Voice systems for intake and routine questions.

Start with the call workflow

Define the questions the system should handle, the information it may collect and when a caller needs a person. Administrative intake and FAQs are the usual starting points; clinical advice is out of scope.

Write the workflow down as a list of call reasons before any provider is chosen. Projects more often fail at this step than at the model. An agent asked to hold an open-ended conversation performs worse than one scoped to the handful of call types the business actually receives, with everything outside that set transferred. A narrow, well-specified scope is also what makes the remaining work testable.

What the implementation can include

Speech recognition and synthesis, turn-taking, interruption handling, an agreed telephony provider and approved backend actions. Recording, transcripts and retention need explicit decisions. Language coverage depends on the selected providers and testing.

Where the response time goes

Response time is the sum of several stages rather than a single model setting: speech recognition, endpoint detection, model inference, speech synthesis and network transport. Each needs measuring separately before anything is optimised, because the dominant cost is rarely where it is assumed to be.

Streaming helps at three points. Streaming recognition produces partial transcripts while the caller is still speaking. Streaming model output lets synthesis begin before the full reply exists. Streaming synthesis returns audio in chunks so playback starts earlier, with trade-offs in phrasing and prosody. Transport then adds telephony codecs, buffering and the network distance between the caller, the providers and your own backend.

We agree a target for time to first audible response and test it on the intended network and telephony path. No sub-second figure is promised here; the number depends on providers, region and what the agent has to do mid-turn.

Telephony, numbers and the channel

A voice agent needs a channel. That usually means a SIP trunk or a programmable telephony provider, a number in the right country, and decisions about inbound routing, outbound dialling permissions and call transfer. Codec choice and audio sample rate affect recognition accuracy, so the channel belongs in the engineering scope rather than being settled later.

Where the agent must coexist with an existing phone system, we scope how calls arrive, which ones the agent answers and what happens to the rest. Warm transfer, cold transfer and voicemail fallback each behave differently and need choosing deliberately.

Interruptions, silence and overlapping speech

Real callers interrupt, pause mid-sentence, speak over hold music and go quiet while they find a document. Each case needs defined behaviour. Interruption handling means deciding how playback stops, how queued audio is cancelled, and whether a tool call already in flight is abandoned or completed.

Endpoint detection decides when the caller has finished speaking. Set too aggressively, the agent interrupts the caller; set too conservatively, every turn gains a noticeable delay. This is tested against pauses, background noise and accents rather than assumed to work from provider defaults.

Hindi, English and mixed-language calls

Callers in India frequently switch language inside a sentence. Mixed Hindi and English is the normal case rather than an edge case, and it affects recognition, synthesis and how numbers, names and addresses are transcribed. Digits spoken in one language and confirmed in another are a common source of errors in booking flows.

We evaluate candidate providers against recordings or scripts that reflect how your callers actually speak, including vocabulary specific to your business. Required language coverage and quality are tested for the project; they are not assumed from a feature list.

Actions the agent may take

Permitted actions determine both how useful the agent is and how much risk it carries, so they are defined as explicit tool contracts with typed inputs, so the agent can check availability, create a booking or look up an order without being able to do anything that was not agreed.

Write-actions get confirmation steps and idempotency keys, so a repeated or half-completed call does not create a duplicate booking. Integration failures are surfaced rather than concealed. When a downstream system is unavailable the agent states this and routes the call, instead of generating a reply around the error.

When the call must reach a person

Escalation is designed before implementation begins rather than added after testing. The agent hands off when the caller asks for a person, when confidence is low, when the request falls outside the agreed scope, or when a repeated misunderstanding suggests the call is going badly. Context collected so far should travel with the transfer so the caller does not start again.

Recording, transcripts and retention

Calls generate personal data. Whether calls are recorded, whether callers are told, how long audio and transcripts are kept, who can read them and where they are stored all need explicit decisions written into the proposal rather than inherited from a provider default. For deployments in India this includes consent and retention expectations under applicable data protection rules; we scope these with you rather than advise on them as a legal matter.

Evaluate with representative calls

Test background noise, interruptions, unsupported questions and service failures. Agree targets for response latency and task completion; there is no universal sub-second promise. The on-site receptionist sample is scripted text and does not demonstrate a deployed voice service.

A useful release gate is a fixed set of representative calls, scored the same way every run: task completion, transfer accuracy, time to first audible response reported at a percentile rather than an average, and recovery after an interruption. The set grows as real calls reveal cases nobody scripted.

What a voice project costs

Every engagement starts with a fixed-price discovery sprint at USD 2,500 over two weeks, which produces the architecture, initial screen designs, acceptance criteria, a running-cost model and a fixed price for the build. The sprint fee is deducted from the build.

Running cost for voice is per minute rather than per request, and it is the sum of telephony, speech recognition, model inference and speech synthesis. The sprint produces that model with your expected call volume and provider choices in it, so the per-call cost is known before the build is committed. What an AI feature costs to run →

Part of the AI Feature Implementation Sprint → · Explore our work → · Read the latency budget write-up →