Voice quality is a performance variable.
A customer does not judge the model. They judge whether the conversation felt like a conversation. That feeling is engineered.
Perception is not cosmetic.
Small failures in voice behaviour change behaviour: customers hang up, repeat themselves or decide the system cannot help them.
The conversation is the product surface.
Latency, turn-taking and interruption handling are not polish added at the end. They decide whether the customer gives the process a chance to reach its actual purpose.
That is why voice behaviour is treated as a measured performance variable, alongside reach and resolution.
Latency
Delay before a response is heard.
Turn-taking
Knowing when the customer has finished.
Interruption
Handling overlap without losing the thread.
Six things voice has to get right.
Each one is observable and measurable in operation.
Natural conversation
Speech that carries a process without sounding scripted at the wrong moment.
Full duplex
Overlapping speech handled like a real exchange, not a walkie-talkie.
Interruption and barge-in
A customer can take the turn without being talked over.
Turn-taking
Recognising a finished sentence, a backchannel or a pause to think.
Latency
Response delay kept within the range that keeps a conversation flowing.
Prosody
Pacing and emphasis that make meaning land, not just words.
Half duplex and full duplex are different experiences.
The difference is audible to the customer long before it shows up in a metric.
Half duplex
Speak, stop, listen
- The customer can hear the mechanism
- Interruptions are missed or clipped
- Pauses are misread
- The exchange feels transactional
Full duplex
Overlapping speech
- Closer to how people actually talk
- Interruptions are handled in context
- Backchannels are recognised
- Harder to build, better to experience
Voice selection and transparency.
Voice choice and disclosure are process decisions, not only a brand decision.
| Decision | Options | Consideration |
|---|---|---|
| Voice identity | Synthetic, recorded, or hybrid | Consistency across volume and locale |
| Language and locale | Per-locale voices and pacing | Recognition and synthesis quality vary by locale |
| Disclosure | How the customer is told this is automated | Policy, jurisdiction and process expectations |
| Fallback | Human handover path | What happens when voice interaction fails |
Voice behaviour is measured, not assumed.
Interaction quality is reported alongside reach and resolution.
Interruption handling
97%
Share of customer interruptions handled without loss of conversation context.
Response latency
620 ms
Median time from customer end-of-speech to first audio response.
Turn confusion
3.1%
Share of turns where the customer had to repeat themselves.
Abandonment in first 30 s
6%
Share of connected calls ended by the customer in the first 30 seconds.
Recognition accuracy
94%
Correctly understood utterances in the reference configuration.
Availability
99.99%
Service level with 0 s average wait time in the referenced inbound configuration.
A convincing voice is not the goal.
The goal is a conversation that reaches the process outcome and leaves the customer in control.
- No claim of being human where disclosure is required.
- Escalation available when the conversation is not working.
- Voice behaviour measurable and changeable under control.
Where to go next.
Voice AI sits between process orchestration and telephony.
Evidence
Hear the difference in your process.
We will show how voice behaviour is measured against your outcome.