A caller pauses, then adds a caveat that changes the meaning of the sentence. A voice agent that starts answering during the pause has already acted on incomplete information. Waiting would have preserved the context, but waiting after every pause can make a conversation feel stalled.
That tension is central to the voice-evaluation work of Kaan Karakas, an applied AI engineer at Delphi, a platform for creating AI representations of people that others can chat with or call. These representations are called Digital Minds. Endpointing, a protocol that determines when the system should begin constructing a response, connects its interpretation of the caller’s speech to the backend work needed to answer.
Karakas’s main contribution to Delphi’s team-owned voice-call project involved evaluations, including studying strong sessions and their endpointing patterns. His work puts a practical question at the center of production voice engineering: what evidence should the system require before it starts answering?
A Pause Is Incomplete Evidence
Karakas illustrates the problem with a hypothetical speaker who pauses before supplying a caveat. An endpoint decision during that silence leaves the response without the qualification that would have changed it. The system can begin producing the wrong answer even though the speaker is about to supply the information it needs.
Silence measures a break in speech, but it doesn’t establish whether the thought is complete. Karakas says a pause can precede a contradictory statement, making semantic information from the call essential to the decision. The system has to interpret the meaning available so far alongside the possibility that the speaker will continue.
“Endpointing can become very detrimental if it happens at the wrong time because the system loses context,” says Karakas. “It might also interrupt someone who is being vulnerable or emotional and wants to talk about something.”
The consequences reach beyond answer content. Interrupting someone who is thinking can make the exchange feel conspicuously automated and weaken confidence in it. A natural-sounding response still needs the speaker’s full meaning. Vocal fluency cannot establish that the system waited for the context its answer required.
When Turn Detection Starts Spending Compute
In the pipeline Karakas describes, endpointing initiates backend work. The request leads to traversal of the person’s graph, which connects the represented person’s knowledge and experiences, and the construction of a response. The timing decision therefore affects both what the system knows and when it starts consuming resources to answer.
Karakas says: “When someone talks to a Delphi, there are moments when they think or pause. At the right moment, we have to send a call to the back end, which traverses the graph and begins constructing the response. If we trigger a lot of response generations that we never use, that wastes resources because we’re using a lot of inference compute and other resources.”
Starting generation before the turn is certain can shorten the silence a caller experiences. If the caller continues, however, the response under construction may never be used. Waiting for greater certainty avoids some of that unused work while increasing the delay before an answer.
Karakas says Delphi has tried models that predict whether a speaker has reached the end of a sentence. Their task includes distinguishing completion from a pause before more context arrives. His evaluation work examined the endpointing structures in strong voice sessions, grounding the timing question in actual exchanges.
For developers, the useful unit of analysis is the decision and its consequences together: the context available when work begins, the delay before a usable answer, and the inference spent on responses that do not reach the caller. Optimizing any one of these alone leaves the others unresolved.
Transcription Determines What Can Be Evaluated
Turn prediction depends on an accurate account of the incoming speech. If the transcription fails to capture the words, the endpointing decision loses the evidence needed to interpret the turn. Karakas says this dependency becomes demanding across the range of subjects discussed by their clients.
A voice agent built for sales can focus on the language and rhythms of a sales call, but Delphi has to accommodate people whose expertise spans many fields, each with its own vocabulary and ways of explaining an idea. As he explains, “We have a huge range of Delphi owners, and the range of topics they discuss is enormous. Therefore, we have to make sure our evaluation pipelines contain a good sampling of all the different types of conversations people might have.”
His work draws on Delphi’s strongest voice sessions to understand what makes their timing feel natural. Studying their endpointing patterns helps connect the caller’s pauses with the moment the Digital Mind begins responding. Trying incoming models across a broad range of conversations also lets the team explore how well they capture the words that guide those decisions.
Live use provides another view of how the system performs. Karakas says Delphi samples production responses and investigates hallucinations or drift, using what happens in ongoing conversations to identify problems and improve its answers.
Voice Traffic Turns Timing Into a Production Constraint
At Delphi, getting the timing right matters across a large volume of everyday conversations, where even a brief interruption or unnecessary wait can become a recurring frustration. “For voice calls, nearly 40% of our platform traffic involves voice calls rather than chat,” Karakas says, adding, “A six-figure number of conversations are happening daily on the Delphi platform, so there’s quite a lot of traffic.”
That daily total includes chat, but voice accounts for a substantial share of the activity. When the system repeatedly starts answering too soon, it spends compute on responses that callers may never hear. Holding back for too long makes people wait through pauses that interrupt the flow of a call. Transcription mistakes add another difficulty: a conversation can go smoothly on one subject and stumble on another if the system struggles with the vocabulary. The calls used to evaluate it need to reflect that variety.
Karakas studies the endpointing patterns in strong voice sessions to understand how the system can make these decisions more reliably. An accurate transcript gives the turn-prediction model the context to distinguish a finished thought from a pause before an important qualification. That judgment determines when the backend starts constructing an answer, so better timing can preserve the caller’s meaning while reducing work on responses that will never be used.
Evaluating those decisions across Delphi’s range of conversations helps the team understand when a pause gives someone room to finish and when it simply leaves them waiting.
Kaan Karakas’s work at Delphi shows how an answer can go wrong before the system has generated a word, simply because it decided too soon that the caller was finished. His voice evaluations address that problem where people experience it: in the pauses, interruptions, and responses that determine whether a conversation feels worth continuing.
Photo by Chase Chappell: Unsplash
Marcus Whitfield writes about developer tools, programming languages, and the software trends shaping how engineers build. Before joining DevX, he spent five years as a full-stack developer and two more running a small dev-tools newsletter that topped 10,000 subscribers.
