Product Update: A New Voice and a Sharper Ear

After the August product update, in September we worked on two things: how the assistant sounds, and how well it listens.

A voice that doesn't stall

Callers judge an assistant in the first few seconds. Not just on what it says, but on how it sounds. A voice that goes quiet mid-sentence sounds like a bad line, even when the rest of the call goes perfectly.

That is exactly what our old voice sometimes did: the first words came straight away, then a pause of sometimes more than two seconds, then the rest of the sentence. So the digital assistant now runs on a new voice. We first measured both voices on the same set of Dutch sentences:

• 9 out of 10 sentences are ready within 1.2 seconds, against 4.3 seconds with the old voice.
• The longest stall we measured: 0.2 seconds, against 2.5 seconds. A gap that short is absorbed before the caller can hear it.

We also worked on the sound itself. On its way to the caller, the audio is already compressed twice. So the voice is now delivered without any extra compression, keeping as much quality as possible along the way. It turned out to be the faster option too.

Numbers the way a person says them

In July the assistant learned to read phone numbers digit by digit. In September we finished the job for the format mobile numbers are usually given in: +31, in groups. "+31 6 12345678" was sometimes read half as a number and half as separate digits. Now every digit is spoken on its own, even when the number is written in groups or with hyphens. It reads a web address like praktijk-oost.nl the way a person would, too.

It sounds like a detail. But anyone writing down a number while it is read out notices straight away.

A transcript that stays closer to the caller

A task from the digital assistant comes with a transcript of the call. That is where the team checks what was actually said.

Behind the scenes these are two different things: what the assistant understands, and what ends up on paper. The language model behind the assistant hears the audio itself and understands Dutch well. But the transcript came from a separate, general-purpose transcription model, which didn't always capture Dutch word for word.

So a separate speech recogniser, set up specifically for Dutch and with a vocabulary from general practice, now writes the transcript. The assistant listens as it did before, so it responds just as quickly. Only what appears on the task has become more reliable.

One call, one assistant

Internally, the digital assistant used to be made up of separate parts: one for reception, and separate parts for, say, a callback request. During a call they passed the baton to each other. At every handover, the next part had to rebuild what had already been discussed.

Now one assistant handles the whole call. The different requests are tasks it takes on, not separate people to talk to. Everything the caller has said stays with it, from the first sentence to the last.

The assistant only confirms something once it has actually been recorded. If a request can't be recorded, it doesn't say it has been noted, and the caller hears that too.

In the portal

• Digital assistant categories stand out. The category picker shows straight away which categories are tasks the assistant creates itself, separate from regular callback requests.
• Withdrawn requests in the history. If a caller withdraws a request with the assistant, it appears with its own label in the task history.
• Xelion: internal numbers can now also be called from the portal.

Where it runs

The improvements to the assistant run at the GP practice we are developing it with. You can also hear the new voice on our demo line: call +31 85 109 6240 or try it at plainwise.com/demo.

As always, your feedback shapes what we build. Use the feedback button in the portal or get in touch with your Plainwise contact. We read everything.

---

Every change, month by month, is in our changelog.