Claude Thailand Community Meetup 3 · 10 Oct 2026
3 Founders + Claude
vs. Google, ElevenLabs, OpenAI
in Thai TTS
in Thai TTS
Phatrasek JirabovonvisutCo-founder, Paxa Labs
Speko TTS benchmark · Thai
Fast
Time until the voice starts talking. The bars grow in real time.
Paxa, typical request
89ms
Faster than a blink.
Slow requests, p90: 122 ms. Still first.
8 to 12 times faster than Google, ElevenLabs and OpenAI.
8 to 12 times faster than Google, ElevenLabs and OpenAI.
Source: benchmarks.speko.ai · Thai voices, measured from Singapore, run 07-03 · CC BY 4.0Speko TTS benchmark · Thai
Robust
Does the voice say the right words for numbers, dates, money and operators?
Paxa, on Thai
1.00
Always right.
Tied for first.
Tied for first.
With Google Gemini, xAI and Soniox.
ElevenLabs: 0.00. OpenAI: 0.25.
ElevenLabs: 0.00. OpenAI: 0.25.
Source: benchmarks.speko.ai · Thai voices, run 07-03 · 1.00 means always right · CC BY 4.0Speko TTS benchmark · Thai
Natural
How human it sounds. 1500 is the field average.
Field average
Google Gemini1722
xAI Grok1719
MiniMax1699
Cartesia1688
ElevenLabs1605
Soniox1591
Paxa1467
Google Chirp 3 HD1442
Microsoft Azure1280
Inworld1202
OpenAI1084
Paxa, today
7th
of 11.
Above OpenAI.
Above Microsoft.
Above Google Chirp.
Above OpenAI.
Above Microsoft.
Above Google Chirp.
Behind Google Gemini and ElevenLabs.
Not best. Yet.
Not best. Yet.
Source: benchmarks.speko.ai · Thai voices, run 07-03 · ratings compare voices within Thai only · CC BY 4.0Speko TTS benchmark · list prices
Cost
US dollars per million characters. Same price in every language.
Paxa
$10
per million characters.
Cheapest on the chart. Half of OpenAI, a third of Google Gemini.
Source: benchmarks.speko.ai · vendor list prices · vendors without a listed price are not shown · CC BY 4.0Same voice, same script
Listen
First V1, the model on those charts. Then V2.
Listen on your phoneyoyoismee.github.io/claude-meetup-3
V1
On the charts today
V2
Coming this month
Each clip is about 18 secondsWho else is on that chart
Google
Microsoft
OpenAI
ElevenLabsTogether: about $9 trillion and 425,000+ people.
Market caps 8 Oct 2026 · latest filings · OpenAI March 2026 · ElevenLabs Sept 2026 · private headcounts estimated
Paxa Labs
3 people
$0 raised
$0 raised
Bootstrapped by the founders.
Sometimes, even flooded.
A disclaimer
This is not “if we can, anyone can.” AI is my home field.
I hope you can do it too.In your own field.
2020 vs. 2026
2020
Big companies. It took a lot of people to build one AI product.
2026
+
3 people + Claude get a lot of stuff done.
My 2020 team could only dream of this speed.
Headcount vs. productivity
Productivity HeadcountSmall team
Big companyMore people, more overhead
Small team + ClaudePlays above the curve
Output of a much bigger team
Claude can write a full squad’s sprint of code in a few hours.
So the key is scaling, and managing overhead.
The key is scaling
“It’s good that you’re good.
But one person can only do so much.”
But one person can only do so much.”
My ex-boss
Higher up, your impact is the difference you make to everyone around you.
That’s the game
DecideUnblockTeachDirectCoordinate
Same game. Now my team includes Claude.
Info flows up. Decisions flow down.
Context sits at the top, detail at the bottom. No one can hold it all.
Info goes upSummarized at every layer
Decisions go downMore detail at every layer
At 100,000 people, most of it isn’t relevant to you anyway.
Then the org grows.
Managing
and organizing
and organizing
← The work.Still down here.
The bigger the org, the more of it manages the work instead of doing it.
At Paxa: split by task
Team of 3BothClaudeHuman, for now
A human where it needs a human, or where we’re best. Often both. Claude everywhere else.
Info goes upClaude summarizes. We read.
