The Story of How Researching 'Crosstalk' for a Radio Show Between My Faves Was Pure Hell [chatGPT]

This is a continuation of this article. This was an attempt to use auto-generated

fave characters paired up to host a radio show via chat.

Manual (incomplete) is here

I am developing this mechanism to have AIs converse with each other, and my ultimate goal is to recreate my favorite ABEMA news reality show,
"
AbemaPrime" on chat. To do that, I believe
a 'crosstalk' effect

is unavoidable. It involves elements that would occur in human conversation, such as:

・Interrupting the other person and talking over them
・Rapid-fire exchanges
・Both parties speaking at the same time
・Small interjections or retorts while the other is speaking

and so on.

In the current radio show between my faves, the specification is basically that they don't interrupt until the other person finishes speaking (since it's chat), making it a very

peaceful and comfortable space.

But, speaking of Abema, wouldn't it be creepy if
Hiroyuki-san politely listened to the other person until the end (lol)?

So, I started verifying a crosstalk feature. I didn't realize at the time that I had opened the
gates of hell...

As usual, the generated characters used for testing are "Ano-chan" and her radio partner, her close friend the actress "Baiku-chan".

1. Pseudo-overlapping effect

Since you naturally cannot overlap words in a chat, I thought of:

① [Pseudo-overlapping of words]

Example:
A: That's definitely—
B: (Cutting in) Exactly!
A: Right?
B: Yeah, yeah

② [Restricted interruption]

A: And then—
B: (Wait, hold on, that's wrong)
A: No, let me finish!

③ [Simultaneous speech expression]

A & B: (Simultaneously) Whaaaat!?

I was thinking of effects using output control like this. I thought about making this a command called "Crosstalk Mode," where executing it would trigger the effects above during the conversation.

My ideal was to keep the overall volume of the conversation high, while occasionally having snappy dialogue, retorts, simultaneous speech, and overlapping occur.

[Results]

・Rather than crosstalk, it just made the tempo faster.
・Every turn became a short sentence prioritizing tempo, making the content thin.
It was actually a negative
・To strictly adhere to the function, the AI began ignoring the characters' relationships and naming rules.

In conclusion, it was a huge failure (crying).

2. Shifting to novel-style dialogue overlap expressions

I asked the AI characters themselves,
'I want to create this kind of crosstalk effect,
could you think of how to do it?'
(no pride left)

As a result, examples of writing that look like words are overlapping, as used in novels, appeared.
I designed this to be executed in each situation.

① [Interruption effect]

Example:
Ano-chan:
It exists.
When I was in middle school,
I tried to make chocolate
and it wouldn't set---Baiku-chan:Ah, I get it

② [Accidental overlapping words]

Example:
Ano-chan:That is---
Baiku-chan:Actually, I---
Ano-chan:Ah, sorry, go ahead
Baiku-chan:No, please go ahead, Ano-chan

It looked cool, so it gave a decent impression.
However, from here on, a hellish phenomenon began to occur.

[Results]

・When I explained the policy with examples as above, the AI characters would keep using the same phrases like demons (crying).
・If I set a rule to prohibit the same words, it would converge into harmless conversation.
・If I made the crosstalk effect mandatory, it would end up in unnatural or repetitive exchanges.
・If I set a rule for it to occur naturally according to the theme, the occurrence rate became abnormally low.

Around this time, I was having a fight with chatGPT.
What I realized up to this point was:

・If set to always-on mode, it is difficult to control
・Give up on diverse crosstalk effects and narrow it down to one
・Create a mechanism to trigger the function.
I narrowed it down to the effect of interrupting the other person and overlapping my own speech.

It is essential for humans to give up.

3. Using radio 'wrap-up' instructions as a trigger

Here, I came up with a brilliant idea (or so I thought at the time (crying)).

How about using a 'wrap-up' instruction like '30 seconds before the end of the segment' on a cue card used in radio or news programs as a trigger?
It was a mechanism like the following.

