TL;DR:
In speech, a pause that is fine on a screen feels awkward, because people expect a reply within moments.
Latency is the wait before the first sound. Streaming speech starts talking before the whole answer is ready.
Perceived speed also depends on signals: showing you that it heard you and is working on it.
Ask a friend a question and wait for the reply. Half a second of quiet is normal. Two seconds starts to feel odd. Five seconds and you wonder if they heard you. Voice is the same with an assistant: timing is part of the experience, and a slow start can make a correct answer feel worse.
This article explains why, what is happening in the delay, and the tricks that make a voice assistant feel quicker.
Why speech is less forgiving than text
When you type a question and wait, you can look at the screen, read something else, or sip your drink. The delay is invisible. In speech, silence is part of the conversation. People use tiny gaps to signal things, such as that they are thinking, did not hear, or have finished. A long gap gets read as a problem.
What happens during the wait
Between the moment you stop speaking and the moment you hear an answer, several steps take place. The system has to decide you have finished, turn your speech into text, work out a reply, and turn that reply back into sound. Each step takes time, and the delays add up.
The wait before the first sound is called latency. A system can lower it by making each step faster, and by overlapping them so the next step starts before the last one is finished.
Walk-through: a pause before an answer
Say a question takes a total of four seconds to be answered in one design, as an example. Here is the same question in two designs.
Design | What happens | What you feel |
|---|---|---|
Wait for everything | It works out the entire answer, then converts all of it to speech, then plays it. First sound after about 4 seconds. | A long, uncertain silence, then the answer |
Stream the speech | It starts speaking the first sentence as soon as it is ready, while still writing the rest. First sound after about 1 second. | A short pause, then a flowing answer |
The total content is the same in both. In the second design, the part you hear first is ready earlier, and the rest is prepared while you listen. The difference between four seconds of silence and one is large, even though the answer finishes at about the same time.
Perceived speed
What matters is how fast it feels, not only how fast it is. A few techniques help:
Show it heard you. A small visual change the moment you stop speaking tells you it is working.
Start with the answer. Opening with filler wastes the time you were waiting for.
Keep answers short by default. You can always ask for more.
Let you interrupt. If you can cut it off, a long answer costs you less.
Speed and trust
Speed changes how much you use a tool. If asking costs a long pause, you will ask only when you must. If it answers quickly, you ask the small questions that add up to real help. Quick replies also keep your attention on your task instead of the waiting.
Note: Fast is not enough on its own. A quick wrong answer is worse than a slow right one. The goal is a short wait to the start of a good answer.
Where the time goes
It helps to know the usual suspects when an assistant feels slow. Slow ends of the chain include a weak internet connection, a long question that takes time to transcribe, and a big request that needs many steps. You cannot change all of these, but you can change some: ask shorter questions, ask for the short answer first, and use a wired or strong connection for long sessions.
And if a voice assistant waits for a long silence before deciding you have finished speaking, it adds to the delay every time. Pressing a key to open a pill avoids that, because closing the pill tells it you are done.
What speed cannot fix
Some questions need time, such as ones that require reading a long document or doing several steps. A good assistant makes the wait visible, saying what it is doing, instead of leaving silence. Honest progress beats pretending it is instant.
Privacy and control
DuckyHelper is built around asking out loud and getting an answer out loud while you keep working. It only listens and looks while the pill is open and for that answer. The blackboard is there for longer explanations that are easier to show.
Try it yourself
Time yourself asking any voice assistant three simple questions. Count the seconds to the first sound. Then ask for a long answer and notice whether it starts talking before it has finished. That gap is the one that decides whether you keep using it.
Frequently asked questions
What is latency?
The delay between when you finish asking and when the answer begins.
Does a longer answer mean a longer wait?
Not with streaming. The first sentence can arrive quickly while the rest follows.
Why does it sometimes pause in the middle?
It may still be preparing the next part. Streaming is quick on average, but not always perfectly smooth.
Get started
