Platform
Every call is instrumented.
Nothing is a black box.
We engineer every millisecond from end-of-utterance to first audio byte. Here is exactly how the pipeline works.
Voice pipeline
From caller to response in one second.
Four components, each chosen for latency. Every stage streams — we don't wait for a complete transcript before starting the LLM, and we don't wait for a complete LLM response before starting TTS.
Voice-activity detection strips silence before any speech is sent downstream.
Real-time, streaming transcription with end-of-utterance detection.
Context-aware response generation with tool calling for actions.
Low-latency streaming synthesis — first audio chunk in ~180 ms.
Response time
Fast enough to feel human.
From the moment a caller stops speaking to the first audio byte back, Televox typically responds in about 600–700 ms — quick enough to feel like a real conversation, not a robot waiting its turn.
Every stage streams — we start the LLM before the transcript is final, and start speaking before the LLM finishes. We instrument every production call end-to-end and watch these numbers internally to catch regressions. Actual latency varies with network conditions and call complexity.
Real call, real actions
Every turn triggers something.
A dental practice call from hello to "booked" — watch the actions fire in real time alongside the transcript.
Under the hood
Six things that separate a good call from a great one.
Barge-in & semantic turn detection
We run Silero VAD continuously to detect when the caller starts speaking mid-response. Rather than a fixed silence timer, our EOU model combines acoustic energy with semantic completeness — so short answers like "yes" or "Thursday" don't trigger premature cut-offs, and callers can interrupt naturally.
Knowledge base & RAG
Each agent carries a vectorised knowledge base built from your uploaded documents — hours, services, pricing, FAQs, policies. At inference time the LLM retrieves the relevant chunks before generating a response, keeping answers grounded and hallucination-resistant.
Actions & tool calling
The LLM has structured tools: check-availability (Cal.com live), create-booking, send-sms, log-to-crm (outbound webhook), warm-transfer. Tools run server-side in the turn pipeline — no round-trips to the client, latency stays sub-second.
Warm transfer
When a caller needs a human the agent announces the transfer, keeps the caller on hold with status, dials the target number, and bridges the call only when the human answers. No caller abandonment from being sent to a ringing line.
Never overloaded
Each plan handles a set number of calls at the same time. If a spike pushes past that, extra callers hear a friendly message instead of dead air — and the system stays fast and stable no matter what. Your capacity is yours; another business's busy hour never touches your calls.
How we measure ourselves
Every turn — EOU, STT, LLM, TTS — is timestamped server-side and stored alongside the transcript. Our internal tooling shows a latency waterfall for any call. This is how we catch regressions and validate our latency claims — not a blind black box.
Live dashboard & human takeover
Watch every conversation live. Step in whenever you want.
AI-first doesn't mean human-never. The Live board shows every active call and chat across every channel, flags the moment someone asks for a person, and lets your team listen, join, or take over from the browser — then hand back without losing context.
One live board, every channel
Every active conversation — phone calls, web calls, WhatsApp calls, WhatsApp chats, and website chat — on one screen, updating every few seconds. Attention flags surface the moments that matter: "wants a human" when someone explicitly asks for a person (any phrasing, voice or chat), and "wobbling" when conservative heuristics suggest a text conversation is drifting.
Take over any chat
Click Take over on a WhatsApp or web-chat thread and the AI goes silent in that thread. You reply from the dashboard; the visitor sees your replies as "‹name› · team". Hand back and the AI resumes with the full conversation context. An optional per-agent setting signs WhatsApp replies with the teammate's name.
Listen to — or join — any live call
Listen to any live call straight from your browser. Or join and take over: the AI announces you by name, goes silent while transcription continues, and you talk to the caller. Hand back and the AI resumes with a context note. Works on phone calls, web calls, and WhatsApp voice calls (beta) — where a traditional cold transfer only works on phone calls.
An escalation ladder, not a blind transfer
If someone from your team is live on the board, a caller's first "get me a human" holds for a live join instead of transferring the caller away. Optional working hours on the transfer action — timezone plus weekly windows — mean the AI never rings a dark office; outside hours it takes a message or callback and flags the board.
Operator seats
Invite staff by email as Operators. They sign in with a magic link that lands straight on the Live board, and they see only the board and call history — no settings, billing, or agent configuration — enforced down to the database layer, not just hidden in the UI.
In your pocket
The dashboard installs to your phone's home screen as a PWA (on iOS, install it first). Opt in to push notifications and it pings you at exactly one moment: when someone asks for a human. No alert spam, by design — you open the Live board to listen, join, or take over.
Stated plainly: WhatsApp voice calls are in beta. Push notifications flag the wants-a-human moment; you listen, join, or take over from the Live board itself.
Reliability
Built for the rush. Never overloaded.
Your plan sets how many calls Televox answers at the same time. When volume spikes past that, nothing slows down or drops — extra callers simply hear a friendly message, and your capacity stays protected from everyone else's busy hours.
- Answers multiple calls at once — up to your plan's limit
- At your limit, extra callers hear a friendly message, never dead air
- Every call is logged, whatever happens
- Stays fast and stable even through a sudden rush
Built on
Best-in-class infrastructure, assembled with care.
We don't build what's already been built well. We compose the fastest, most reliable components in the voice-AI space.
FAQ