say
• Agent context, which feeds the model your agent's own question so it hears the caller's reply through that lens (it cut word error rate by 10.2% across a 20,000-file voice-agent benchmark)
• A mode selector — min_latency, balanced, or max_accuracy — so you can trade off speed and accuracy without hand-tuning individual parameters

None of that requires touching a config file. It's all sitting in the Vapi dashboard now.

How to turn it on

Getting Universal-3.5 Pro Realtime running on an existing Vapi assistant takes about a minute:

1. Log into your Vapi dashboard and go to Assistants.
2. Open one of your active assistants.
3. Click the Transcriber card.
4. In the provider dropdown, select AssemblyAI.
5. Choose Universal 3.5 Pro from the model dropdown.

That's the whole setup. From there, the interesting part is what you do with the parameters underneath it.

Configuring the model

Once Universal-3.5 Pro Realtime is selected, Vapi exposes the parameters that actually move accuracy and latency for your use case.

Keyterms. You can boost up to 100 key terms — names, jargon, product terms — to steer transcription toward the vocabulary your callers use. This is the fastest lever for domain accuracy: a healthcare receptionist agent benefits from boosting drug and provider names, a fintech agent from boosting account and product terms.

Mode (under Advanced). Instead of guessing which low-level parameters to tune for better accuracy or lower latency, you pick a mode: min_latency for the fastest possible response, balanced as a sensible default, or max_accuracy when getting every word right matters more than shaving off milliseconds.

Prompt. Universal-3.5 Pro Realtime is a promptable model — you can describe the domain, topic, or scenario in plain language, and that context steers transcription toward what's actually being discussed on the call.

Language codes. You can pass in the specific language codes your callers will use, which biases transcription toward those languages. This also helps with code-switching: if you know your callers might move between, say, English and Spanish mid-call, specifying both languages helps the model track the switch instead of guessing.

Seeing it in action

The clearest way to understand what this unlocks is to watch an agent use it. In the walkthrough, David Lange built a quick demo assistant on top of Universal-3.5 Pro Realtime and talked to it about a receptionist use case:

That exchange is a good snapshot of how the pieces fit together for a real use case: agent context for a more personalized conversation, keyterms and prompting for domain accuracy, and language codes waiting in the wings for whenever the receptionist agent needs to pick up a call in Spanish, French, German, or any of the other 18 supported languages.

What's next

Universal-3.5 Pro Realtime in Vapi is live today — if you already have a Vapi assistant, switching your transcriber is a dropdown away. If you're starting from scratch, it's the same five steps above, applied to a brand-new assistant.

For the full parameter reference, see the prompting and keyterms docs, or read the Universal-3.5 Pro Realtime launch post for what's under the hood.

Frequently asked questions