1. The user instructs Ano-chan to '30 seconds until the segment ends'.
2. Bike-chan is unaware of the instruction and continues talking at length.
3. To end the segment as instructed, Ano-chan tries to interrupt and end the conversation when the other character talks too long..
4. Conversely, Bike-chan is given a boost to keep talking for a long time.
5. As a result, wouldn't a comical sense of humor be expressed due to the discrepancy between the character who wants to talk at length and the character who wants to end it?

[Output Image]
Bike-chan: So, for my ideal holiday plan, I'd like to talk about 'nap time' next, but (Ano-chan: Ah, Bike-chan! Wait a second!) Eh, what? I'm at a good part right now—

Ano-chan: Hey, the director is giving me a huge '30 seconds until commercial' wrap-up signal! The nap talk is (Bike-chan: Eh, but I'm just getting started!) —Shut up! Please wrap it up, you're an adult, right!?

[Results]
・Even though the instructions are given to Ano-chan, Bike-chan ends up executing the 'wrap-up' instruction on her own, causing a reversal of roles.
・Perhaps because the normal radio structure is strong when prioritizing naturalness, the staging is barely noticeable at all
.・Demanding over-the-top reactions
leads to repetitive behavior and lines (always falling back on 'Let's stop for a moment' or 'We're out of time, so let's wrap it up').・The impression of interrupting the other person or overlapping speech is weak.
.・Even when the rules are followed correctly,
the conversation isn't interesting at all (crying)..・It becomes a dialogue like a philosophical debate.

I was starting to think, let's just give up on crosstalk if it's going to be this much trouble.

That was when it happened.
Could I be a genius?
The moment of realization arrived (a bit of an exaggeration).

4. An Overwhelming Flash of Inspiration

Output is forcibly stopped in the middle of a sentence, and then the output is resumed later, so wouldn't it look like
"interrupting the speech and overlapping words"—something that looks like crosstalk?

The mechanism is like this.

1. When the user inputs a theme,
specify the turn number to forcibly stop the output

Example
Free talk: How to use SNS
Crosstalk occurrence: 3rd turn

2. Output begins, and it is forcibly stopped at the 3rd turn's statement.
At this time, add a performance where the statement 'stops in the middle'.

Example
・Incomplete sentence stop type
'So at that time—'

・Emotional stop type
'Eh, wait a...'

・ Content interruption type
"So, actually, that—"

③ Output resumes when the user inputs "Continue" or "NEXT"
④ Output resumes from the last sentence of ②. If that statement is by Ano-chan, then Bike-chan is set to interrupt that statement and overlap with it.

Example:
Previous output:
Ano-chan: "It's such a pain, that—"
Resumption:
Bike-chan: "Wait a second. Isn't it a bit mean to call it a pain?"

And this mechanism hooked me more than I imagined.

5. Dialogue Example

① I executed the following input instructions.

Free talk: Should idols prioritize cuteness?
Or prioritize performance?
Crosstalk occurrence: 3rd turn

② When the user inputs "Continue" or "NEXT", the output resumes.

Because the speech stops visually, it creates a "cut off" feeling, and since the other character returns a rebuttal after resuming, it looks like they are talking over each other.
Also, if you shorten the number of turns and repeat it, you can expect the effect of it looking like they are arguing.

Furthermore, as a secondary effect, if you keep inputting "Continue" without setting a limit, they will talk about the same theme infinitely (though they will run out of material midway).

Additionally, this can be used not only in radio mode but also in normal conversation, so you can enjoy a taste like
"interrupting your favorite character's statement to interject your own opinion." Probably (lol)

6. Regarding the use of this project

It is possible to generate character files from the input form below, but
for now, you cannot use it unless I issue a ticket code.

The reason is that this mechanism passes the input content to chatGPT via API,
but it
costs money to execute (crying). It's a bit scary to release it for free, so

for now, it's for my personal use only.
However,
if you want to use it, please contact me via comments or something!
I won't charge you any money.

Also, I forgot to write this in a previous article, but
for now, this feature is intended for female characters.
(I think it's possible for males too, but it might not work as expected.)

7. Summary

I realized that generating crosstalk in AI dialogue is quite difficult.

Thinking about it calmly, I don't think this feature is that effective of a performance either.

However, I was happy that I was able to produce a similar effect by changing my perspective, and I decided to continue this pathological verification without giving up in the future.

Related articles