Decisions go downWe spec it. Claude builds it.
More Claudes without direction is just another big org.
* Where our next hire would go, if we ever hire.
How we split the work
Four questions for every block.
Block
1Who’s better?
2Important?
3Risk if wrong?
4Can we mitigate it?
So
Research direction
Us, for now
Very
Game over
Hard to check
Team
Code
Claude
Yes
Bugs
Specs, evals, exact-match tests
Claude
Writing
Both
Yes
Reputation
We review every draft
Both
Claude by default. A human where it’s important, we’re better, and it’s hard to check.
Claude isn’t human. Optimize for that.
FasterProcesses information far faster than any of us.
Actually readsWe skip the docs and ignore email. Claude reads everything.
Artifact-drivenSpecs, feedback, evals. Documents are its native language.
ZoomableIt makes the work human-friendly at any level. HTML for humans, Markdown if simple.
Bureaucracy is your friend
Humans hate it because it’s slow. But it’s what keeps big orgs from breaking.
Document everythingPaper trails everywhere
Every department reviews itNitpicking random stuff
It reaches meI skim. Maybe I say no.
I can burn the whole project down. Claude isn’t mad, or burnt out.In a big org, that’s 40 people, a few months of work, and maybe some layoffs.For us, it’s one afternoon with Claude.
Burn tokens, not humans.
First, a safety net: same output.
Before anything changed, we had Claude build the tests.
Answer key
The version that works=
Every new version
Built by Claude“Make it faster.
Every sound must come out exactly the same.”
Every sound must come out exactly the same.”
What we told Claude
Then we kept asking.
Claude did the work. We kept raising the bar.
01
Rewrite it allin Rust.Even the libraries that
don’t exist in Rust yet.
NormallyA department’s worth of work.
02
Tune it forthe machine.Once it’s live.
Every chip has its own tricks.
NormallyLots of specialist work.
Wasted if you switch machines.
Wasted if you switch machines.
03
Squeeze. Again.And again.Every last millisecond.
Normally“Good enough” ships.
Every change, every time.
Remember the bureaucracy? This is it, for real.
TestedSame output as the answer key, or it’s out.
ReviewedClaude Code reviews all of it.
A lot.
A lot.
“Documented”Paper trails everywhere.
We steer. Claude does the volume.
Claude
“Porting the whole library is a significant undertaking.
I’d recommend using bindings for now,
and leaving the full port for a future phase.”
Us
“No. You do it.”
Speko TTS benchmark · Thai
Remember this?
Now you know how we got here.
Paxa, typical request
89ms
About half the speedup:
Claude alone.
Claude alone.
And every sound still comes out
exactly the same.
exactly the same.
Source: benchmarks.speko.ai · Thai voices, measured from Singapore, run 07-03 · CC BY 4.0Ceiling: our architecture and tech choices
Probably top half
Fastest
Without Claude
With Claude
Our architecture set the ceiling.Claude drove us to it.
Speed run
Research
Team of 3BothClaude
Claude reads and runs. We decide what’s worth running.
01
WatchClaude monitors papersand news.
Daily brief, or on demand.
02
PlanTry this. Mix that. Adopt this. Or design our own.03
ExperimentClaude implements, runs itand reports back.
We supervise.
04
ShipSome of it makes it into the next real release.Why we stay in the loop
Real costsGPU time costs real money.
Better oddsClaude alone succeeds less often than with us supervising. For now.
ContextWorld, business and tech.
Research is people talking, too.
Research is people talking, too.
Speed run
Business
Team of 3BothClaude
Claude scouts. Some leads get a template. Some get us.
ScoutClaude scouts a lot and reports potential partners.
Template + ClaudeOur idea as a template. Claude personalizes each one.
Human to humanWe write it. We talk to them.
Business is more than content. Human + human is still magical.
Your domain.
Your context.
Your org design.Spend the time. Totally worth it.
Your context.
Your org design.Spend the time. Totally worth it.
Your home field.
GLHF
Paxa V2 is coming this month. Stay tuned.
Speech and language AI, native in English, Thai, and Mandarin.
Facebookpaxalabsth
LinkedInpaxalabs
GitHubpaxalabs
Questions? Catch me after the talk.
Try Paxa Labs
100 free credits to start,
no card required.
no card required.
paxalabs.com