Introducing Ovation™ — the applaud Voice Engine

MALIBU — August 2026.Today we're switching every applaud call over to Ovation™, a real-time voice engine we built ourselves. Until recently we ran our calls the way most companies do, on rented voice infrastructure of the Retell variety. This post is about why we stopped, and what we learned building the replacement.

The phone call that broke our trust

The moment we decided to build our own engine wasn't a strategy meeting. It was a test call where the agent said, warmly and confidently, "that's just gone to your phone" — and no text had been sent. The model had narrated an action it never took. If you work with language models you'll recognise the failure; if you don't, imagine an employee who describes doing their job instead of doing it, in a tone so pleasant you almost don't check.

That single call set the design rule Ovation is built on: the model proposes, the engine disposes. The AI chooses the words. It does not get to decide whether a text was sent, whether the caller agreed to anything, or when the call ends. Deterministic code owns every action, and the agent is only allowed to say a message arrived after the engine has confirmed delivery. Boring, verifiable machinery around an eloquent model — that's the whole philosophy.

Latency is a rabbit hole with a floor

Everyone in voice AI talks about latency, usually with one suspiciously round number. Here's what we actually found. When our replies took two seconds, calls felt broken. When we forced them under half a second, something worse happened: the agent started answering people who hadn't finished talking. One of our first test callers got cut off mid-thought three times in a minute and told the agent, verbatim, that it was impossible to talk to. He was right.

The uncomfortable truth is that a silence timer cannot tell a breath from a finished thought. A person who says "the staff were lovely and…" is not done. A person who says "the staff were lovely." probably is. So Ovation stopped treating silence as a number and started treating it as a judgment. Finished sentences get answered in about a quarter of a second. A sentence trailing off on "and" or "so" or "um" gets patience. Genuine hesitation gets grace. It sounds obvious written down. It took us hundreds of test calls to stop getting it wrong.

The speed itself comes from a trick we call speculation: while you're still mid-sentence, Ovation is already drafting and voicing its reply in reserve. If you stop, the answer is simply there — confirmed turns land in as little as 300 milliseconds, and typical replies arrive in under a second. If you keep talking, the draft is thrown away and nobody ever knows it existed.

The bugs that taught us the most

A few discoveries from the build, in the spirit of honesty about how this work actually goes:

• For a while, callers reported the voice sounded faintly "fuzzy". The cause was a race between two threads sharing an audio queue — every so often a few milliseconds of speech simply vanished mid-word. The fix was three lines. Finding it was not.
• Our first echo protection worked by ignoring everything the engine heard while it was speaking. Effective against echo; also effective against humans. Anything a caller said over the agent's voice — "yes please", "no thanks, bye" — went straight into the void, and both sides waited politely for the other. Ovation now listens while it talks, tells its own echo apart from a real voice, and stops mid-sentence when you interrupt it.
• A "yeah" buried in the middle of a sentence is not an agreement. Early versions treated it as one and acted on it. The engine now only accepts a clear, standalone yes — and if you say "uh-huh" while the agent is talking, it understands you mean keep going, not stop.
• Pause halfway through a sentence and most transcription pipelines will quietly hand the AI only the second half. Ours did too, which is why an early build once responded to the fragment "the clinic was" with cheerful nonsense. Ovation stitches the whole thought back together before the model ever sees it.

A voice that adjusts to the person

People don't talk at one speed. Ovation measures each caller's pace as the conversation goes and drifts toward it — a little quicker with someone brisk, unhurried with someone who takes their time, with response timing adjusted to match. It filters out the television in the background rather than answering it. And when it needs a moment to think, it says so out loud, briefly, the way a person would, instead of leaving that strange dead air that gives call bots away.

How we test it

Before any change ships, it runs a gauntlet: simulated callers played by AI with assigned personalities — the rambler, the sceptic who opens with "is this a real person?", the driver who wants everything done in ten seconds, the one who just says "fine" to everything. Every conversation is scored for interruptions, repeated lines, dead air, timing, and how natural it reads to a blind judge. Ovation has been hardened across more than 5,000 simulated conversational exchanges this way, and the gauntlet grows every time a real call teaches us something new.

The numbers

Everything runs end to end on infrastructure covered by applaud's signed BAAs, so the whole pipeline is HIPAA-compliant — not eligible, compliant.

Ovation is live today with our founding practices. We could tell you it feels different, but the honest version is simpler: you'll know it when you don't notice it.