Voice agents are not just chat with audio attached. They are realtime systems that need to listen, decide, speak back, and stay usable when people interrupt, pause, switch devices, or speak in bad conditions. The hard part is not only the model. It is the full path around it: STT, TTS, transport, latency, consent, transcript policy, and failure handling.
That makes this a practical systems problem, not a demo problem. A non-technical buyer can still be the right fit if the business wants voice to handle real work without turning into a messy stack of partial transcripts, awkward synthetic speech, and brittle handoffs.
Voice breaks at the seams first
Most voice demos fail in the gaps between components. Turn-taking feels off. Interruptions arrive late. The agent speaks too long. Audio drops. Transcript quality slips under noise or accent variation. The fallback path is weak when speech fails. That is why we treat voice as one controlled operating layer, not as one speech model plus a nice voice.
STT and TTS are only two layers of the stack
Speech-to-text and text-to-speech matter, but they are only part of the job. We have worked with live transcription, streaming playback, voice activity detection, barge-in, latency budgeting, and browser audio paths with WebRTC, TURN, and STUN. The point is not only getting words in and out. The point is making the conversation feel usable while privacy, retention, and abuse boundaries still hold.
The platform choice changes the operating model
Voice work usually means choosing a stack, not one vendor. STT, TTS, and realtime orchestration can sit on different surfaces depending on latency, language coverage, voice quality, routing, and ownership needs. In practice, teams often compare or mix platforms such as ElevenLabs, Deepgram, Cartesia, OpenAI voice surfaces, browser speech, telephony layers, and custom transport around them. The useful question is not which provider sounds best in isolation. It is which combination gives the business the right control over speed, interruption behavior, transcript handling, and cost.
What we stabilize before rollout
The weight is in the operating layer around the voice. We can shape consent and recording rules, transcript retention, raw-audio handling, latency budgets, fallback to typed chat, handoff behavior when the agent should stop, and what the system should do when speech confidence drops. That is what turns voice from a flashy feature into something the team can actually own.
Strong fit, weak fit
The strongest fit is a team that already knows why voice matters and needs the system around it made more reliable. The weak fit is a team chasing voice because the demo feels modern, while ownership of privacy, escalation, and failure modes is still vague. In those cases, the speech stack is usually not the real blocker.
