Conversation wait: reading that does not land on top of you

Comeyomi now has a conversation wait. While somebody is speaking it does not begin reading, and it starts once the conversation breaks. This addresses synthesised speech landing on top of the host in a busy Space.

What was happening

Reading a comment as soon as it arrives is the obvious design, and it has one weakness. A comment that arrives mid-explanation gets read over the top of it.

A Space is a single stream of audio, so two people talking at once means neither is intelligible. From the listener's side the host is drowned out; from the host's side they are talked over. The more people there are and the more often comments arrive, the more often this happens. Improving the quality of the reading does not touch it.

Waiting for a gap

With the wait on, Comeyomi listens to the Space and holds the reading while anyone is talking. It treats about two seconds of quiet as a break in the conversation and starts there.

There is no cap on how long it will wait. An earlier design gave up after a fixed interval and read anyway, which meant it still landed in the middle of a long explanation — which defeats the point. It now waits until the room is actually quiet.

So that its own speech is not mistaken for someone talking, audio is ignored while it plays and briefly afterwards. A run of comments will not keep stopping itself.

Only the newest of what piled up

Comments keep arriving while it waits. Reading all of them in order would fill the gap you finally got with synthesised speech.

So while the wait is active, it reads only the newest one at the point the room goes quiet. Older comments that were overtaken are skipped. Returning to the flow of the conversation, the reaction that just arrived is the more useful one.

Skipped comments are still in the web chat, so they can be read there. Sessions where the host talks at length will skip more, so switch the wait off during stretches where you want every comment picked up.

On, off, automatic

In the Space's web chat, "settings" at the top has the conversation wait. From the chat, !vad 入れる, !vad 切る and !vad 自動 do the same.

Setting	Behaviour	Suits
On	Always wait for a break	Explanation or interviews, where overlap hurts
Off	Read as soon as a comment arrives	Small Spaces with unhurried conversation
Auto	Switch according to how busy it is	When unsure (the default)

The default is automatic.

How automatic decides

It looks at the last minute. From off, it turns on when any of these is reached: ten or more listeners, ten or more comments per minute, or eight or more joins per minute.

Once on, it stays on until listeners fall below five, comments below four and joins below three. The thresholds for turning on and off are deliberately different, so the setting does not flicker on the boundary.

Whether it is currently waiting is shown at the top of the chat.

Choosing

For a small Space with unhurried talk, off feels better. Comments get read straight away and the exchange keeps its pace.

For a presentation or an interview, where one person speaks at length, choose on. The reading is delayed, but nothing gets trampled. If you cannot decide, automatic is fine.

The conversation wait changes when things are read, not how much. For stretches where knowing a reaction arrived is enough, combine it with SE mode, which plays sounds without speaking the text.

The reading modes are covered in the usage guide, and what to check when there is no sound is in the FAQ.