Voicebots

A voice conversation that can listen, clarify and hand over.

Spoken interaction adds timing and uncertainty to a conversation. We design the listening, response and handoff together, so the system can collect the right information, handle corrections and report only actions that have been confirmed.

LISTEN → UNDERSTAND → RESPOND

Where this capability helps.

01

Service enquiries

Provide spoken guidance, collect the details of a request and route it with the context a service team needs.

02

Guided intake

Gather required fields in a short dialogue, read back important details and prepare a structured request.

03

Voice inside an application

Add a spoken interaction to an agreed task, with the microphone, playback and visual state designed together.

A correction changes the request, not just the transcript.

In this illustrative service-intake dialogue, the caller changes their preferred day while the assistant is speaking. The intended workflow stops the old response, updates the request and confirms the corrected preference.

ILLUSTRATIVE DESIGN / SAMPLE INFORMATION

Illustrative dialogue. This example contains no recorded or generated audio.

  1. Caller

    “I need a service visit. Tuesday afternoon would work.”

  2. Assistant

    “I have Tuesday afternoon as your—”

  3. Caller interrupts

    “Actually, make that Thursday morning.”

  4. INTENDED SYSTEM BEHAVIOUR

    Stop playback. Discard the unfinished confirmation and update the preferred time.

  5. Assistant

    “Thursday morning is your preference. The service team will confirm an available time.”

PREPARED REQUEST
Preferred time
Thursday morning
Appointment status
Awaiting confirmation
What if the service system does not respond?

The assistant preserves the request and explains the handoff. It does not announce a booking. Whether a retry or transfer is appropriate depends on the channel and agreed process.

A preferred time is different from a confirmed appointment. The assistant should describe the state returned by the booking system or the team handling the request.

See this in a customer support and voice workflow. The spoken response has its own pronunciation, pacing and playback decisions. Explore text to speech.

A conversation connected to the next step.

  • A spoken interaction for an agreed task, language and channel.
  • Handling for corrections, important details, confirmed actions and human handoff.
  • A call evaluation covering timing, understanding and the outcome the caller reaches.

Decisions that shape the system.

  1. 01

    Manage turns and interruptions

    We define when listening starts, how an end of turn is detected and what happens when someone speaks during playback. Stopping sound is one part; the conversation must also discard or revise the interrupted response.

  2. 02

    Treat important details differently

    Names, addresses, identifiers and quantities deserve explicit checks. We design clarification and read-back around the consequences of a mistake, rather than asking the same confirmation after every sentence.

  3. 03

    Connect the right channel

    A browser microphone and a telephone connection have different audio and transport constraints. We assess the available channel, recording conditions and integration options as part of the scope.

  4. 04

    Make waiting and handoff understandable

    A slow lookup needs a useful conversational state. A transfer needs the request, confirmed details and reason for handoff. We also define the response to lost audio or an unavailable business system.

How we would evaluate it.

Criteria are agreed for the intended task and tested on representative material. The items below describe the evaluation approach.

Responsiveness
Measure the time from the end of speech to useful output across recognition, model, tool and speech-generation steps.
Corrections
Test interruptions, changed details, short answers and pauses. Check that the intended request is preserved.
Task outcome
Review captured fields, confirmed actions and successful handoffs across representative calls.
Audio conditions
Test the selected language and channel with realistic noise, speech pace, accents and connection behaviour.

Practical questions.

Is a voicebot just STT followed by TTS?

Speech recognition and synthesis are important parts of many voice systems. The conversation also needs turn detection, interruption handling, state, confirmations and tool results. Some models combine audio tasks, but these interaction decisions still need to be designed.

Can it book appointments or make account changes?

That depends on an available integration and the permissions agreed for it. We define the required checks and the meaning of a confirmed result. If the destination cannot confirm an action, the conversation should not claim it is complete.

Which voices and languages can we use?

We select and evaluate candidate models for the intended language, voice and channel. Pronunciation, comprehension and recording conditions are tested with representative material before making coverage or quality claims.

What would make this useful for you?

Bring a short dialogue, the channel it needs to run on and the outcome a caller should reach. We can define an initial voice workflow around those requirements.

We build client solutions using established models. Integrations, data access and operating arrangements are agreed for the project. Our proprietary model programme explores future deployment options.

Discuss your projectExplore this in a walkthrough