Realtime Voice Pipelines: xAI and OpenAI on Taliiq
Taliiq's realtime voice mode routes speech straight through a single speech-to-speech model from xAI or OpenAI, cutting the latency of a traditional STT-LLM-TTS pipeline.
Two ways to build a voice agent
Every voice agent has to answer the same question: how does audio become a decision, and how does that decision become audio again? Taliiq gives you two ways to do it, and you choose per agent.
The Standard Pipeline is the familiar chain. Speech-to-text turns the caller's audio into text, an LLM reasons over that text and decides what to say (calling tools along the way), and text-to-speech turns the reply back into audio. Three specialized models, three vendors if you want, full control over every stage. This is where most of Taliiq's catalog lives: Deepgram, Soniox, and Hakim for STT; dozens of LLMs; ElevenLabs, Cartesia, and Groq for TTS.
The Realtime Model mode skips the relay race entirely. One speech-to-speech model listens to raw audio and speaks a raw audio reply directly, reasoning and tool calls happening inside that same model. No text handoff between three separate network hops means turn-taking feels tighter and interruptions land faster, something callers notice even if they can't explain why.
xAI and OpenAI, the two realtime providers
xAI's Grok Voice models (Grok Voice Fast 1.0, and the reasoning-capable Grok Voice Think Fast 1.0) are multilingual out of the box, with keyword biasing and function calling in the same session. One xAI key covers your whole voice stack, since Taliiq wires it through the same BYOK credential you'd already use for Grok's LLMs and STT/TTS.
OpenAI's Realtime API (the gpt-realtime family) brings a mature, widely deployed speech-to-speech stack with a broad voice library and strong function calling. It's the natural default when your other providers are already OpenAI-based.
What this means for your call quality
Realtime mode is built for the calls where every half-second matters: live phone support, fast-paced ordering, anything conversational. The tradeoff is real too. Your Brain LLM tab goes quiet in this mode, since the realtime model owns reasoning, and provider choice narrows to xAI or OpenAI. If you need a specific STT accuracy profile, a particular TTS voice from a third vendor, or a non-realtime LLM, the Standard Pipeline is still the right call.
Both modes live in the same Voice tab, so testing one against the other on your own workload takes a few clicks, not a rebuild.
See which one fits your calls
The fastest way to know is to hear both on a real call. Spin up an agent on either pipeline, connect it to a number, and compare.