2.4.2. Speech Integration for Agentic Interactions
💡 First Principle: Speech is two complementary services — speech-to-text (STT) turns spoken input into text the model can reason over, and text-to-speech (TTS) turns the model's text response into audio. An agent gains voice by wrapping its text core with STT on the way in and TTS on the way out.
Why care: voice agents are a named scenario in the refreshed exam ("convert speech to text and text to speech for agentic interactions"). The key insight is that the agent's reasoning stays text-based; speech is an envelope around it. This keeps the design clean — you don't need a special "voice model," you need STT and TTS bracketing your existing agent.
⚠️ Exam Trap: Don't imagine the language model itself "hears" audio in a voice scenario. The standard pattern is STT → text reasoning/agent → TTS. A design that tries to feed raw audio directly into a text completion is usually the wrong answer for a voice-agent requirement.
Reflection Question: In a voice assistant, why does keeping the reasoning text-based (STT in, TTS out) make the system easier to evaluate and ground than treating it as one opaque audio-to-audio step?
