Testing AI voices

all of it in one video

0:00 six models, one line
1:00 six voices
1:57 one line, four moods
2:13 a short scene

Every clip is also below on its own, so you can play two side by side. Headphones help.

Same voice (SG Manager), same words, same direction in brackets. Only the model changes.

what I typed

[excited] Good news, team! Our proposal was approved this morning. [laughing] I honestly did not expect it to happen so fast.

Drama 3
The newest model, released 13 October 2026.
Drama 3 (preview)
The same model under its earlier preview name.
S2.1 Pro
The current main model.
S2.1 Pro Free
The same model, offered free for testing.
S2 Pro
The previous version.
S1
The oldest, due to retire at the end of 2026. It reads directions in round brackets, so its line said (excited) and (laughing).

Singapore voices from Fish Audio’s public voice library, made by other users. I wrote each one a line that suits it.

SG Manager
[warm, encouraging] Okay team, quick one before lunch. The client loved the draft, so let’s tidy the last two slides and send it by four.
SG Woman Podcast
[relaxed, conversational] Welcome back to the show. Today we’re talking about the question every small business owner asks me: where do I even start with AI?
SG Guy
[casual, friendly] Eh bro, the new system quite easy to use leh. I tried it this morning, ten minutes only, done already.
SG Male
[calm, clear, like a documentary narrator] Every morning at six, the hawker centre wakes up before the rest of the city does.
Singlish Uncle
[cheerful, teasing] Aiyah, last time we write everything by hand. Now you all press one button, the computer talk for you. Not bad ah!
SG Aunty
[panicked] Wah, I thought I lost the file already! [sigh of relief] Okay, okay, found it. Don’t scare me like that.

Same voice (SG Woman Podcast), same words: “I didn’t expect to see you here today.” Only the direction in brackets changes.

Two voices, four lines, made one at a time and joined with short gaps. A café owner and a new staff member in his first week.

what I typed

Aunty [firm but kind] Jun Wei, the queue is very long already. Can you help me at the counter?

Jun Wei [nervous, apologetic] Sorry Aunty, I’m still learning the new till. [pause] Which button for takeaway ah?

Aunty [laughs softly] The green one lah. Don’t worry, first week everybody also like that.

Jun Wei [relieved, grateful] Okay, got it! Thank you, Aunty.

Voices: SG Aunty and SG Guy, both on Drama 3.

my pick
Drama 3: very good. The excitement sounds real, the laugh lands, and it takes its time. S2.1 Pro was good. The free version of the same model was decent.
brackets
The direction in brackets changes the delivery a lot. None of the voices read the brackets out loud, which I checked by running every clip back through speech-to-text.
singlish
The speech-to-text check stumbled on “leh”, “lah” and “not bad ah”. That says more about the checker than the voices. For local speech, trust your ears over a transcript.
a surprise
I sent one request with a made-up model name. It still came back with audio, no error. So a clip that plays is no proof of which model made it. Listen to two side by side before you trust a setting.
two twins
S2 Pro and S2.1 Pro came back the same length to the hundredth of a second. They may be the same model behind two names now.

1 · one line
Pick a line you’d really use, with a feeling in it. A flat sentence makes every model sound the same.
2 · one change
Change one thing at a time: the model, then the voice, then the mood. Keep everything else the same.
3 · listen blind
Rename the files so you can’t see which is which, then rank them. Your favourite may not be the newest.
4 · check rights
Library voices belong to the people who made them. Personal tests are fine. For work or paid content, read the terms or use a voice you own.