TECHNOLOGY
Conversation timing that decides when to speak and when to wait, modeling that captures people, and purpose-built evaluation. The technology we’ve built by running two live services ourselves.
CONVERSATION TIMING ENGINE
Doing one well is easy: always wait long and you’ll never cut anyone off — but you’ll be slow. Playmore aims to do all four at once, the way people do.
A support bot is explaining a delivery when the customer says “mm-hm.”
Breaks off at the “mm-hm,” goes quiet for over a second, then starts the explanation over.
Recognises it as a listening cue and carries right on with the explanation.
An English learner says “I went to the… um…” and pauses for over 2 s to find the word.
Cuts in after 1.45 s, so when the learner carries on, the two talk over each other for 1.55 s.
Judges the turn unfinished and waits; once the learner is done, it answers in 0.2 s.
A restaurant booking call: name, party size, time — each answered briefly.
Waits 1.45 s after every answer — over 4 s of dead air across three exchanges.
Each answer is clearly over, so it moves on within 0.2 s every time.
The bot is explaining a premium renewal when the customer cuts in: “Wait, it’s going up?”
Keeps talking for 3.6 s after the customer cuts in, right over everything they say.
Stops right after “Wait,” then apologises and explains 0.2 s after the customer finishes.
Scenes use synthetic voices. Playmore’s reply and stop times are exactly what the real engine decided on hearing the same audio; the typical rule answers after a fixed 1.45 s silence and stops once the transcript is finalised. Reply generation time is left out of both.
BENCHMARK
On human-judged real service calls, Playmore told when a turn had ended more accurately than every system compared. It waits for people who are still finding their words, and answers quickly when they’re done.
At every pause in real service calls, people judged whether it was time to answer, to wait, or to backchannel. We played the same recordings to each system and measured how precisely it picks out the moments to answer. On the representative items, Playmore scored higher than every system compared.
Systems that wait well answer late; systems that answer fast cut people off. Up and to the right means doing both well, and Playmore is the only one there.
On real Korean phone-line calls, Playmore is also near the top: waits 85%, answers 92%.
“Before, I could predict exactly when it would answer after I finished, so it felt mechanical. Now it responds with slightly different timing depending on the moment — it feels like talking to a real person.”
Most industry benchmarks are built from synthesized speech. But simply swapping the same words into a synthetic voice drops end-of-turn scores sharply and reshuffles the rankings — a brief pause before someone continues starts to sound like the end of a turn. So Playmore measures on real service calls, against what human listeners judged.
We compared the same seven systems on both tests. The system that ranked first on the public synthetic-voice benchmark drops to fourth on real calls — and Playmore moves from fifth to first.
We played the same recordings to 12 voice AI systems. Most stop quickly when interrupted — but also stop 88–100% of the time at a simple “mm-hmm.”
In Korean conversation, someone pausing to find words should be given time, and someone who has finished should get an answer right away. Doing just one of these is easy.
People react even while listening — “yeah,” “mm-hmm.” Playmore adds backchannels at the right moments, choosing one that fits and even the emotion of the next reply. In our listening sessions, no round ever preferred having no backchannels.
Each decision takes about 0.07 seconds — less than the blink of an eye, and roughly half the time of leading open turn-detection models. It’s built light enough never to break the flow: the moment the user stops speaking, it decides whether to answer or wait.
Measured October 2026. Service-call metrics come from our own call data judged by human listeners, and from real-call replay simulation.
TRANSCRIPT CORRECTION
The more a word matters — a name, a number, a technical term — the more precisely it’s understood. Because it truly understands the user, the conversation adapts to them naturally, and accurate call records make summaries, evaluations and analysis more reliable.
Recognition errors on key words in call records dropped by up to 68%.
Across insurance consultations, senior check-in calls and English conversation, key-word errors fell by more than half in every domain (52–68% fewer).
In conversations where something was misheard, the AI answers with wrong information about 63% less often.
In-house evaluation. Tested on real calls in insurance consultation, senior check-ins and English conversation with Deepgram Nova-3 live transcription, comparing results before and after correction.
We capture what someone has experienced, how they decide, and how they react in different situations. How we capture it depends on the purpose — a conversation partner, a customer, a specific expert. Expert Twins are corrected, reviewed and approved by the experts themselves.
Telling whether someone is done or just thinking, stopping when interrupted, nodding along with “mm-hm” and “right” while listening — Playmore’s conversation timing engine makes these calls, and it gets better every day in real service calls.
Would that person really answer this way? Would a customer really object like that? Does the conversation flow? We evaluate conversations against purpose-built criteria, find where they drifted, and fix it in the next version.
People don’t talk to us from a quiet studio. Playmore builds and runs its own noise suppression engine for real service conditions. The same voice, recorded in a noisy café, processed by each engine.
Playmore builds everything from the design that captures people to conversational flow and timing, evaluation and operations. Whatever models come next, we focus on making conversation more human.
Two patent applications filed in 2025: memory-based realtime conversation generation, and multi-dimensional conversation coaching and evaluation.
